Multi-component analysis method and system for aircraft exhaust smoke using micro-on-chip spectral imaging

Abstract

This invention discloses a method and system for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging. The method involves acquiring multispectral image data of the exhaust plume region using a micro-on-chip spectral imaging system. This invention unmixes the mixed spectra using a non-negative matrix factorization algorithm and, combined with a pre-built fuel component spectral library, accurately identifies and quantitatively calculates the concentration distribution of key components such as carbon dioxide, carbon monoxide, nitrogen oxides, and unburned hydrocarbons in the exhaust plume. This method overcomes the errors caused by component overlap in traditional spectral analysis. By establishing a three-dimensional spatial coordinate model, high-precision extraction of the exhaust plume region is achieved, avoiding background interference. Furthermore, convolutional neural networks are used to perform spatiotemporal correlation modeling of the multi-component concentration data, generating a three-dimensional dynamic evolution map of the exhaust plume chemical components. This map reflects the diffusion patterns of exhaust plume components over time and space in real time, providing crucial data support for environmental impact assessment, aircraft emission control, and accident emergency response.

Description

Technical Field

[0001] This invention relates to the field of spectral image processing technology, specifically to a method and system for multi-component analysis of aircraft exhaust plumes using microchip-based on-chip spectral imaging. Background Technology

[0002] Traditional aircraft exhaust analysis relies on large laboratory equipment, which suffers from poor portability, low real-time performance, and insufficient accuracy in component analysis. In existing technologies:

[0003] 1. Portable spectrometers are limited by their optical structure, making it difficult to simultaneously meet the requirements of wide-band coverage (200-2500nm) and high spectral resolution (<5nm);

[0004] 2. Most exhaust smoke dynamic diffusion models use simplified assumptions and do not incorporate real-time concentration data to correct for turbulent transport processes;

[0005] 3. Multi-component quantitative algorithms are susceptible to spectral aliasing interference, especially for low-concentration pollutants (such as nitrogen oxides) with insufficient detection sensitivity. Summary of the Invention

[0006] To address the aforementioned technical problems, a method and system for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging are provided. This technical solution resolves the issues raised in the background section.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] In a first aspect of the invention, a method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging is provided, comprising:

[0009] Multispectral image data of the aircraft's exhaust region is acquired through a micro on-chip spectral imaging system. The multispectral image data is a three-dimensional spectral data cube. The system integrates a tunable filter module and an array detector.

[0010] Spatial alignment and radiometric correction are performed on the acquired multispectral images to eliminate image distortion caused by aircraft vibration;

[0011] An adaptive threshold segmentation algorithm was used to extract the exhaust smoke region, and a three-dimensional spatial coordinate model of exhaust smoke diffusion was established.

[0012] The spectral data is demixed using a nonnegative matrix factorization algorithm to obtain the spectral feature vectors of each component.

[0013] By combining a pre-constructed spectral library of aircraft fuel components for matching analysis, the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides and unburned hydrocarbons in the combustion products is quantitatively calculated.

[0014] A three-dimensional dynamic evolution map of the chemical composition of exhaust smoke was generated by using a convolutional neural network model to perform spatiotemporal correlation analysis on multi-component concentration data.

[0015] Preferably, the acquisition of multispectral image data of the aircraft contrail region via the micro-on-chip spectral imaging system specifically includes:

[0016] Initialize the parameters of the spectral imaging system, set the wavelength scanning range of the tunable filter module to 200-2500nm, the spectral resolution to 5nm, and adjust the integration time of the array detector to 10-1000μs to adapt to the dynamic range of the exhaust smoke.

[0017] Wavelength scanning is performed by driving a tunable Fabry-Pérot interferometer through a microelectromechanical system (MEMS), and its cavity length L is adjusted according to the following formula:

[0018] ;

[0019] In the formula, m is the interference order, m=1, n is the refractive index of the medium, and λ is the current scanning wavelength;

[0020] The array detector is controlled to synchronously acquire image data at various wavelengths, and the detector is made of CMOS material.

[0021] Vibration compensation is implemented by monitoring the aircraft's vibration displacement in real time using an inertial measurement unit. And adjust the cavity length L of the filter module to compensate for changes in optical path;

[0022] ;

[0023] In the formula, The displacement caused by vibration;

[0024] Spatiotemporal alignment of the acquired multi-frame images was performed using a feature point-based registration algorithm, with the matching degree threshold set to a correlation coefficient ρ > 0.95.

[0025] Generate a three-dimensional spectral data cube V∈ℝ^{w×h×λ}, where w and h are spatial dimensions and λ is the spectral dimension.

[0026] Preferably, the spatial alignment and radiometric correction processing of the acquired multispectral image specifically includes:

[0027] The inertial measurement unit monitors the triaxial acceleration and angular velocity of the aircraft in real time, with a sampling frequency of 1,000 times per second and a measurement range covering displacements of ±500 micrometers.

[0028] A vibration displacement model is constructed based on monitoring data. The Kalman filter algorithm is used to estimate image distortion. The position and angle state parameters are iteratively updated through the state transition equation.

[0029] Spatial alignment is performed, and a feature point detection method based on the SIFT algorithm is used to extract 128-dimensional feature descriptors. A matching threshold with a correlation coefficient greater than 0.95 is set to filter valid matching points.

[0030] Affine transformation model is applied to correct image distortion. Rotation, scaling, and translation parameters are optimized using the least squares method. A random sampling consensus algorithm is used to eliminate mismatched points, requiring an inlier ratio of over 90%.

[0031] Dark current correction was performed by acquiring 100 frames of data in the absence of light, with the detector temperature stabilized at -20 degrees Celsius, calculating the average dark current and subtracting it from the original image.

[0032] Flat field correction is performed by generating a uniform radiation field using an integrating sphere light source, calculating the response coefficient of each pixel, and dividing the original image by the coefficient to achieve brightness normalization.

[0033] To compensate for the influence of atmospheric transmittance, the MODTRAN model is used to input parameters such as flight altitude, visibility, and aerosol type to generate atmospheric transmittance curves for each wavelength, thereby correcting the original spectral intensity.

