Multi-dimensional feature extraction method for thermal jet FTIR (Fourier Transform Infrared Spectrometry) telemetering spectrum of aero-engine

Abstract

The invention provides an aero-engine thermal jet FTIR telemetry spectrum multi-dimensional feature extraction method, which comprises the following steps: firstly, carrying out precise external field measurement on two types of aero-engine thermal jets by using a Fourier infrared spectrometer to obtain thermal jet spectrum data independently generated by each type of engine; secondly, a basis is provided for thermal jet characteristic analysis of different types of engines by establishing a four-stage processing framework of coarse detection, local optimization, dynamic screening and intelligent combination; and finally, combining the statistical kurtosis and skewness, systematically integrating the basic attribute and distribution rule of the spectral data, and carrying out structured and high-dimensional feature extraction with comprehensive characterization capability on the thermal jet data of two types of different models of aero-engines, thereby laying a solid data foundation for researches such as subsequent thermal jet characteristic comparison, fault diagnosis and the like.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing spectrum processing, and in particular relates to a multi-dimensional feature extraction method for FTIR remote sensing spectrum of aero-engine thermal jet. Background Art

[0002] In recent years, with the rapid development of artificial intelligence (AI), deep learning, with its powerful feature extraction and pattern recognition capabilities, has gradually become a cutting-edge technology for fault diagnosis in this field. Seiari et al. proposed a method for fault diagnosis of aircraft engine actuators using a model-observer combination. They designed a Luneberger model observer that can detect actuator faults using observer residuals. However, this method relies heavily on the accuracy of the established aircraft engine model. Furthermore, it can only detect relatively obvious actuator faults. For complex, multi-factorial faults, or those with relatively mild fault severity, relying solely on observer residuals may not be sensitive or accurate enough, making it difficult to accurately diagnose the specific type and severity of the fault. Chen et al. proposed an improved SUKF algorithm and applied it to engine performance degradation estimation. This algorithm effectively improved the accuracy of the airborne adaptive model's estimation of the engine's true health parameters and its performance parameter degradation. However, while their improved SUKF algorithm improved estimation accuracy, it generally increased computational complexity. In an airborne environment, computing resources are often limited, which can affect the algorithm's real-time performance and prevent rapid estimation of engine performance degradation, hindering timely maintenance measures. Fan et al. designed an adaptive law based on interval observer theory that effectively compresses the error interval of output estimation and improves sensitivity to sudden faults. A flexible event-triggered fault detection mechanism was designed based on the adaptive interval observer (AIO). While the adaptive law based on interval observer theory can compress the error interval of output estimation, its design can be complex and may require extensive parameter adjustments and optimization for different engine models and operating conditions.

[0003] Feature extraction can reduce dimensionality, remove noise, enhance data interpretability and model performance, and uncover hidden information in the data, making it a crucial step in data analysis. Yang et al., incorporating adaptive spectral pattern extraction (ASME) theory, proposed a fast Fourier transform (FFG) method for quantitatively assessing bearing damage. They also established an early fault identification mechanism to determine the optimal time for the first prediction. Gao et al. proposed an autocorrelation multi-head attention transformer (AMTrans) algorithm for infrared spectral sequence deconvolution. They utilized an attention mechanism for feature extraction and an autocorrelation function for attention computation. The autocorrelation attention model exploits the inherent sequential nature of spectral data and effectively recovers spectra by capturing autocorrelation patterns in the sequence. This model is trained using supervised learning and demonstrates promising results in infrared spectral recovery. Sun et al. proposed a HAD background reconstruction method based on contrastive self-supervised learning. By constructing a self-supervised pre-trained model based on a pixel-level masking strategy and a dual attention network (DAN) encoder, the pre-trained model can learn general background representations without generating labeled samples. The DAN encoder consists of a visual Transformer and a channel attention module to extract global context information and correlation between spectra.

