Engine identification method based on aero-engine thermal jet mixed spectral characteristics

Abstract

The invention provides an engine identification method based on aero-engine thermal jet mixed spectrum characteristics, and the method comprises the steps: firstly carrying out the precise measurement of the thermal jets of different models of aero-engines through employing a Fourier infrared spectrometer, and enabling collected data to comprise the thermal jet spectrums independently generated by the engines of different models; spectral information of mutual mixing of thermal jets of various types of engines is obtained; then, the mixed spectrum of the thermal jets of various types of aero-engines is subjected to unmixing treatment operation through a VCA-LS algorithm, independent pure spectrums of the thermal jets of various types of engines are successfully obtained through VCA, the proportion of each pure spectrum in the mixed spectrum is deduced, and a solid foundation is laid for subsequent in-depth analysis; and finally, through a one-dimensional convolution deep learning MHSA-CNN algorithm, feature extraction is performed on different types of aero-engine thermal jets subjected to unmixing processing, so that internal characteristics and differences of the aero-engine thermal jets can be grasped more accurately, and powerful data support and technical support are provided for fault diagnosis of the aero-engine.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing spectrum processing, and in particular relates to an engine identification method based on the mixed spectrum characteristics of thermal jets of an aero-engine. Background Art

[0002] In recent years, the field of artificial intelligence has experienced rapid development, and a growing number of researchers have explored aeroengine fault diagnosis methods based on machine learning and deep learning. Chen et al. innovatively proposed a method that compensates for quantified actuator faults through dynamic effects. They constructed component-level models capable of fault identification for four typical actuators in high-bypass turbofan engines under both dynamic and steady-state conditions. Shen et al. proposed a novel generative transfer learning framework that leverages hard-constrained cycle-consistent adversarial networks (CycleGAN-HCs) to generate unpaired mechanical fault signals for training classifiers capable of operating across domains. This framework effectively facilitates feature transfer and data augmentation, and its effectiveness and reliability have been demonstrated in four cross-domain fault diagnosis tasks for piston aeroengines, demonstrating superior performance and enhanced generalization in complex mechanical fault diagnosis scenarios. Liao et al. innovatively proposed a new hybrid diagnosis algorithm. This algorithm utilizes an engine model, a feature extractor, a feature filter, a fault classifier, and several feature mappers. The feature mapper generates fault features based on incoming fault-free data, thereby rebalancing the tuning dataset. However, their research focuses solely on using big data technology to process and analyze massive amounts of data to uncover potential failure modes. They fail to clearly explain the principles of fault diagnosis, such as the data features and logical deductions used. Furthermore, ensuring real-time data analysis remains a pressing challenge.

[0003] The Fourier transform infrared spectrometer is designed based on the principle of dual-beam interferometry. After the interfering light illuminates the sample, the detector generates an interference pattern. By performing a Fourier transform on the interference pattern, an infrared spectrum is obtained, thereby analyzing and determining the elements, components, and molecular structure of the substance being measured. It requires no sampling or sample pretreatment, enables real-time long-distance monitoring, and can also perform rapid analysis of multiple components simultaneously. Passive remote sensing methods, among others, do not require the addition of an active infrared light source. The spectrometer directly receives infrared radiation signals, allowing for flexible changes in detection locations and detection distances of up to several kilometers.