[0034] To perform nonlinear response correction, a cubic polynomial mapping model is established. The curve relationship between radiation intensity and output signal is fitted using the least squares method, and the goodness of fit R is required to be... 2 Exceeding 0.999;

[0035] Dynamic range adjustment is implemented, and adaptive logarithmic transformation is used to enhance weak signals. The transformation parameters are automatically optimized according to the real-time scene brightness. Image quality is evaluated through indicators such as peak signal-to-noise ratio and structural similarity to ensure that the corrected data meets the preset accuracy requirements.

[0036] Preferably, the step of extracting the exhaust smoke region using an adaptive threshold segmentation algorithm and establishing a three-dimensional spatial coordinate model of exhaust smoke diffusion specifically includes:

[0037] The multispectral image was preprocessed by using a Gaussian filter to eliminate high-frequency noise. The filter kernel size was set to 5×5 pixels and the standard deviation σ=1.5.

[0038] Image texture features are calculated based on the gray-level co-occurrence matrix, with a matrix step size d=1, direction θ=0°, and the gray level is compressed to 16 levels to reduce computational complexity.

[0039] A dynamic threshold segmentation strategy is adopted, and the initial threshold is determined by the Otsu algorithm;

[0040] The segmentation results are optimized by applying morphological opening and closing operations. The structuring element uses a circular template with a radius of r=3 pixels to eliminate isolated regions with an area of ​​less than 10 pixels.

[0041] A smoke exhaust region mask is established, and the largest area region is selected as the effective target through connected component analysis. The connectivity judgment adopts the 8-neighborhood criterion.

[0042] Spatial coordinate transformation is performed by combining the aircraft attitude data. The pixel coordinates (u,v) are mapped to the world coordinate system (x,y,z) through camera calibration parameters. The transformation matrix includes rotation, translation, and scaling parameters.

[0043] A three-dimensional voxel model was constructed with a voxel resolution of 0.1m×0.1m×0.1m. The smoke concentration was linearly mapped to the voxel gray value through spectral intensity.

[0044] Diffusion simulation was conducted, and the Gaussian fuzzy algorithm was used to simulate the smoke diffusion process. The fuzzy radius increased linearly with time t: σ(t) = 0.5 + 0.1·t;

[0045] A three-dimensional isosurface is generated, and a concentration threshold surface is extracted using the moving cube algorithm. The threshold is set to 30%-70% of the maximum concentration. The model accuracy is verified by comparing the measured data from ground-based lidar with the requirements of spatial overlap rate >85% and concentration error <15%.

[0046] Preferably, the step of unmixing the spectral data using a nonnegative matrix factorization algorithm to obtain the spectral feature vectors of each component specifically includes:

[0047] The original spectral data were preprocessed, and wavelet transform was used to remove high-frequency noise. The transform basis function was selected as the Daubechies fourth-order wavelet, and the number of decomposition layers was set to 3.

[0048] The number of endmembers k is estimated using a virtual dimension algorithm. After noise whitening, the eigenvalue ratio is calculated, and a threshold is set when the rate of change of the eigenvalue descent slope exceeds 50%.

[0049] Initialize matrices W and H using a non-negative random number generation strategy. The matrix dimensions are n×k and k×g, respectively, where n is the number of spectral channels and g is the number of spatial pixels.

[0050] Perform iterative optimization, using a multiplicative update rule to alternately update W and H:

[0051] Implement nonnegativity constraints by setting the negative elements of W and H to zero after each iteration to ensure that all matrix elements are nonnegative;

[0052] Determine the termination condition: stop iterating when the number of iterations reaches 2000, and perform abundance normalization on the H matrix to ensure that the sum of the abundance coefficients of each pixel is 1;

[0053] The unmixing results were verified by matching the endmember spectra with a standard spectral library. The matching degree was evaluated using the correlation coefficient method, with a threshold of ρ>0.85.

[0054] Output the spectral feature vector W and the corresponding abundance coefficient matrix H of each component. The spectral feature vectors are arranged in descending order of correlation coefficient.

[0055] Preferably, the alternating update of W and H using the multiplicative update rule specifically includes:

[0056] Fix H, update W: W←W⊙[(VHT) . / (WHHT)];

[0057] Fixed W, updated H: H←H⊙[(WTV) . / (WTWH)];

[0058] Where "⊙" represents the Hadamard product and ". / " represents element-wise division.

[0059] Preferably, the step of performing matching analysis using a pre-constructed spectral library of aircraft fuel components to quantitatively calculate the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides, and unburned hydrocarbons in the combustion products specifically includes:

[0060] A fuel component spectral library was constructed, including spectral data of the following typical fuels: aviation kerosene, covering characteristic absorption peaks at 3.3 μm (CH stretching vibration) and 4.7 μm (CO2 symmetric stretching vibration); hydrogen fuel, containing characteristic emission lines at 656 nm (Hα line) and 486 nm (Hβ line); and solid propellants, covering Al2O3 particle scattering bands in the 9-11 μm band.

[0061] The measured spectrum was preprocessed, and a Savitzky-Golay filter was used to eliminate high-frequency noise. The filter window size was set to 11 points and the polynomial order was 3.

[0062] Spectral characteristic bands were extracted, specific bands were selected for each combustion product, and spectral matching analysis was performed. The similarity between the measured spectrum and the library spectrum was calculated using the correlation coefficient method.

[0063] A quantitative calibration model was established using a partial least squares regression algorithm. The calibration sample set included 30-50 sets of standard spectra with different concentration gradients. Model training was carried out, and the number of principal components was determined through cross-validation. The validation index was the sum of squared predicted residuals. The number of principal components retained 95% of the variance information.

[0064] The concentration distribution is calculated by inputting the measured spectrum into the trained PLSR model and outputting the concentration prediction values ​​of each component. The predicted values ​​are limited to the physical reasonable range of 0-100%. Post-processing correction is performed to spatially smooth the concentration prediction values ​​and use a Gaussian filter to eliminate isolated outliers. The filter standard deviation σ = 1.5 pixels.