[0004] During feature extraction using deep learning, some original information may be compressed, filtered, or lost, making it difficult to recover data details. Furthermore, redundant features may be extracted, increasing model complexity and reducing training efficiency. This can easily lead to overfitting, resulting in a model that performs well on training data but generalizes poorly to new data, preventing accurate feature extraction. In spectral data analysis, spectral peaks, as key characteristic parameters, can intuitively reflect the advantages of a substance's absorption or emission properties. They have become one of the most representative and widely used features, often used in important research and practical scenarios such as component identification, concentration determination, and structural analysis. Kurtosis and skewness are two important statistics that describe data distribution characteristics. Kurtosis describes the peak shape of a data distribution, i.e., the degree of concentration of the data distribution around the mean. It is often used to measure the steepness or flatness of a distribution compared to a normal distribution. Skewness is a statistic used to describe the symmetry of a data distribution. It measures the degree to which the data distribution deviates from a symmetric distribution (such as a normal distribution). Cluster analysis can uncover hidden natural grouping structures in data and reveal the inherent connections and similarities between data points.

[0005] As the heart of an aircraft, the reliability of an aircraft engine plays a decisive role in flight safety. With the continuous innovation and development of aviation technology, engine design and manufacturing processes are becoming increasingly sophisticated and complex, and the corresponding fault diagnosis technology is also keeping pace with the times. Currently, aircraft engine fault diagnosis mainly uses the following technical means:

[0006] Manual inspection technology: This technology builds fault diagnosis models based on the expertise and practical experience of domain experts, providing technicians with a basis for fault diagnosis. However, because the diagnostic process relies heavily on expert experience, it inevitably suffers from a high degree of subjectivity, which affects the objectivity and accuracy of the diagnosis.

[0007] Sensor technology: By deploying sensors in key engine locations, key parameters such as temperature, pressure, and vibration can be monitored in real time, allowing for the rapid detection of anomalies. However, this technology faces numerous challenges. High-quality sensors are not only expensive to purchase but also require substantial ongoing maintenance. Furthermore, sensors generate massive amounts of abnormal data during operation, requiring the use of complex signal processing algorithms for effective analysis.

[0008] Remote Monitoring Technology: Leveraging satellite communications and ground stations, remote monitoring technology enables real-time monitoring and remote diagnosis of aircraft engines worldwide. However, this technology presents data transmission security risks and a strong reliance on satellite communications systems. Furthermore, the remote monitoring system itself is technically complex, and its operation, maintenance, and upgrades present significant challenges.

[0009] Traditional data feature extraction methods often only focus on basic statistics of the data, such as mean and variance, while ignoring the high-order features of the data. Summary of the Invention

[0010] In order to achieve in-depth mining of high-order features of aircraft engine heat jet FTIR data, the present invention provides a multidimensional feature extraction method for aircraft engine heat jet FTIR telemetry spectra, focusing on peak position features, intensity features corresponding to peak position features, peak kurtosis features, and peak skewness features, which can reveal the characteristics of heat jets more comprehensively and accurately.

[0011] A multidimensional feature extraction method for FTIR telemetry spectra of aircraft engine thermal jets, wherein the multidimensional features of each telemetry spectrum used to identify the aircraft engine type include a peak position feature P, an intensity feature V corresponding to the peak position feature P, a peak kurtosis feature k, and a peak skewness feature γ;

[0012] The method for obtaining the peak position feature P and the corresponding intensity feature V of any remote sensing spectrum sample is as follows:

[0013] S1: determine in sequence whether the intensity corresponding to each wavenumber of the current telemetry spectrum sample is greater than the intensity corresponding to the wavenumbers of its two left and right neighbors. If so, the wavenumber is the candidate peak wavenumber;

[0014] S2: determining the expansion range of each candidate peak wave number with each candidate peak wave number as the center, and taking the wave number corresponding to the maximum intensity in each expansion range as the preferred peak wave number;

[0015] S3: Select the top N with the largest intensity from all the preferred peak wave numbers as the initial peak position feature P;

[0016] S4: Sort the wavenumbers in the initial peak position feature P in ascending order, and determine whether the absolute value of the difference between adjacent wavenumbers is not greater than the set threshold. Merge the two wavenumbers with a yes judgment result into one wavenumber, and keep the two wavenumbers with a no judgment result unchanged to obtain the final peak position feature P and the corresponding intensity feature V.