[0004] In statistical methods, the unmixing of hyperspectral images is usually regarded as a blind source separation problem. Blind source separation can separate multiple mixed signals without prior knowledge and recover their respective original signals. There are many methods for blind source separation, and the more commonly used methods include independent component analysis (ICA), vertex component analysis (VCA), and non-negative matrix factorization (NMF). Spectral feature extraction refers to the process of finding features that can represent the essential information of the spectrum and have unique identifiability from complex spectral data. Common feature extraction methods include PCA, PLS, LDA, etc. In recent years, deep learning methods have also been widely applied to spectral feature extraction. By constructing a multi-layer neural network structure, they can automatically learn the deep features in spectral data and show strong advantages and potential when dealing with large-scale and complex spectral data. For example, good results have been achieved in fields such as hyperspectral image classification, substance composition recognition, and speech processing. Alaghbari et al. proposed a new integrated model based on a deep autoencoder (AE) for anomaly detection and feature extraction. First, the AE is trained and then used to detect anomalies. The trained AE model is then used again to extract useful low-dimensional features for the anomaly data. Demir et al. used the Resnet101 network in the construction of an example projector, and the feature generator created 6,000 microplastic features. Using neighborhood component analysis (NCA), the best 1,000 features were selected from the 6,000 feature pool. Then the k-nearest neighbor (kNN) algorithm was used to classify the generated feature vectors. Dong et al. proposed a three-residual fusion network called MSCR-FuResNet (multi-scale feature extraction and enhanced fusion of channel and residual block network), which enhances detailed feature extraction through multi-scale feature extraction. Then, by suppressing channels and pixels, the discrimination of similar features is improved. Finally, by modifying the activation function and residual block, low-contrast feature extraction is increased.

[0005] As the core component of an aircraft, the reliability of an aero-engine is directly related to flight safety. With the continuous progress of aviation technology, the design and manufacturing processes of engines have become increasingly complex, and fault diagnosis technologies have also been continuously developing. Currently, the fault diagnosis of aero-engines mainly relies on the following technical means: 1) Manual detection technology, which constructs a fault diagnosis model by combining the knowledge and experience of domain experts to assist technicians in making fault judgments. However, the diagnosis process overly relies on the knowledge and experience of experts, and has disadvantages such as subjectivity. 2) Sensor technology, which installs sensors at key parts of the engine to monitor parameters such as temperature, pressure, and vibration in real time, and can detect abnormalities in a timely manner. However, the cost of high-quality sensors is relatively high, and high maintenance costs are required. Moreover, sensors generate a large amount of abnormal data, and a large number of signal processing algorithms are also needed for analysis and processing. 3) Remote monitoring technology, which realizes real-time monitoring and remote diagnosis of aircraft engines worldwide through satellite communication and ground stations. However, it has disadvantages such as data transmission security, over-reliance on satellite communication, and the complexity of remote monitoring system technology. Summary of the Invention

[0006] To solve the above problems, the present invention provides an engine identification method based on the mixed spectral characteristics of aero-engine hot jets, which can obtain the independent pure spectra of the hot jets of each type of engine, and perform feature extraction on the hot jets of different types of aero-engines after unmixing processing, and can improve the identification accuracy of each type of engine.

[0007] An engine identification method based on the mixed spectral characteristics of aero-engine hot jets includes the following steps:

[0008] S1: Use a Fourier transform infrared spectrometer to collect multiple mixed spectral samples of the hot jets of the aero-engine to be measured;

[0009] S2: Use vertex component analysis method to perform unmixing processing on each mixed spectral sample of the hot jets to obtain the independent pure spectra that make up each mixed spectral sample of the hot jets;

[0010] S3: Input each independent pure spectrum into the MHSA-CNN network model respectively to obtain the types of the aero-engines to be measured corresponding to each independent pure spectrum.

[0011] Further, using the vertex component analysis method to perform unmixing processing on each mixed spectral sample of the hot jets to obtain the independent pure spectra that make up each mixed spectral sample of the hot jets specifically includes the following steps:

[0012] S21: Calculate all the mixed spectral samples R of the hot jets at different wavenumbers m =[r m1 ,r m2 ,r m3 ,…,r mnAverage of the brightness temperature where m = 1, 2, …, L, L is the number of wavenumber types included in all hot jet mixed spectrum samples, and R m is the set of hot jet mixed spectrum samples at the m-th wavenumber, and r m1 ~r mn are n hot jet mixed spectrum samples in the set R m ;

[0013] S22: Remove the DC components of all hot jet mixed spectrum samples at different wavenumbers respectively:

[0014]

[0015] where R m0 is the set of hot jet mixed spectrum samples at the m-th wavenumber after removing the DC component;

[0016] S23: Project R m and R m0 at different wavenumbers into a two-dimensional subspace respectively:

[0017] r mp = U d ′ * R m0

[0018] r m = U d ′ * R m

[0019] where U d ′ is the transposed two-dimensional projection matrix, r mp is the set of hot jet mixed spectrum samples at the m-th wavenumber after removing the DC component in the two-dimensional subspace, and r m is the set of hot jet mixed spectrum samples at the m-th wavenumber in the two-dimensional subspace;

[0020] S24: Obtain the signal-to-noise ratio SNR m corresponding to different wavenumbers:

[0021]

[0022] where SNR m is the signal-to-noise ratio corresponding to the hot jet mixed spectrum samples at the m-th wavenumber, P rm is the mean vector related to r mp ; P ym is the mean vector related to R m ; and there is:

[0023]

[0024]

[0025] where r pFmi represents flattening the i-th hot jet mixed spectral sample in the set r mp into a one-dimensional vector; represents the transposed R Fmi represents flattening the i-th hot jet mixed spectral sample in the set R m into a one-dimensional vector;

[0026] S25: Respectively determine whether the signal-to-noise ratio SNR m corresponding to each wavenumber is greater than the set signal-to-noise ratio threshold SNR th . Project the set r m corresponding to the wavenumbers with a judgment result of yes onto a two-dimensional subspace, and project the set r m corresponding to the wavenumbers with a judgment result of no onto a one-dimensional subspace;

[0027] S26: Normalize all the sets of hot jet mixed spectral samples projected onto the two-dimensional or one-dimensional subspace;

[0028] S27: Respectively obtain the projection vectors of all the sets of normalized hot jet mixed spectral samples:

[0029]

[0030] where P0 is the initial projection matrix, is the i-th hot jet mixed spectral sample in the set of hot jet mixed spectral samples at the m-th wavenumber after normalization; Z mi is corresponding projection vector;

[0031] S28: Respectively obtain the Euclidean norms of all the projection vectors Z mi , and use the one corresponding to the maximum Euclidean norm as the estimated value of the first independent pure spectrum ;

[0032] S29: Let and update the initial projection matrix P0 according to U1:

[0033]

[0034] where U1 is the set of independent pure spectra, I is the identity matrix, T represents transpose, and P1 is the updated projection matrix;

[0035] S210: Use the updated projection matrix P1 and the remaining after removing the hot jet mixed spectral samples corresponding to Repeat steps S27 - S28 to determine the estimated value of the second independent pure spectrum And so on until all independent pure spectra are determined.

[0036] Furthermore, the average value of the brightness temperature in step S21 is calculated as follows:

[0037]

[0038] where h represents the Planck constant value, c represents the speed of light, v m represents the m-th wave number, k represents the Boltzmann constant value, L(v m ) represents the radiation flux of the unit beam corresponding to the m-th wave number, and ln{·} represents taking the logarithm.

[0039] Furthermore, after all independent pure spectra are determined, the mixing ratios of each independent pure spectrum in each wave number are as follows:

[0040]

[0041] where A is the mixing ratio matrix of each independent pure spectrum in each wave number, M is the matrix composed of the estimated values of each independent pure spectrum, is the matrix composed of the normalized thermal jet mixing spectrum samples corresponding to the estimated values of each independent pure spectrum.

[0042] Furthermore, the MHSA - CNN network model includes three convolutional layers with different scales, a multi - head self - attention mechanism module, a max - pooling layer, a flattening operation, and a fully - connected layer.

[0043] Beneficial effects:

[0044] The present invention provides an engine identification method based on the thermal jet mixing spectrum characteristics of an aero - engine. First, a Fourier transform infrared spectrometer is used to accurately measure the thermal jets of different models of aero - engines. The collected data includes the thermal jet spectra generated by each model of engine alone and the spectral information of the thermal jet mixtures of each type of engine. Then, the VCA - LS algorithm is used to perform unmixing processing on the mixed spectra of the thermal jets of each type of aero - engine. Through VCA, the independent pure spectra of the thermal jets of each type of engine are successfully obtained, and the proportion of each pure spectrum in the mixed spectrum is deduced, laying a solid foundation for subsequent in - depth analysis. Finally, the present invention uses a one - dimensional convolutional deep - learning MHSA - CNN algorithm to extract features from the unmixed thermal jets of different types of aero - engines, so as to more accurately grasp their internal characteristics and differences, providing strong data support and technical guarantee for the fault diagnosis of aero - engines. Description of the Drawings