[0065] To verify the quantitative results, the measured concentration of standard gas samples was compared with the predicted concentration. The average absolute error was required to be <5%, and the coefficient of determination was required to be >0.95.

[0066] Output a concentration distribution map and map the quantitative results to a three-dimensional spatial coordinate model of the aircraft exhaust. The concentration values ​​are encoded using pseudo-color with a color scale range of 0-255.

[0067] Preferably, the selection of specific wavelengths for each combustion product specifically includes:

[0068] CO2 selects a specific wavelength band of 4.2-4.4 μm;

[0069] CO is selected in the specific wavelength range of 4.6-4.8 μm;

[0070] NOx is selected in the specific band of 5.3-5.5 μm;

[0071] For unburned hydrocarbons, a specific wavelength band of 3.2-3.5 μm is selected.

[0072] Preferably, the step of using a convolutional neural network model to perform spatiotemporal correlation analysis on multi-component concentration data to generate a three-dimensional dynamic evolution map of the chemical composition of exhaust smoke specifically includes:

[0073] The input is multi-component concentration data processed by the non-negative matrix factorization algorithm. The multi-component concentration data includes CO2 / CO / NOx / unburned hydrocarbons. The data format is 40×40 pixel spectral image blocks. A convolutional neural network with U-Net++ architecture is constructed, which includes an encoder-decoder structure. The hidden layer is set with 8 residual blocks, each block containing 2 convolutional layers. The input layer receives the concentration data, and the output layer generates a 512-dimensional feature vector for concentration inversion.

[0074] The model incorporates a turbulence model and chemical transport equations, solves the partial differential equations using the fourth-order Runge-Kutta method with a time step of 0.1 seconds, constructs a diffusion path by combining the vehicle's motion trajectory data, and employs an attention mechanism to enhance spatiotemporal feature extraction. The attention weights are calculated using a space-channel joint attention module.

[0075] The network was trained using a pre-labeled exhaust smoke diffusion dataset. The loss function was defined as a weighted combination of mean squared error and Dice coefficient. Dynamic learning rate adjustment was implemented, and the early stopping strategy was set to terminate training when the verification loss did not decrease for 5 consecutive rounds.

[0076] The feature vectors output by the model are mapped to a 3D voxel model. The density values ​​are encoded using pseudo-color. The Marching Cubes algorithm is used to extract iso-density surfaces. Volume rendering is achieved by combining a ray casting algorithm. The viewpoint transformation is controlled by a 3D rotation matrix.

[0077] The accuracy of the map was verified by comparing the measured data from ground-based lidar. The requirements were that the spatial overlap rate was >85% and the concentration error was <15%. The final output was a three-dimensional dynamic evolution map containing the concentration gradient field, diffusion velocity field and chemical reaction rate, with an update frequency of 10Hz.

[0078] In a second aspect of the invention, a multi-component analysis system for aircraft exhaust plumes based on micro-on-chip spectral imaging is also provided, comprising:

[0079] The acquisition module is used to acquire multispectral image data of the aircraft's exhaust region through a micro on-chip spectral imaging system. The multispectral image data is a three-dimensional spectral data cube. The system integrates a tunable filter module and an array detector.

[0080] The processing module is used to perform spatial alignment and radiometric correction processing on the acquired multispectral images to eliminate image distortion caused by aircraft vibration;

[0081] The model building module is used to extract the exhaust smoke region using an adaptive threshold segmentation algorithm and to build a three-dimensional spatial coordinate model of exhaust smoke diffusion.

[0082] The feature vector extraction module is used to unmix the spectral data using a non-negative matrix factorization algorithm to obtain the spectral feature vectors of each component.

[0083] The analysis module is used to perform matching analysis with a pre-built spectral library of aircraft fuel components to quantitatively calculate the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides and unburned hydrocarbons in the combustion products.

[0084] The generation module is used to perform spatiotemporal correlation analysis on multi-component concentration data using a convolutional neural network model to generate a three-dimensional dynamic evolution map of the chemical composition of exhaust smoke.

[0085] Compared with existing technologies, this invention provides a method and system for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging, which has the following advantages:

[0086] This invention utilizes a nonnegative matrix factorization (NMF) algorithm to unmix the mixed spectra. Combined with a pre-built fuel component spectral library, it can accurately identify and quantitatively calculate the concentration distribution of key components such as carbon dioxide, carbon monoxide, nitrogen oxides, and unburned hydrocarbons in exhaust smoke. This method overcomes the errors caused by component overlap in traditional spectral analysis, improving the accuracy of quantitative results. By establishing a three-dimensional spatial coordinate model, high-precision extraction of the exhaust smoke region is achieved, avoiding background interference and further ensuring the reliability of concentration calculations. Furthermore, a convolutional neural network (CNN) is used to perform spatiotemporal correlation modeling of multi-component concentration data, generating a three-dimensional dynamic evolution map of exhaust smoke chemical components. This map can reflect the diffusion patterns of exhaust smoke components over time and space in real time, providing crucial data support for environmental impact assessment, aircraft emission control, and accident emergency response. Through spatial alignment and radiometric correction processing, image distortion caused by aircraft vibration is effectively eliminated, ensuring the stability of dynamic monitoring. Attached Figure Description

[0087] Figure 1 This is a schematic diagram of the method flow of S101-S106 in this invention;

[0088] Figure 2 This is a schematic diagram of the method flow for S201-S209 in this invention;

[0089] Figure 3 This is a schematic diagram of the method flow of S301-S309 in this invention;

[0090] Figure 4 This is a schematic diagram of the method flow of S401-S408 in this invention;

[0091] Figure 5 This is a schematic diagram of the method flow for S501-S507 in this invention;

[0092] Figure 6 This is a schematic diagram of the method flow for S601-S605 in this invention. Detailed Implementation

[0093] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0094] Example 1

[0095] Please refer to Figure 1 As shown, in a first aspect of the present invention, a method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging is provided, comprising:

[0096] S101. Multispectral image data of the aircraft's exhaust region is acquired through a micro on-chip spectral imaging system. The multispectral image data is a three-dimensional spectral data cube. The system integrates a tunable filter module and an array detector.