[0017] Furthermore, the method for merging two wave numbers with a judgment result of yes into one wave number is:

[0018] The average value of the two wavenumbers is taken as the new wavenumber after merging, and the maximum value of the intensity values ​​corresponding to the two wavenumbers is taken as the intensity value corresponding to the new wavenumber.

[0019] Furthermore, the FTIR telemetry spectra of the aircraft engine heat jet were obtained by conducting field measurements of the heat jets of two types of aircraft engines using a Fourier transform infrared spectrometer.

[0020] Furthermore, the GMM model is used to cluster the multidimensional features of each telemetry spectrum to obtain the multidimensional feature categories corresponding to each type of aircraft engine.

[0021] Furthermore, the peak kurtosis feature is used to reflect the concentration degree of the spectral absorption characteristic peak of a specific component in the heat jet. If the peak kurtosis value is greater than the peak kurtosis of the normal distribution, it indicates that the spectral absorption characteristic peak of the specific component in the heat jet is concentratedly distributed; if the peak kurtosis value is not greater than the peak kurtosis of the normal distribution, it indicates that the composition of the heat jet is complex and contains a variety of different substances.

[0022] Furthermore, the peak skewness feature is used to reflect the direction in which specific substances in the thermal jet affect the spectrum. When the peak skewness value is positive, the peak is right-skewed, indicating that the influence of specific substances in the thermal jet on the spectrum tends to be in the high-value direction; when the peak skewness value is negative, the peak is left-skewed, indicating that the influence of specific substances in the thermal jet on the spectrum tends to be in the low-value direction; when the peak skewness value is close to 0, the peaks are approximately symmetrically distributed, indicating that the intensity distribution on both sides of the peak is relatively balanced.

[0023] Beneficial effects:

[0024] The present invention provides a method for extracting multidimensional features from FTIR telemetry spectra of aero-engine hot jets. First, a Fourier transform infrared spectrometer is used to perform precise field measurements of the hot jets of two types of aero-engines to obtain the hot jet spectral data independently generated by each type of engine. Secondly, a four-level processing architecture of "coarse detection-local optimization-dynamic screening-intelligent merging" is established to provide a basis for analyzing the characteristics of hot jets of different types of engines. Finally, the basic properties and distribution laws of the spectral data are systematically integrated by combining the statistical quantities of kurtosis and skewness, and structured, high-dimensional features with comprehensive characterization capabilities are extracted from the hot jet data of two different types of aero-engines, laying a solid data foundation for subsequent research on hot jet characteristics comparison, fault diagnosis, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flowchart of a method for extracting multidimensional features from FTIR telemetry spectra of aero-engine thermal jets provided by the present invention;

[0026] Figure 2 The peak extraction result diagram provided by the present invention;

[0027] Figure 3 This is the clustering result diagram provided by the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0029] The present invention generally comprises four parts: aircraft engine thermal jet spectrum data acquisition, data preprocessing, multi-dimensional feature extraction and clustering.

[0030] The technical solution is detailed as follows:

[0031] 1. Data Collection

[0032] The present invention uses the field experiment method to collect the thermal jet data of the aircraft engine. Figure 1 The measurement distance between the spectrometer and the aircraft engine ranges from 127 to 280 meters. The outdoor temperature during measurement is 20-25°C, and the humidity is 40% to 73% RH. The measurement instrument used is the EM27 Fourier transform infrared spectrometer, operating in passive mode. The spectral resolution is 1 cm⁻¹, the spectral measurement range is 2.5-12 μm, and the full viewing angle can reach 30 mrad. The EM27 spectrometer first performs offset subtraction on the spectral signal entering the instrument and calculates the brightness temperature according to the Planck equation.