[0045] Figure 1 An engine recognition framework diagram provided by the present invention based on the mixed spectral characteristics of the hot jet of an aeroengine;

[0046] Figure 2 A schematic diagram of the principle of the vertex component analysis method provided by the present invention;

[0047] Figure 3 A structure diagram of the MHSA-CNN network model provided by the present invention;

[0048] Figure 4 A schematic diagram of the process of extracting characteristic peaks by the attention mechanism provided by the present invention;

[0049] Figure 5 A schematic diagram of the type-I spectral characteristics extracted by the MHSA-CNN provided by the present invention;

[0050] Figure 6 A schematic diagram of the type-II spectral characteristics extracted by the MHSA-CNN provided by the present invention. Detailed implementation manners

[0051] 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.

[0052] The technical solution of the present invention generally includes remote sensing spectral data acquisition, mixed spectral unmixing, and spectral feature extraction; specifically, an engine recognition method based on the mixed spectral characteristics of the hot jet of an aeroengine includes the following steps:

[0053] S1: Use a Fourier transform infrared spectrometer to collect multiple hot jet mixed spectral samples of the aeroengine to be measured;

[0054] It should be noted that the present invention performs field measurements on the hot jets of two types of aeroengines. As Figure 1 shown, the spectrometer is respectively kept at distances of 5 m and 10 m from the aeroengine, perpendicular to the nozzle outlet of the exhaust flow. During the measurement, the outdoor temperature is 26.3 °C and the humidity is 56.4% Rh. The measuring instrument is an EM27 Fourier transform infrared spectrometer, and the measurement method is a passive measurement mode, and the spectral resolution is 1 cm -1 , and the spectral measurement range is 2.5 - 12 μm, and the full viewing angle can reach 30 mrad.

[0055] S2: Use the vertex component analysis method to perform unmixing processing on each hot jet mixed spectral sample to obtain each independent pure spectrum that composes each hot jet mixed spectral sample;

[0056] S3: Input each independent pure spectrum into the MHSA-CNN network model to obtain the types of each aero-engine to be measured corresponding to each independent pure spectrum.

[0057] It should be noted that the spectral samples collected in the present invention only contain data of two types of engines. Based on this, in the process of unmixing using the vertex component analysis method, it is assumed that the kth mixed spectral data vector (i.e., the vector corresponding to the kth sample) is represented by r k (k = 1, 2,... n), then r k can be expressed as follows:

[0058]

[0059] Among them, taking the pure spectrum as the endmember, m j represents the endmember spectrum, a j is the abundance value of the pure spectrum corresponding to the endmember, p is the number of endmembers. In this embodiment, since only two pure spectra are mixed, p is equal to 2, and n k is the noise in the spectrum m j . The physical meaning of the abundance a j is the proportion occupied by each endmember at each wavenumber. Among them, the wavenumber is the reciprocal of the wavelength. Therefore, it is necessary to satisfy the "sum to one" constraint and the "non-negative" constraint. The formula is as follows:

[0060]

[0061] In this embodiment, it is necessary to satisfy a1 + a2 = 1.

[0062] It should be noted that the vertex component analysis VCA is based on the linear mixing model assumption, which believes that the observed mixed spectrum is linearly mixed by multiple pure endmembers in a certain proportion. Correspondingly, in the research of the present invention, the present invention believes that the mixed spectrum of the two types of engine hot jets is mixed by the pure spectra of these two types of engine hot jets in a certain proportion. After unmixing and extracting the endmembers by the VCA method, that is, extracting the pure spectra of the two types of engine hot jets from the mixed spectrum, the present invention selects the least squares method to determine the proportion of each endmember in the mixed spectrum, that is, the abundance value.