[0097] S102. Perform spatial alignment and radiometric correction on the acquired multispectral images to eliminate image distortion caused by aircraft vibration;

[0098] S103. An adaptive threshold segmentation algorithm is used to extract the exhaust smoke region and a three-dimensional spatial coordinate model of exhaust smoke diffusion is established.

[0099] S104. The spectral data is demixed using a non-negative matrix factorization algorithm to obtain the spectral feature vectors of each component.

[0100] S105. By combining the pre-constructed spectral library of aircraft fuel components for matching analysis, the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides and unburned hydrocarbons in the combustion products is quantitatively calculated.

[0101] S106. Using a convolutional neural network model, spatiotemporal correlation analysis is performed on the concentration data of multiple components to generate a three-dimensional dynamic evolution map of the chemical composition of exhaust smoke.

[0102] As will be understood by those skilled in the art, this invention uses a nonnegative matrix factorization (NMF) algorithm to unmix the mixed spectra, and combined with a pre-built fuel component spectral library, to accurately identify and quantitatively calculate the concentration distribution of key components such as carbon dioxide, carbon monoxide, nitrogen oxides, and unburned hydrocarbons in the exhaust smoke. This method overcomes the errors caused by component overlap in traditional spectral analysis and improves the accuracy of quantitative results. By establishing a three-dimensional spatial coordinate model, high-precision extraction of the exhaust smoke region is achieved, avoiding background interference and further ensuring the reliability of concentration calculation. Furthermore, a convolutional neural network (CNN) is used to perform spatiotemporal correlation modeling of multi-component concentration data to generate a three-dimensional dynamic evolution map of the chemical composition of the exhaust smoke. This spectrum can reflect the diffusion patterns of exhaust components over time and space in real time, providing key data support for environmental impact assessment, aircraft emission control, and emergency response. Through spatial alignment and radiometric correction, it effectively eliminates image distortion caused by aircraft vibration, ensuring the stability of dynamic monitoring. Furthermore, by adopting an on-chip spectral imaging system that integrates a tunable filter module and an array detector, it significantly reduces the size and weight of the equipment, making it suitable for real-time online monitoring of aircraft and overcoming the limitations of traditional large-scale spectrometers in terms of portability.

[0103] Acquiring multispectral image data of the aircraft's exhaust region using a miniature on-chip spectral imaging system specifically includes:

[0104] Initialize the parameters of the spectral imaging system, set the wavelength scanning range of the tunable filter module to 200-2500nm, the spectral resolution to 5nm, and adjust the integration time of the array detector to 10-1000μs to adapt to the dynamic range of the exhaust smoke.

[0105] Wavelength scanning is performed by driving a tunable Fabry-Pérot interferometer through a microelectromechanical system (MEMS), and its cavity length L is adjusted according to the following formula:

[0106] ;

[0107] In the formula, m is the interference order, m=1, n is the refractive index of the medium, and λ is the current scanning wavelength;

[0108] The array detector is controlled to synchronously acquire image data at various wavelengths, and the detector is made of CMOS material.

[0109] Vibration compensation is implemented by monitoring the aircraft's vibration displacement in real time using an inertial measurement unit. And adjust the cavity length L of the filter module to compensate for changes in optical path;

[0110] ;

[0111] In the formula, The displacement caused by vibration;

[0112] Spatiotemporal alignment of the acquired multi-frame images was performed using a feature point-based registration algorithm, with the matching degree threshold set to a correlation coefficient ρ > 0.95.

[0113] Generate a three-dimensional spectral data cube V∈ℝ^{w×h×λ}, where w and h are spatial dimensions and λ is the spectral dimension.

[0114] Please refer to Figure 2 As shown, the spatial alignment and radiometric correction processing of the acquired multispectral images specifically includes:

[0115] S201: The inertial measurement unit monitors the three-axis acceleration and angular velocity of the aircraft in real time, with a sampling frequency of 1,000 times per second and a measurement range covering displacements of ±500 micrometers.

[0116] S202. Construct a vibration displacement model based on monitoring data, use the Kalman filter algorithm to estimate image distortion, and iteratively update the position and angle state parameters through the state transition equation.

[0117] S203. Perform spatial alignment operation, adopt the feature point detection method based on SIFT algorithm, extract 128-dimensional feature descriptors, and set a matching threshold with a correlation coefficient greater than 0.95 to filter valid matching points;

[0118] S204. Apply an affine transformation model to correct image distortion, optimize rotation, scaling, and translation parameters using the least squares method, and use a random sampling consensus algorithm to eliminate mismatched points, requiring an inlier ratio of over 90%.

[0119] S205. Perform dark current correction. With the detector temperature stable at -20 degrees Celsius, collect 100 frames of data without illumination, calculate the average dark current, and subtract it from the original image.

[0120] S206. Implement flat field correction, use an integrating sphere light source to generate a uniform radiation field, calculate the response coefficient of each pixel, and divide the original image by the coefficient to achieve brightness normalization.

[0121] S207. To compensate for the influence of atmospheric transmittance, the MODTRAN model is used to input flight altitude, visibility, and aerosol type parameters to generate atmospheric transmittance curves for each wavelength and correct the original spectral intensity.

[0122] S208. Perform nonlinear response correction, establish a cubic polynomial mapping model, and fit the curve relationship between radiation intensity and output signal using the least squares method, requiring a goodness of fit R. 2 Exceeding 0.999;

[0123] S209. Implement dynamic range adjustment, adopt adaptive logarithmic transformation to enhance weak signals, and automatically optimize transformation parameters according to real-time scene brightness. Evaluate image quality through indicators such as peak signal-to-noise ratio and structural similarity to ensure that the corrected data meets the preset accuracy requirements.