[0033]

[0034] Where h represents the Planck constant, h = 6.62607015 × 10 -34 J·s. c represents the speed of light, c = 2.998 × 108 m / s. v represents the wave number, the unit is cm -1 k represents the Boltzmann constant value, k = 1.380649 × 10 -23 J / k. L(v) represents the radiation flux per unit beam.

[0035] The present invention collected a total of 295 spectral data samples of the thermal jets of two types of aircraft engines, including 151 pure spectral samples of the thermal jet of the Type I engine, 41 pure spectral samples of the thermal jet of the Type II engine, and 103 mixed spectral samples of the thermal jets of the two types of engines.

[0036] This paper constructs a multidimensional feature vector F for the collected FTIR data of aero-engine heat jets by organically combining low-order peak position and intensity characteristics with statistical features. This provides structured and comprehensive features for subsequent analysis of two different types of aero-engine heat jet data. The specific form of the feature vector F is as follows:

[0037] F=[P,V,k,γ] T

[0038] Among them, P represents the position feature of the peak, V represents the intensity feature corresponding to P, k represents the kurtosis feature, and γ represents the skewness feature.

[0039] In response to the demand for high-precision detection of aircraft engine thermal jet characteristics, the present invention provides a method for extracting multidimensional features from FTIR telemetry spectra of aircraft engine thermal jets. This method establishes a four-level processing architecture of "coarse detection-local optimization-dynamic screening-intelligent merging" to extract multidimensional features of each telemetry spectrum for identifying the type of aircraft engine, including the peak position feature P, the intensity feature V corresponding to the peak position feature P, the peak kurtosis feature k, and the peak skewness feature γ, thereby providing a basis for analyzing the thermal jet characteristics of different types of engines.

[0040] Among them, such as Figure 1 As shown, the method for obtaining the peak position feature P and the corresponding intensity feature V of any remote sensing spectrum sample is:

[0041] S1: determine in sequence whether the intensity corresponding to each wavenumber of the current telemetry spectrum sample is greater than the intensity corresponding to the wavenumbers of its two left and right neighbors. If so, the wavenumber is the candidate peak wavenumber;

[0042] Specifically, assume that each row of spectral data represents a sample, and assume that the spectral signal of each sample is recorded as s = [s1, s2, ..., s n ], n is the wave number, the peak is preliminarily detected, and the method for obtaining the candidate peak wave number is:

[0043] Filter out the peak positions that meet the following conditions as candidate peak wave numbers. Assuming that there are m wave numbers that meet the conditions, the intensity corresponding to each candidate peak position is V = {s p1 ,s p2 ,…,s pm}

[0044] p i ={p||s pi >s pi±1}

[0045] Where p = 1, 2,…, n, i = 1, 2,…, m.

[0046] S2: determining the expansion range of each candidate peak wave number with each candidate peak wave number as the center, and taking the wave number corresponding to the maximum intensity in each expansion range as the preferred peak wave number;

[0047] Specifically, for each peak value p detected initially i , the present invention will search for the true maximum point within a specific range centered on the peak value. This specific range is determined by two parameters, Δl and Δr, where Δl represents the range to the left, corresponding to the left_range in the code; Δr represents the range to the right, corresponding to the right_range in the code. The present invention will search for the true maximum point within the interval [p i -Δl,p i +Δr] to find the maximum value. But it should be noted that this interval cannot exceed the range of the spectral data. Therefore, the actual search interval will be limited to [max(0,p i -Δl),min(n,p i +Δr+1)], where n represents the length of the spectral data.

[0048]

[0049] Among them, s j Indicates the intensity value of the jth point in the spectral data. The argmax() function is used to find the value of the independent variable that makes the function reach the maximum value. In this formula, it will traverse the interval [max(0,p i -Δl),min(n,p i +Δr+1)], calculate the corresponding s j , then return so that s j The largest j value is the optimized peak position p′ i .