[0063] The vertex component analysis method is based on the concept of simplex. In the mixed spectral data of the present invention, each spectrum can be regarded as a point in a high-dimensional space, and the pure spectrum is located at the vertex of this space. After the affine transformation of the projection, the spectrum representing the pure substance is still located at the vertex. Therefore, the present invention projects the data onto an orthogonal subspace direction and calculates the vector with the largest projection distance value, which is the endmember to be found. As Figure 2As shown, the black dots represent all mixed spectral data vectors, and the singlet S p is the projection of the u vector onto the hyperplane, S x The vertices of the projection are still the simplex S p The vertex of the end member can be found based on z1 and z2 in the projection vector and

[0064] Specifically, the vertex component analysis method is used to unmix each heat jet mixed spectrum sample to obtain each independent pure spectrum constituting each heat jet mixed spectrum sample, which specifically includes the following steps:

[0065] S21: Calculate all thermal jet mixed spectrum samples R at different wave numbers m =[r m1 ,r m2 ,r m3 ,…,r mn ]The average brightness temperature Where m = 1, 2, ..., L, L is the number of wave number species contained in all thermal jet mixed spectrum samples, R m is the thermal jet mixing spectrum sample set at the mth wave number, r m1 ~r mn For the set R m n samples of thermal jet mixing spectra;

[0066] It should be noted that the present invention uses the M27 spectrometer to collect spectra. The EM27 spectrometer first performs bias subtraction on the spectral signal entering the instrument, and then obtains the brightness temperature according to the Planck formula. Specifically, the average brightness temperature The calculation formula is:

[0067]

[0068] Where h represents the Planck constant, h = 6.62607015 × 10 -34 , c represents the speed of light, c = 2.998 × 10 8 m / s,v m Indicates the mth wave number, in cm -1 , k represents the Boltzmann constant value, k = 1.380649 × 10 -23 J / k,L(v m ) represents the radiation flux of the unit beam corresponding to the mth wave number, and ln{·} represents the logarithm.

[0069] A total of 295 spectral data samples of the hot jets of two types of aero - engines were collected in this invention. Among them, there are 151 pure spectral samples of the hot jets of Type - I engines, 41 pure spectral samples of the hot jets of Type - II engines, and 103 mixed spectral samples of the hot jets of the two types of engines.

[0070] S22: Remove the DC components of all the mixed spectral samples of the hot jets at different wavenumbers respectively:

[0071]

[0072] Among them, R m0 is the set of mixed spectral samples of the hot jets at the m - th wavenumber after removing the DC component;

[0073] That is to say, in step S22, the data at each wavenumber is subtracted by the mean value corresponding to that wavenumber to remove the DC component in the data, obtaining zero - mean data; the purpose of doing this is to reduce the errors that may be caused by data offset or background light, etc., so that the model of this invention only focuses on the inherent change characteristics of the data.

[0074] S23: Project R m and R m0 at different wavenumbers into a two - dimensional subspace respectively:

[0075] r mp = U d ′ * R m0

[0076] r m = U d ′ * R m

[0077] Among them, U d ′ is the transposed two - dimensional projection matrix, and the two - dimensional projection matrix U d is calculated through singular - value SVD decomposition. r mp is the set of mixed spectral samples of the hot jets at the m - th wavenumber after removing the DC component in the two - dimensional subspace, and r m is the set of mixed spectral samples of the hot jets at the m - th wavenumber in the - dimensional subspace;

[0078] S24: Obtain the signal - to - noise ratio SNR m corresponding to different wavenumbers:

[0079]

[0080] Among them, SNR m is the signal - to - noise ratio corresponding to the mixed spectral samples of the hot jets at the m - th wavenumber, and P rm is the mean vector related to r mp ; P ymis the mean vector related to R m ; and there is:

[0081]

[0082] where r pFmi represents flattening the i-th hot jet mixed spectral sample in the set r mp into a one-dimensional vector; represents the transposed R Fmi represents flattening the i-th hot jet mixed spectral sample in the set R m into a one-dimensional vector;