[0124] As will be understood by those skilled in the art, this invention monitors aircraft vibration through high-frequency sampling at 1000 times per second, and combines this with Kalman filtering to dynamically estimate distortion, achieving sub-pixel-level (<500μm) displacement compensation, significantly reducing the impact of motion blur on spectral resolution. It employs an affine transformation model (optimized rotation, scaling, and translation parameters) combined with the RANSAC algorithm to eliminate mismatched points (inlier rate >90%), ensuring geometric correction error <0.5 pixels and improving the spatial consistency of multispectral data. It extracts 128-dimensional feature descriptors, setting the matching threshold to a correlation coefficient of 0.95, exhibiting strong robustness to scale changes (±20%), rotation (±30°), and illumination fluctuations (±40%), adapting to dynamic flight scenarios. The registration error between multispectral images is <1 pixel, eliminating spectral-spatial data misalignment caused by viewpoint changes or vibration, providing an accurate foundation for subsequent 3D modeling.

[0125] Please refer to Figure 3 As shown, the adaptive threshold segmentation algorithm is used to extract the exhaust smoke region, and a three-dimensional spatial coordinate model of exhaust smoke diffusion is established, specifically including:

[0126] S301. Perform preprocessing on the multispectral image, using a Gaussian filter to eliminate high-frequency noise, with the filter kernel size set to 5×5 pixels and the standard deviation σ=1.5;

[0127] S302. Calculate image texture features based on gray-level co-occurrence matrix, with matrix step size d=1, direction θ=0°, and gray level compressed to 16 levels to reduce computational complexity;

[0128] S303. A dynamic threshold segmentation strategy is adopted, and the initial threshold is determined by the Otsu algorithm.

[0129] S304. Apply morphological opening and closing operations to optimize the segmentation results. Use a circular template for the structural element with a radius of r=3 pixels to eliminate isolated regions with an area of ​​less than 10 pixels.

[0130] S305. Establish a smoke exhaust area mask, and use connected component analysis to select the largest area region as the effective target. The connectivity judgment adopts the 8-neighborhood criterion.

[0131] S306. Combine the aircraft attitude data to perform spatial coordinate transformation. Map the pixel coordinates (u,v) to the world coordinate system (x,y,z) through camera calibration parameters. The transformation matrix includes rotation, translation and scaling parameters.

[0132] S307. Construct a three-dimensional voxel model with a voxel resolution of 0.1m×0.1m×0.1m. The exhaust concentration is linearly mapped to the voxel gray value through spectral intensity.

[0133] S308. Implement diffusion simulation, using Gaussian fuzzy algorithm to simulate the exhaust smoke diffusion process, with the fuzzy radius increasing linearly with time t: σ(t) = 0.5 + 0.1·t;

[0134] S309. Generate a three-dimensional isosurface and extract the concentration threshold surface using the moving cube algorithm. The threshold is set to 30%-70% of the maximum concentration. The model accuracy is verified by comparing the measured data from ground-based lidar with the required spatial overlap rate >85% and concentration error <15%.

[0135] Please refer to Figure 4 As shown, the spectral data is unmixed using a nonnegative matrix factorization algorithm to obtain the spectral feature vectors of each component, specifically including:

[0136] S401. Preprocess the original spectral data by using wavelet transform to remove high-frequency noise. The transform basis function is selected as the Daubechies fourth-order wavelet, and the decomposition level is set to 3 levels.

[0137] S402. Estimate the number of endmembers k. Use the virtual dimension algorithm to calculate the eigenvalue ratio after noise whitening. Set a threshold when the rate of change of the eigenvalue descent slope exceeds 50%.

[0138] S403. Initialize matrices W and H using a non-negative random number generation strategy. The matrix dimensions are n×k and k×g, respectively, where n is the number of spectral channels and g is the number of spatial pixels.

[0139] S404. Perform iterative optimization, using a multiplicative update rule to alternately update W and H:

[0140] S405. Implement non-negativity constraints by setting the negative elements of W and H to zero after each iteration to ensure that all matrix elements are non-negative.

[0141] S406. Determine the termination condition. Stop iterating when the number of iterations reaches 2000. Perform abundance normalization on the H matrix to ensure that the sum of the abundance coefficients of each pixel is 1.

[0142] S407. Verify the unmixing results by matching the endmember spectra with the standard spectral library. The matching degree is evaluated using the correlation coefficient method, with a threshold ρ>0.85 set.

[0143] S408. Output the spectral feature vector W and the corresponding abundance coefficient matrix H of each component. The spectral feature vectors are arranged in descending order of correlation coefficient.

[0144] The alternating updates of W and H using the multiplicative update rule specifically include:

[0145] Fix H, update W: W←W⊙[(VHT) . / (WHHT)];

[0146] Fixed W, updated H: H←H⊙[(WTV) . / (WTWH)];

[0147] Where "⊙" represents the Hadamard product and ". / " represents element-wise division.

[0148] Please refer to Figure 5 As shown, by combining a pre-constructed spectral library of aircraft fuel components for matching analysis, the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides, and unburned hydrocarbons in the combustion products was quantitatively calculated, specifically including:

[0149] S501. Construct a fuel component spectral library, including spectral data of the following typical fuels, including aviation kerosene: covering characteristic absorption peaks at 3.3 μm (CH stretching vibration) and 4.7 μm (CO2 symmetric stretching vibration); hydrogen fuel: containing characteristic emission lines at 656 nm (Hα line) and 486 nm (Hβ line); solid propellants: covering Al2O3 particle scattering bands in the 9-11 μm band;

[0150] S502. Preprocess the measured spectrum by using a Savitzky-Golay filter to eliminate high-frequency noise. The filter window size is set to 11 points and the polynomial order is 3.

[0151] S503. Extract spectral characteristic bands, select specific bands for each combustion product, and perform spectral matching analysis. Calculate the similarity between the measured spectrum and the library spectrum using the correlation coefficient method.