[0050] S3: Select the top N with the largest intensity from all the preferred peak wave numbers as the initial peak position feature P;

[0051] It should be noted that for each peak p′ after optimization i , the corresponding spectral intensity value is s p′ Here s represents the spectral data sequence, through the index p′ i The intensity value at the corresponding peak can be obtained. The intensity index of the optimized peak is sorted using the argsort() function. The function returns an index sequence arranged according to the intensity. Since the present invention needs to filter out the peak with the largest intensity, the intensity needs to be arranged in descending order, that is, the index is reorganized in order from large to small intensity. The "↓" symbol in the formula clearly indicates a descending operation. From the index sequence after descending order, the first N indexes are selected, and the peaks corresponding to these indexes are the N peaks with the strongest intensity. Among them, the parameter N corresponds to num_largest_peaks in the code.

[0052] p″ i =argsort(s p′ )↓[:N]

[0053] Among them, p″ i Indicates the index set of the strongest peaks finally filtered out. argsort(s p′ ) corresponds to the intensity s of the optimized peak (indexed as p′) p′ Sort the indexes. ↓ indicates that the sorted index sequence is sorted in descending order. [:N] indicates that the first N indexes are intercepted from the descending index sequence to complete the screening of the strongest peaks.

[0054] S4: Sort the wavenumbers in the initial peak position feature P in ascending order, and determine whether the absolute value of the difference between adjacent wavenumbers is not greater than the set threshold. Merge the two wavenumbers with a yes judgment result into one wavenumber, and keep the two wavenumbers with a no judgment result unchanged to obtain the final peak position feature P and the corresponding intensity feature V.

[0055] The method for merging two wavenumbers with a judgment result into one wavenumber is: taking the average value of the two wavenumbers as the new wavenumber after the merger, and taking the maximum value of the intensity values ​​corresponding to the two wavenumbers as the intensity value corresponding to the new wavenumber.

[0056] That is to say, according to the previous three steps, the position and intensity of the peak (i.e., brightness temperature) have been obtained:

[0057] pos=W p″

[0058] int=s p″

[0059] Among them, W p″ Indicates p″ i , s p″ Indicates intensity;

[0060] For two adjacent peak points (pos a ,int a ) and (pos b ,int b ), where pos represents the position of the peak (such as wave number), and int represents the intensity of the peak. The present invention compares the absolute value of the difference between the two peak positions |pos a -pos b | and a pre-set threshold △m (corresponding to the merge_range parameter in the code) to decide whether to merge. If |pos a -pos b |≤△m, indicating that the distance between the two adjacent peaks is close enough. The present invention believes that they represent the same physical feature and need to be merged (if they are not close, they are not merged). The new peak point after merging is That is, the new peak position is the average of the two original peak positions, and the new peak intensity is the maximum of the two original peak intensities. The reason for this is that taking the position average can integrate the position information of the two peaks, while taking the maximum intensity can retain the most significant signal features. a -pos b |>△m, indicating that the two adjacent peaks are far apart and represent different physical features, so they are not merged.

[0061]

[0062] in, Indicates a merge.

[0063] The results of peak extraction are shown in the figure Figure 2 shown.

[0064] Furthermore, the following details the method for extracting peak kurtosis and skewness features.

[0065] Kurtosis is a statistic used to measure the sharpness of the data distribution relative to the normal distribution. In the FTIR data of aircraft engine hot jets, the peak kurtosis can reflect the concentration of the spectral absorption characteristic peaks of specific components in the hot jet. Specifically, if the kurtosis value is greater than the kurtosis of the normal distribution (usually 3), it means that the peak is sharper, which means that the spectral absorption characteristic peaks of specific components in the hot jet are more concentrated; if the kurtosis value is less than the kurtosis of the normal distribution, it indicates that the peak is flatter and the data distribution is more dispersed, that is, the range of variation of the data values ​​near and on both sides of the peak is wider, which suggests that the composition of the hot jet is complex and contains more substances.