[0083] S25: Respectively judge whether the signal-to-noise ratio SNR m corresponding to each wavenumber is greater than the set signal-to-noise ratio threshold SNR th . Project the set r m corresponding to the wavenumbers with the judgment result being yes onto a two-dimensional subspace, and project the set r m corresponding to the wavenumbers with the judgment result being no onto a one-dimensional subspace;

[0084] It should be noted that if SNR m ≤SNR th , it indicates that the signal-to-noise ratio is low, then project the set r m onto a one-dimensional subspace to filter noise; if SNT m >SNR th , it indicates that the signal-to-noise ratio is high, then project the set r m onto a two-dimensional subspace to retain the complete spectral features; meanwhile, SNR th = 15 + 10 * log10(2);

[0085] S26: Perform normalization processing on all sets of hot jet mixed spectral samples projected onto a two-dimensional subspace or a one-dimensional subspace;

[0086] S27: Respectively obtain the projection vectors of all sets of hot jet mixed spectral samples after normalization processing:

[0087]

[0088] where P0 is the initial projection matrix, and P0 can be set as the identity matrix I, is the i-th hot jet mixed spectral sample in the set of hot jet mixed spectral samples at the m-th wavenumber after normalization processing; Z mi is corresponding projection vector;

[0089] S28: Respectively obtain all projection vectors Z miEuclidean norm, and take the one corresponding to the maximum Euclidean norm as the estimated value of the first independent pure spectrum

[0090] That is to say, in step S28, among the projection results of all samples, it is necessary to find the vector corresponding to the most "prominent" sample in this projection direction, that is, the Euclidean norm is used as the measurement criterion for selection; Z mi The calculation method of the Euclidean norm is as follows:

[0091]

[0092] S29: In order to project all data vectors onto the direction orthogonal to the first endmember that has been found Let And update the initial projection matrix P0 according to U1:

[0093]

[0094] where U1 is the set of independent pure spectra, I is the identity matrix, T represents the transpose, and P1 is the updated projection matrix;

[0095] S210: Use the updated projection matrix P1 and the remaining after removing the corresponding hot jet mixed spectrum samples to re - execute steps S27 - S28 to determine the estimated value of the second independent pure spectrum And so on until all independent pure spectra are determined.

[0096] It should be noted that for the two types of spectra corresponding to the two types of engines in this embodiment, finding the estimated value and the estimated value means that all independent pure spectra are determined.

[0097] Furthermore, after all independent pure spectra are determined, the mixing ratio of each independent pure spectrum at each wavenumber is calculated by the least - squares method as follows:

[0098]

[0099] where A is the mixing ratio matrix of each independent pure spectrum at each wavenumber, M is the matrix composed of the estimated values of each independent pure spectrum, is the matrix composed of the normalized hot jet mixed spectrum samples corresponding to the estimated values of each independent pure spectrum.

[0100] Furthermore, as the network architecture deepens, both the number of channels and redundant information during the training process increase. Inspired by the human visual sense, the attention mechanism is utilized to optimize features, thereby improving the efficiency of network training. The present invention proposes a one-dimensional deep learning MHSA-CNN algorithm for feature extraction and visualization analysis of the pure spectra of two types of aero-engine hot jets after unmixing. The structure diagram of the one-dimensional convolutional neural network model MHSA-CNN constructed by the present invention is as shown in Figure 3 shown. This model includes three convolutional layers of different scales, a multi-head self-attention mechanism module, a max pooling layer, a flattening operation, and a fully connected layer, which perform feature extraction on the pure spectra extracted by two VCAs. The specific network structure is shown in the following table. In order for the network to extract multi-scale and local feature information of two different types of aero-engine hot jets, convolutional kernels of different sizes are used in the three convolutional layers. The specific parameters of the convolutional layers are shown in Table 1. During the operation of the deep learning network, a large number of parameters will be generated. In order to retain key features and reduce redundancy, the convolutional kernel size of the max pooling layer is set to 2, halving the feature length, so that the convolution and pooling operations compress the feature length.