[0152] S504. Establish a quantitative calibration model, using partial least squares regression algorithm. The calibration sample set contains 30-50 sets of standard spectra with different concentration gradients. Implement model training. The number of principal components is determined through cross-validation. The validation index is the sum of squared predicted residuals. The number of principal components retains 95% of the variance information.

[0153] S505. Calculate the concentration distribution. Input the measured spectrum into the trained PLSR model and output the predicted concentration values ​​of each component. The predicted values ​​are limited to the 0-100% physical reasonable range. Perform post-processing correction, perform spatial smoothing on the concentration prediction values, and use a Gaussian filter to eliminate isolated outliers. The filter standard deviation σ = 1.5 pixels.

[0154] S506. Verify the quantitative results by comparing the measured concentration with the predicted concentration of standard gas sampling. The average absolute error should be <5% and the coefficient of determination should be >0.95.

[0155] S507 Output concentration distribution map, mapping the quantitative results to the three-dimensional spatial coordinate model of the aircraft exhaust. The concentration values ​​are encoded using pseudo-color, with a color scale range of 0-255.

[0156] Specific wavelength bands are selected for each combustion product, including:

[0157] CO2 selects a specific wavelength band of 4.2-4.4 μm;

[0158] CO is selected in the specific wavelength range of 4.6-4.8 μm;

[0159] NOx is selected in the specific band of 5.3-5.5 μm;

[0160] For unburned hydrocarbons, a specific wavelength band of 3.2-3.5 μm is selected.

[0161] Please refer to Figure 6 As shown, the spatiotemporal correlation analysis of multi-component concentration data using a convolutional neural network model generates a three-dimensional dynamic evolution map of the chemical composition of exhaust smoke, specifically including:

[0162] S601. Input the multi-component concentration data processed by the non-negative matrix factorization algorithm. The multi-component concentration data includes CO2 / CO / NOx / unburned hydrocarbons. The data format is 40×40 pixel spectral image blocks. Construct a U-Net++ architecture convolutional neural network, which includes an encoder-decoder structure. The hidden layer is set with 8 residual blocks, each block containing 2 convolutional layers. The input layer receives the concentration data, and the output layer generates a 512-dimensional feature vector for concentration inversion.

[0163] S602. Embed the turbulence model and chemical transport equations into the model, solve the partial differential equations using the fourth-order Runge-Kutta method, with a time step of 0.1 seconds, construct the diffusion path by combining the aircraft motion trajectory data, and use an attention mechanism to enhance the extraction of spatiotemporal features. The attention weights are calculated through the space-channel joint attention module.

[0164] S603. Train the network using a pre-labeled exhaust smoke diffusion dataset. The loss function is defined as a weighted combination of mean squared error and Dice coefficient. Implement dynamic learning rate adjustment. The early stopping strategy is set to terminate training when the verification loss does not decrease for 5 consecutive rounds.

[0165] S604. Map the feature vectors output by the model to a three-dimensional voxel model. The density values ​​are encoded using pseudo-color. The Marching Cubes algorithm is used to extract iso-density surfaces. The volume is drawn by combining the ray casting algorithm. The viewpoint is controlled by a three-dimensional rotation matrix.

[0166] S605. The accuracy of the map is verified by comparing the measured data of the ground lidar. The spatial overlap rate is required to be >85% and the concentration error is <15%. The final output is a three-dimensional dynamic evolution map containing the concentration gradient field, diffusion velocity field and chemical reaction rate, with an update frequency of 10Hz.

[0167] In a second aspect of the invention, a multi-component analysis system for aircraft exhaust plumes based on micro-on-chip spectral imaging is also provided, comprising:

[0168] The acquisition module is used to acquire multispectral image data of the aircraft's exhaust region through a micro on-chip spectral imaging system. The multispectral image data is a three-dimensional spectral data cube. The system integrates a tunable filter module and an array detector.

[0169] The processing module is used to perform spatial alignment and radiometric correction on the acquired multispectral images to eliminate image distortion caused by aircraft vibration.

[0170] The model building module is used to extract the exhaust smoke region using an adaptive threshold segmentation algorithm and to build a three-dimensional spatial coordinate model of exhaust smoke diffusion.

[0171] The feature vector extraction module is used to unmix spectral data using a non-negative matrix factorization algorithm to obtain the spectral feature vectors of each component.

[0172] The analysis module is used to perform matching analysis with a pre-built spectral library of aircraft fuel components to quantitatively calculate the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides and unburned hydrocarbons in combustion products.

[0173] The generation module is used to perform spatiotemporal correlation analysis on multi-component concentration data using a convolutional neural network model to generate a three-dimensional dynamic evolution map of the chemical composition of exhaust smoke.

[0174] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging, characterized in that, include: Multispectral image data of the aircraft's exhaust region is acquired through a micro on-chip spectral imaging system. The multispectral image data is a three-dimensional spectral data cube. The system integrates a tunable filter module and an array detector. Spatial alignment and radiometric correction are performed on the acquired multispectral images to eliminate image distortion caused by aircraft vibration; An adaptive threshold segmentation algorithm was used to extract the exhaust smoke region, and a three-dimensional spatial coordinate model of exhaust smoke diffusion was established. The spectral data is demixed using a nonnegative matrix factorization algorithm to obtain the spectral feature vectors of each component. By combining a pre-constructed spectral library of aircraft fuel components for matching analysis, the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides and unburned hydrocarbons in the combustion products is quantitatively calculated. A three-dimensional dynamic evolution map of the chemical composition of exhaust smoke was generated by using a convolutional neural network model to perform spatiotemporal correlation analysis on multi-component concentration data. The acquisition of multispectral image data of the aircraft contrail region via a micro on-chip spectral imaging system specifically includes: Initialize the parameters of the spectral imaging system, set the wavelength scanning range of the tunable filter module to 200-2500nm, the spectral resolution to 5nm, and adjust the integration time of the array detector to 10-1000μs to adapt to the dynamic range of the exhaust smoke. Wavelength scanning is performed by driving a tunable Fabry-Pérot interferometer through a microelectromechanical system (MEMS), and its cavity length L is adjusted according to the following formula: ； In the formula, m is the interference order, m=1, n is the refractive index of the medium, and λ is the current scanning wavelength; The array detector is controlled to synchronously acquire image data at various wavelengths, and the detector is made of CMOS material. Vibration compensation is implemented by monitoring the aircraft's vibration displacement in real time using an inertial measurement unit. And adjust the cavity length L of the filter module to compensate for changes in optical path; ； In the formula, The displacement caused by vibration; Spatiotemporal alignment of the acquired multi-frame images was performed using a feature point-based registration algorithm, with the matching degree threshold set to a correlation coefficient ρ > 0.