[0066]

[0067] Among them, μ4 represents the fourth-order center distance; μ2 represents the second-order center distance; n represents the number of data points, s i represents the value of the i-th data point, Represents the mean of the data. "-3" is used to set the kurtosis of the normal distribution to 0, which allows for a more intuitive comparison of the differences between the kurtosis of different data distributions and the normal distribution.

[0068] Skewness is used to describe the asymmetry of data distribution. In aircraft engine hot jet FTIR data, peak skewness can reflect the direction in which certain special substances in the hot jet affect the spectrum. When the skewness value is positive, the peak is right-skewed, that is, the long tail on the right side of the peak (towards the larger value) is longer, indicating that the influence of certain special substances in the hot jet on the spectrum is biased towards high values. When the skewness value is negative, the peak is left-skewed, and the long tail on the left side of the peak (towards the smaller value) is longer, indicating that the influence of certain special substances in the hot jet on the spectrum is biased towards low values. When the skewness value is close to 0, the peak is approximately symmetrically distributed, indicating that the data distribution on both sides of the peak is relatively balanced.

[0069]

[0070] Among them, μ3 represents the third-order center distance; μ2 represents the second-order center distance.

[0071] So far, the present invention has obtained F = [P, V, k, γ] corresponding to 295 spectral samples. T , the following 295 F=[P,V,k,γ] T Perform feature clustering.

[0072] Specifically, the present invention has extracted the key features of peak position, intensity, kurtosis and skewness from the FTIR spectrum data of the aircraft engine wake to construct a multidimensional feature vector. In order to verify the validity of the feature vector, the present invention uses an unsupervised clustering algorithm. These features of each sample are combined into a vector to construct a multidimensional feature matrix F = {f1, f2, ..., f N}, where N represents the number of samples (295), f N It is the multidimensional feature vector of the Nth sample. The extracted multidimensional feature matrix F={f1,f2,…,f N}Input into the GMM model for clustering.

[0073] The GMM model uses the EM algorithm to iteratively estimate the parameters (mean, covariance, and weight) of each Gaussian distribution to maximize the likelihood function of the data. After the model is trained, cluster predictions are performed on new data points. Each aircraft engine wake sample is assigned to the cluster with the highest probability based on its probability of belonging to each Gaussian distribution.

[0074] In the present invention, the probability distribution model of the multidimensional feature matrix F can be expressed as:

[0075]

[0076] Where K represents the number of Gaussian distributions, that is, the number of clusters. k Represents the coefficient of the kth Gaussian distribution, satisfying Indicates the contribution ratio of the kth Gaussian distribution in the mixture model. φ(f|θ k ) is the probability density function of the kth Gaussian distribution, where u k is the mean of the kth Gaussian distribution, σ k is the standard deviation, θ k =(u k ,σ k ) represents the parameters of the k-th Gaussian distribution.

[0077] Assume that the multidimensional feature matrix has a mean of u k , with a standard deviation of σ k Gaussian distribution, the implicit parameters of each spectrum in the feature matrix can be solved according to the expected maximum algorithm, that is, assuming that the multidimensional feature matrix f1,f2,…,f N , the corresponding parameters are θ1,θ2,θ3,····θ N , the Gaussian model parameters θ can be estimated by the EM algorithm. First, initialize the model parameters, and calculate the multidimensional feature matrix f from the current model parameters θ k Probability from the kth sub-model:

[0078]

[0079] Among them, γ jk Represents sample f j The probability of being in the kth cluster.

[0080] Then the model parameters are updated and the iterative calculation is repeated many times, and then the clustering results of the multidimensional feature moments are obtained according to the results of each parameter.

[0081]

[0082] in, represents the effective number of samples in the kth cluster; α k It represents the proportion of the k-th cluster in all samples.

[0083] The clustering results are shown in the figure Figure 2 shown.