[0101] Table 1 The architecture of CNN

[0102]

[0103] The present invention believes that the largest peak in the spectral information is the feature peak. The local peaks extracted by the convolutional neural network CNN are input into the attention mechanism and divided into three parts: query (Q), key (K), and value (V). The query represents the feature peak that the current network is concerned about, the key represents all the peaks of the input spectral data, and the value represents the actual feature peak. First, the similarity between the query and the key is calculated through a dot product operation, that is, the correlation between the two is judged. The higher the similarity, the more relevant the currently concerned feature peak is. Then, the similarity is converted into probability weights through the softmax function, and then these weights are used to perform weighted summation on all the feature peaks. The process of extracting feature peaks by the attention mechanism is as shown in Figure 4 shown.

[0104] The core formula of the self-attention mechanism is as follows:

[0105]

[0106] where Q, K, and V represent the query (Query), key (Key), and value (Value) matrices respectively, and d k is the dimension of the key vector, which is used to scale the dot product to stabilize the softmax function.

[0107] In the present invention, the number of heads of the multi-head self-attention mechanism is set to 4. By performing different linear transformations on the input one-dimensional spectral data, 4 attention heads are obtained. These 4 attention heads concurrently focus on the peaks in different bands, enabling the model to capture the peak information in the spectral data from multiple perspectives.

[0108] head i =Attention(QW i Q ,KW i K ,VW i V )

[0109] where W i Q ,W i K ,W i V are the transformation matrices for the query, key, and value of the i-th head respectively.

[0110] The outputs of all heads are concatenated and then, through a final linear transformation and layer normalization, the output of the multi-head attention is obtained. This step integrates the feature information extracted from different subspaces, enhances the expressive power of the model, and finally extracts the most critical feature peaks in the spectrum.

[0111] Multihead(Q,K,V)=Concat(head1,…head4)W o

[0112] where W o is the final output transformation matrix.

[0113] As Figure 5 and Figure 6 shown, the extracted feature range is at 2282 - 2283 cm -1 as well as at 2388 - 2389 cm -1 bands. The peak at 2282 - 2283 cm -1 mainly represents functional groups related to cyano (-C≡N) or certain specific isocyanates (-N=C=O). The characteristic absorption peaks in this range are mainly caused by the stretching vibration of triple bonds (C≡N or N=C=O); the range of 2388 - 2389 cm -1 mainly represents CO2.

[0114] In summary, the present invention provides an engine identification method based on the mixed spectral characteristics of aeroengine hot jets. First, a Fourier transform infrared spectrometer is used to accurately measure the hot jets of different models of aeroengines. The collected data includes the spectra of the hot jets generated separately by each model of engine and the spectral information of the mixed hot jets of each type of engine. Then, the VCA-LS algorithm is used to unmix the mixed spectra of the hot jets of each type of aeroengine. Through VCA, the pure spectra of the hot jets of each type of engine are successfully obtained, and the proportion of each pure spectrum in the mixed spectrum is deduced, laying a solid foundation for subsequent in-depth analysis. Finally, the present invention uses the one-dimensional convolutional deep learning MHSA-CNN algorithm to extract features from the unmixed hot jets of different types of aeroengines, so as to more accurately grasp their internal characteristics and differences, providing strong data support and technical guarantee for the fault diagnosis of aeroengines.

[0115] Of course, the present invention may 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. However, these corresponding changes and deformations should all fall within the protection scope of the appended claims of the present invention.

Claims

1. An engine identification method based on the mixed spectral characteristics of the hot jet of an aeroengine, characterized in that, It includes the following steps: S1: Collect multiple hot jet mixing spectral samples of the aircraft engine to be measured using a Fourier transform infrared spectrometer; S2: Perform unmixing processing on each hot jet mixing spectral sample using vertex component analysis to obtain each independent pure spectrum that composes each hot jet mixing spectral sample; S3: Input each independent pure spectrum into the MHSA-CNN network model respectively to obtain the types of each aircraft engine to be measured corresponding to each independent pure spectrum.