95. Generate a three-dimensional spectral data cube V∈ℝ^{w×h×λ}, where w and h are spatial dimensions and λ is the spectral dimension.

2. The method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging according to claim 1, characterized in that, The spatial alignment and radiometric correction processing of the acquired multispectral image specifically includes: The inertial measurement unit monitors the triaxial acceleration and angular velocity of the aircraft in real time, with a sampling frequency of 1,000 times per second and a measurement range covering displacements of ±500 micrometers. A vibration displacement model is constructed based on monitoring data. The Kalman filter algorithm is used to estimate image distortion. The position and angle state parameters are iteratively updated through the state transition equation. Spatial alignment is performed, and a feature point detection method based on the SIFT algorithm is used to extract 128-dimensional feature descriptors. A matching threshold with a correlation coefficient greater than 0.95 is set to filter valid matching points. Affine transformation model is applied to correct image distortion. Rotation, scaling, and translation parameters are optimized using the least squares method. A random sampling consensus algorithm is used to eliminate mismatched points, requiring an inlier ratio of over 90%. Dark current correction was performed by acquiring 100 frames of data in the absence of light, with the detector temperature stabilized at -20 degrees Celsius, calculating the average dark current and subtracting it from the original image. Flat field correction is performed by generating a uniform radiation field using an integrating sphere light source, calculating the response coefficient of each pixel, and dividing the original image by the coefficient to achieve brightness normalization. To compensate for the influence of atmospheric transmittance, the MODTRAN model is used to input parameters such as flight altitude, visibility, and aerosol type to generate atmospheric transmittance curves for each wavelength, thereby correcting the original spectral intensity. To perform nonlinear response correction, a cubic polynomial mapping model is established. The curve relationship between radiation intensity and output signal is fitted using the least squares method, and the goodness of fit R is required to be... 2 Exceeding 0.999; Dynamic range adjustment is implemented, and adaptive logarithmic transformation is used to enhance weak signals. The transformation parameters are automatically optimized according to the real-time scene brightness. Image quality is evaluated through peak signal-to-noise ratio and structural similarity indicators to ensure that the corrected data meets the preset accuracy requirements.

3. The method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging according to claim 2, characterized in that, The process of extracting the exhaust smoke region using an adaptive threshold segmentation algorithm and establishing a three-dimensional spatial coordinate model for exhaust smoke diffusion specifically includes: The multispectral image was preprocessed by using a Gaussian filter to eliminate high-frequency noise. The filter kernel size was set to 5×5 pixels and the standard deviation σ=1.

5. Image texture features are calculated based on the gray-level co-occurrence matrix, with a matrix step size d=1, direction θ=0°, and the gray level is compressed to 16 levels to reduce computational complexity. A dynamic threshold segmentation strategy is adopted, and the initial threshold is determined by the Otsu algorithm; The segmentation results are optimized by applying morphological opening and closing operations. The structuring element uses a circular template with a radius of r=3 pixels to eliminate isolated regions with an area of ​​less than 10 pixels. A smoke exhaust region mask is established, and the largest area region is selected as the effective target through connected component analysis. The connectivity judgment adopts the 8-neighborhood criterion. Spatial coordinate transformation is performed by combining the aircraft attitude data. The pixel coordinates (u,v) are mapped to the world coordinate system (x,y,z) through camera calibration parameters. The transformation matrix includes rotation, translation, and scaling parameters. A three-dimensional voxel model was constructed with a voxel resolution of 0.1m×0.1m×0.1m. The smoke concentration was linearly mapped to the voxel gray value through spectral intensity. Diffusion simulation was conducted, and the Gaussian fuzzy algorithm was used to simulate the smoke diffusion process. The fuzzy radius increased linearly with time t: σ(t) = 0.5 + 0.1·t; A three-dimensional isosurface is generated, and a concentration threshold surface is extracted using the moving cube algorithm. The threshold is set to 30%-70% of the maximum concentration. The model accuracy is verified by comparing the measured data from ground-based lidar with the requirements of spatial overlap rate >85% and concentration error <15%.

4. The method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging according to claim 3, characterized in that, The process of unmixing the spectral data using a nonnegative matrix factorization algorithm to obtain the spectral feature vectors of each component specifically includes: The original spectral data were preprocessed, and wavelet transform was used to remove high-frequency noise. The transform basis function was selected as the Daubechies fourth-order wavelet, and the number of decomposition layers was set to 3. The number of endmembers k is estimated using a virtual dimension algorithm. After noise whitening, the eigenvalue ratio is calculated, and a threshold is set when the rate of change of the eigenvalue descent slope exceeds 50%. Initialize matrices W and H using a non-negative random number generation strategy. The matrix dimensions are n×k and k×g, respectively, where n is the number of spectral channels and g is the number of spatial pixels. Perform iterative optimization, using a multiplicative update rule to alternately update W and H: Implement nonnegativity constraints by setting the negative elements of W and H to zero after each iteration to ensure that all matrix elements are nonnegative; Determine the termination condition: stop iterating when the number of iterations reaches 2000, and perform abundance normalization on the H matrix to ensure that the sum of the abundance coefficients of each pixel is 1; The unmixing results were verified by matching the endmember spectra with a standard spectral library. The matching degree was evaluated using the correlation coefficient method, with a threshold of ρ>0.

85. Output the spectral feature vector W and the corresponding abundance coefficient matrix H of each component. The spectral feature vectors are arranged in descending order of correlation coefficient.