[0084] In summary, the present invention provides a method for extracting multidimensional features from FTIR telemetry spectra of aero-engine hot jets. First, a Fourier transform infrared spectrometer is used to perform precise field measurements of the hot jets of two types of aero-engines to obtain the hot jet spectral data independently generated by each type of engine. Secondly, a four-level processing architecture of "coarse detection-local optimization-dynamic screening-intelligent merging" is established to provide a basis for the analysis of hot jet characteristics of different types of engines. Finally, combined with the statistical quantities of kurtosis and skewness, the basic properties and distribution laws of the spectral data are systematically integrated to perform structured, high-dimensional and comprehensive characterization feature extraction on the hot jet data of two different types of aero-engines, laying a solid data foundation for subsequent research on hot jet characteristics comparison, fault diagnosis, etc.

[0085] 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 may of course make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for extracting multidimensional features from FTIR telemetry spectra of aero-engine thermal jets, characterized in that: The multidimensional features of each telemetry spectrum used to identify the type of aircraft engine include peak position feature P, intensity feature V corresponding to the peak position feature P, peak kurtosis feature k, and peak skewness feature γ; The method for obtaining the peak position feature P and the corresponding intensity feature V of any remote sensing spectrum sample is as follows: S1: determine in sequence whether the intensity corresponding to each wavenumber of the current telemetry spectrum sample is greater than the intensity corresponding to the wavenumbers of its two left and right neighbors. If so, the wavenumber is the candidate peak wavenumber; S2: determining the expansion range of each candidate peak wave number with each candidate peak wave number as the center, and taking the wave number corresponding to the maximum intensity in each expansion range as the preferred peak wave number; S3: Select the top N with the largest intensity from all the preferred peak wave numbers as the initial peak position feature P; S4: Sort the wavenumbers in the initial peak position feature P in ascending order, and determine whether the absolute value of the difference between adjacent wavenumbers is not greater than the set threshold. Merge the two wavenumbers with a yes judgment result into one wavenumber, and keep the two wavenumbers with a no judgment result unchanged to obtain the final peak position feature P and the corresponding intensity feature V.

2. The multidimensional feature extraction method for FTIR telemetry spectrum of aircraft engine thermal jet according to claim 1, characterized in that: The method to merge two wave numbers with a judgment result into one wave number is: The average value of the two wavenumbers is taken as the new wavenumber after merging, and the maximum value of the intensity values ​​corresponding to the two wavenumbers is taken as the intensity value corresponding to the new wavenumber.

3. The multidimensional feature extraction method for FTIR telemetry spectrum of aircraft engine thermal jet according to claim 1, characterized in that: The FTIR telemetry spectra of aircraft engine heat jets were obtained by field measurement of two types of aircraft engine heat jets using a Fourier transform infrared spectrometer.

4. The multidimensional feature extraction method for FTIR telemetry spectrum of aircraft engine thermal jet according to claim 1, characterized in that: The GMM model is used to cluster the multidimensional features of each telemetry spectrum to obtain the multidimensional feature categories corresponding to each type of aircraft engine.

5. The multidimensional feature extraction method for FTIR telemetry spectrum of aircraft engine thermal jet according to claim 1, characterized in that: The peak kurtosis characteristic is used to reflect the concentration degree of the spectral absorption characteristic peak of a specific component in the heat jet. If the peak kurtosis value is greater than the peak kurtosis of the normal distribution, it indicates that the spectral absorption characteristic peak of the specific component in the heat jet is concentrated. If the peak kurtosis value is not greater than the peak kurtosis of the normal distribution, it indicates that the composition of the heat jet is complex and contains a variety of different substances.

6. The multidimensional feature extraction method for aircraft engine thermal jet FTIR telemetry spectrum according to claim 1, characterized in that: The peak skewness characteristic is used to reflect the direction in which a specific substance in the heat jet affects the spectrum. When the peak skewness value is positive, the peak is right-skewed, indicating that the influence of the specific substance in the heat jet on the spectrum is biased toward high values. When the peak skewness value is negative, the peak is left-skewed, indicating that the influence of the specific substance in the heat jet on the spectrum is biased toward low values. When the peak skewness value is close to 0, the peaks are approximately symmetrically distributed, indicating that the intensity distribution on both sides of the peak is relatively balanced.

Images