2. The engine identification method based on the mixed spectral characteristics of the hot jet of an aeroengine according to claim 1, wherein Performing unmixing processing on each hot jet mixing spectral sample using vertex component analysis to obtain each independent pure spectrum that composes each hot jet mixing spectral sample specifically includes the following steps: S21: Calculate the average brightness temperature of all hot jet mixing spectral samples R at different wavenumbers m = [r m1 , r m2 , r m3 , …, r mn where m = 1, 2, …, L, L is the number of wavenumber types included in all hot jet mixing spectral samples, R m is the set of hot jet mixing spectral samples at the m-th wavenumber, and r m1 ~r mn are n hot jet mixing spectral samples in the set R m ; S22: Remove the DC components of all hot jet mixing spectral samples at different wavenumbers respectively; Among them, R m0 is the set of thermal jet mixing spectral samples at the m-th wavenumber after removing the DC component; S23: Project R at different wavenumbers m and R m0 onto a two-dimensional subspace respectively: r mp = U d ′ * R m0 r m = U d ′ * R m Among them, U d ′ is the transposed two-dimensional projection matrix, r mp is the set of thermal jet mixing spectral samples at the m-th wavenumber after removing the DC component in the two-dimensional subspace, r m is the set of thermal jet mixing spectral samples at the m-th wavenumber in the -dimensional subspace; S24: Obtain the signal-to-noise ratio SNR corresponding to different wave numbers m : where, SNR m is the signal-to-noise ratio corresponding to the thermal jet mixing spectrum sample at the m-th wave number, P rm is the mean vector related to r mp ; P ym is the mean vector related to R m ; and there is: Among them, means flattening the i-th hot jet mixing spectrum sample in the set r mp into a one-dimensional vector; means the transposed R Fmi means flattening the i-th hot jet mixing spectrum sample in the set R m into a one-dimensional vector; S25: Determine the signal-to-noise ratio SNR corresponding to each wavenumber m whether it is greater than the set signal-to-noise ratio threshold SNR th , and project the set r of wavenumbers corresponding to the judgment result being yes m onto a two-dimensional subspace, and project the set r of wavenumbers corresponding to the judgment result being no m onto a one-dimensional subspace; S26: Normalize the set of hot jet mixing spectral samples projected onto a two-dimensional subspace or a one-dimensional subspace; S27: Obtain the projection vectors of all sets of normalized hot jet mixing spectral samples respectively; where P0 is the initial projection matrix, is the i-th hot jet mixing spectrum sample in the set of hot jet mixing spectrum samples at the m-th wavenumber after normalization; Z mi is the corresponding projection vector; S28: Obtain all projection vectors Z respectively mi for their Euclidean norms, and use the one corresponding to the maximum Euclidean norm as the estimated value of the first independent pure spectrum S29: Let and update the initial projection matrix P0 according to U1: where, U1 is the set of independent pure spectra, I is the identity matrix, T represents transpose, and P1 is the updated projection matrix; S210: Use the updated projection matrix P1 and the remaining after removing the corresponding hot jet mixed spectral samples to re - execute steps S27 - S28 to determine the estimated value of the second independent pure spectrum And so on until all independent pure spectra are determined.

3. The engine identification method based on the mixed spectral characteristics of the hot jet of an aeroengine as claimed in claim 2, wherein The average of the brightness temperatures in step S21 is calculated as follows: where h represents the Planck constant value, c represents the speed of light, v m represents the m-th wave number, k represents the Boltzmann constant value, L(v m ) represents the radiant flux of the unit beam corresponding to the m-th wave number, and ln{·} represents taking the logarithm.

4. The engine identification method based on the mixed spectral characteristics of the hot jet of an aeroengine as claimed in claim 2, wherein After determining all the independent pure spectra, the mixing ratios occupied by each independent pure spectrum at each wavenumber are as follows: Among them, A is the mixing ratio matrix of each independent pure spectrum in each wavenumber, and M is the matrix composed of the estimated values of each independent pure spectrum. is the matrix composed of the normalized hot jet mixing spectrum samples corresponding to the estimated values of each independent pure spectrum.

5. The engine recognition method based on the mixed spectral characteristics of the hot jet of an aeroengine according to claim 1, wherein The MHSA-CNN network model includes three convolutional layers with different scales, a multi-head self-attention mechanism module, a max pooling layer, a flattening operation, and a fully connected layer.

Images