5. The method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging according to claim 4, characterized in that, The method of alternately updating W and H using a multiplicative update rule specifically includes: Fix H, update W: W←W⊙[(VHT) . / (WHHT)]; Fixed W, updated H: H←H⊙[(WTV) . / (WTWH)]; Where "⊙" represents the Hadamard product and ". / " represents element-wise division.

6. The method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging according to claim 5, characterized in that, The specific steps of matching and analyzing a pre-constructed spectral library of aircraft fuel components to quantitatively calculate the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides, and unburned hydrocarbons in combustion products include: A fuel composition spectral library was constructed, including spectral data of the following typical fuels: aviation kerosene (covering characteristic absorption peaks at 3.3 μm and 4.7 μm); hydrogen fuel (containing characteristic emission lines at 656 nm and 486 nm); and solid propellants (covering Al2O3 particle scattering bands in the 9-11 μm band). The measured spectrum was preprocessed, and a Savitzky-Golay filter was used to eliminate high-frequency noise. The filter window size was set to 11 points and the polynomial order was 3. Spectral characteristic bands were extracted, specific bands were selected for each combustion product, and spectral matching analysis was performed. The similarity between the measured spectrum and the library spectrum was calculated using the correlation coefficient method. A quantitative calibration model was established using a partial least squares regression algorithm. The calibration sample set included 30-50 sets of standard spectra with different concentration gradients. Model training was carried out, and the number of principal components was determined through cross-validation. The validation index was the sum of squared predicted residuals. The number of principal components retained 95% of the variance information. The concentration distribution is calculated by inputting the measured spectrum into the trained PLSR model and outputting the concentration prediction values ​​of each component. The predicted values ​​are limited to the physical reasonable range of 0-100%. Post-processing correction is performed to spatially smooth the concentration prediction values ​​and use a Gaussian filter to eliminate isolated outliers. The filter standard deviation σ = 1.5 pixels. To verify the quantitative results, the measured concentration of standard gas samples was compared with the predicted concentration. The average absolute error was required to be <5%, and the coefficient of determination was required to be >0.

95. Output a concentration distribution map and map the quantitative results to a three-dimensional spatial coordinate model of the aircraft exhaust. The concentration values ​​are encoded using pseudo-color with a color scale range of 0-255.

7. The method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging according to claim 6, characterized in that, The specific selection of specific wavelengths for each combustion product includes: CO2 selects a specific wavelength band of 4.2-4.4 μm; CO is selected in the specific wavelength range of 4.6-4.8 μm; NOx is selected in the specific band of 5.3-5.5 μm; For unburned hydrocarbons, a specific wavelength band of 3.2-3.5 μm is selected.

8. The method for multi-component analysis of aircraft exhaust plumes using micro-on-chip spectral imaging according to claim 7, characterized in that, The step of using a convolutional neural network model to perform spatiotemporal correlation analysis on multi-component concentration data to generate a three-dimensional dynamic evolution map of the chemical composition of exhaust smoke specifically includes: The input is multi-component concentration data processed by the non-negative matrix factorization algorithm. The multi-component concentration data includes CO2 / CO / NOx / unburned hydrocarbons. The data format is 40×40 pixel spectral image blocks. A convolutional neural network with U-Net++ architecture is constructed, which includes an encoder-decoder structure. The hidden layer is set with 8 residual blocks, each block containing 2 convolutional layers. The input layer receives the concentration data, and the output layer generates a 512-dimensional feature vector for concentration inversion. The model incorporates a turbulence model and chemical transport equations, solves the partial differential equations using the fourth-order Runge-Kutta method with a time step of 0.1 seconds, constructs a diffusion path by combining the vehicle's motion trajectory data, and employs an attention mechanism to enhance spatiotemporal feature extraction. The attention weights are calculated using a space-channel joint attention module. The network was trained using a pre-labeled exhaust smoke diffusion dataset. The loss function was defined as a weighted combination of mean squared error and Dice coefficient. Dynamic learning rate adjustment was implemented, and the early stopping strategy was set to terminate training when the verification loss did not decrease for 5 consecutive rounds. The feature vectors output by the model are mapped to a 3D voxel model. The density values ​​are encoded using pseudo-color. The density surface is extracted using the MarchingCubes algorithm. Volume rendering is achieved by combining the ray casting algorithm. The viewpoint transformation is controlled by a 3D rotation matrix. The accuracy of the map was verified by comparing the measured data from ground-based lidar. The requirements were that the spatial overlap rate was >85% and the concentration error was <15%. The final output was a three-dimensional dynamic evolution map containing the concentration gradient field, diffusion velocity field and chemical reaction rate, with an update frequency of 10Hz.

9. A multi-component analysis system for aircraft exhaust plumes using micro-on-chip spectral imaging, used to implement the multi-component analysis method for aircraft exhaust plumes using micro-on-chip spectral imaging as described in any one of claims 1-8, characterized in that, include: The acquisition module is used to acquire multispectral image data of the aircraft's exhaust region through a micro on-chip spectral imaging system. The multispectral image data is a three-dimensional spectral data cube. The system integrates a tunable filter module and an array detector. The processing module is used to perform spatial alignment and radiometric correction processing on the acquired multispectral images to eliminate image distortion caused by aircraft vibration; The model building module is used to extract the exhaust smoke region using an adaptive threshold segmentation algorithm and to build a three-dimensional spatial coordinate model of exhaust smoke diffusion. The feature vector extraction module is used to unmix the spectral data using a non-negative matrix factorization algorithm to obtain the spectral feature vectors of each component. The analysis module is used to perform matching analysis with a pre-built spectral library of aircraft fuel components to quantitatively calculate the concentration distribution of carbon dioxide, carbon monoxide, nitrogen oxides and unburned hydrocarbons in the combustion products. The generation module is used to perform spatiotemporal correlation analysis on multi-component concentration data using a convolutional neural network model to generate a three-dimensional dynamic evolution map of the chemical composition of exhaust smoke.

Images