An on-chip integrated system for multispectral and hyperspectral component analysis of aircraft exhaust plumes

Abstract

The application discloses a kind of on-chip integrated systems of aircraft tail smoke multispectral hyperspectral component analysis, including construction module, the construction module is used to construct the tail smoke spectral radiation transfer equation considering gas molecule and alumina solid particle coupling action.In the present application, the tail smoke spectral radiation transfer equation considering coupling is constructed, which more comprehensively describes the complex radiation transfer process in the tail smoke, can more accurately reflect the spectral radiation characteristics of tail smoke, lays the foundation for subsequent accurate analysis of tail smoke composition and flow field parameters, and the physically constrained machine learning model combines the advantages of physical law and machine learning algorithm, can utilize the rich spatial and spectral information in fusion image, while following the physical mechanism of tail smoke composition change, avoiding the overfitting or not conforming to the physical actual problem that may appear in pure machine learning model, so as to improve the accuracy of tail smoke CO2 、 H2O and unburned hydrocarbon content prediction.

Description

Technical Field

[0001] This invention relates to the field of spectral image processing technology, specifically to an on-chip integrated system for multispectral and hyperspectral component analysis of aircraft exhaust. Background Technology

[0002] In existing technologies, the analysis of aircraft exhaust composition mainly faces the following technical bottlenecks:

[0003] 1. Limitations of data fusion methods: Traditional multispectral and hyperspectral image fusion often uses simple overlay or principal component analysis, which makes it difficult to simultaneously maintain the spatial resolution of the original multispectral image (e.g., 1m level) and the spectral resolution of the hyperspectral image (e.g., 2.5nm level), resulting in spatial blurring or spectral distortion in the fused image.

[0004] 2. Insufficient accuracy of physical models: The calculation of exhaust radiation transmission usually ignores the coupling effect between gas molecules and solid particles, and only uses a single gas absorption model or particle scattering model. Under high temperature (>1500K) and high pressure (>3atm) conditions, the calculation error exceeds 10%.

[0005] 3. Low reliability of component prediction: Existing machine learning models (such as CNN and random forest) lack physical constraints and are prone to prediction results that violate physical laws, such as non-conservation of total component amount and abnormal carbon-hydrogen ratio (CH2 / CH3<1.5 or>3.0).

[0006] 4. Inadequate verification methods: Most methods only verify a single indicator (such as RMSE) and have not established a multi-dimensional verification system for radiative transfer error (<2%), spatial resolution (maintaining the original 1m level), and spectral resolution (spectral angular distance <5°). Summary of the Invention

[0007] To address the aforementioned technical problems, an on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust is provided. This technical solution solves the problems mentioned in the background section.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] An on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust plumes includes:

[0010] The construction module is used to construct the exhaust spectral radiative transfer equation considering the coupling effect between gas molecules and alumina solid particles. The equation simultaneously includes gas absorption terms, particle scattering terms, and medium extinction terms.

[0011] The solution module is used to numerically solve the exhaust spectral radiative transfer equation using a spherical harmonic-discrete coordinate hybrid algorithm to obtain exhaust flow field parameters, wherein the spherical harmonic coordinate system is used to calculate the radiation source function, and the discrete coordinate system is used to iteratively update the radiation intensity in each direction;

[0012] The partitioning module is used to perform three-dimensional meshing of the exhaust flow field parameters. Each mesh point stores data on temperature, pressure, gas mole fraction, and alumina particle size distribution.

[0013] The preprocessing module is used to perform atmospheric correction on multispectral images and endmember extraction on hyperspectral images;

[0014] The fusion module is used to superimpose and fuse the preprocessed multispectral image and the hyperspectral image in the spectral channel dimension, and generate a fused image with both spatial resolution and spectral resolution through a generative adversarial network.

[0015] The prediction module is used to predict the content of CO2, H2O and unburned hydrocarbons in the exhaust smoke based on the fused image and a machine learning model that combines physical constraints.

[0016] Preferably, the steps for constructing the gas-solid coupled exhaust spectral radiative transfer equation specifically include:

[0017] Define the coupling relationship between the gas molecule emission term and the solid particle scattering term, where the gas emission term satisfies Kirchhoff's laws and the particle scattering term adopts the Mie scattering model;

[0018] By introducing the assumption of local thermodynamic equilibrium, the temperature fields of gas and particles are unified to ensure the consistency of energy transfer;

[0019] The comprehensive transport equation, which includes the absorption coefficient, extinction coefficient, and scattering albedo, is derived as follows:

[0020] ;

[0021] in, Representative wavelength is λ Direction is Ω, position is s The radiation intensity at that location Represents the total extinction coefficient, derived from the gas absorption coefficient. With particle extinction coefficient We get the weighted sum. = + , Represents the scattering phase function, describing the scattering from the direction Scattered to direction The probability distribution, and This represents the path length of radiation propagation.

[0022] Preferably, the numerical solution of the exhaust spectral radiative transfer equation using a spherical harmonic-discrete coordinate hybrid algorithm specifically includes:

[0023] Expand the radiation source function in spherically harmonic coordinates, retaining it up to the Lth order spherically harmonic function to ensure computational accuracy. satisfy:

[0024] ;

[0025] in, The absorption coefficient is... The scattering coefficient is... The blackbody radiation intensity. The average radiation intensity in the direction;

[0026] The radiation intensity of each grid point is iteratively updated along N detection directions in the discrete coordinate system. The Gauss-Seidel method is used to accelerate convergence in each iteration. The iterative formula is as follows:

[0027] ;

[0028] in, This is the nth discrete direction;

[0029] The convergence criterion is set as the difference in spectral radiance between two adjacent iterations being less than 10. −5 Or the maximum number of iterations reaches 10 4 Second-rate.

[0030] Preferably, the three-dimensional meshing of the exhaust flow field parameters specifically includes:

[0031] The computational domain size was determined based on the flow field data. A grid of 200 layers was created along the flight direction, and 50 layers each were created laterally and vertically. The grid size was [not specified]. = =0.1m, =0.05m;

[0032] Each grid point stores the following parameters: temperature T, measured and interpolated by a thermocouple; pressure P, measured and interpolated by a pressure sensor; mole fraction of CO2, measured and interpolated by a gas chromatograph; mole fraction of H2O, measured and interpolated by an infrared spectrometer; and density of alumina particles, measured and interpolated by a laser particle size analyzer.

[0033] The alumina particle size distribution was fitted using a log-normal distribution, with distribution parameters including the geometric mean diameter. Geometric standard deviation The fitting formula is:

[0034] .

[0035] Preferably, performing atmospheric correction on the multispectral image specifically includes:

[0036] Atmospheric correction was performed on the multispectral images to eliminate the effects of solar radiation and atmospheric scattering. The correction model adopted the MODTRAN atmospheric radiative transfer model, and the input parameters included observation geometry, atmospheric model, aerosol type, and visibility.

[0037] Preferably, the endmember extraction of the hyperspectral image specifically includes:

[0038] Minimum Noise Fraction (MNF) transform is performed on hyperspectral images to reduce the impact of noise;

[0039] In the data after MNF transformation, pure endmembers are found by iterative projection method;

[0040] The extracted endmembers were spectrally unmixed to obtain the abundance maps of each endmember.

[0041] Preferably, the step of superimposing and fusing the preprocessed multispectral image and the hyperspectral image in the spectral channel dimension specifically includes:

[0042] Ten spectral channels (wavelength range: 400-900nm, bandwidth: 50nm) of the multispectral image and 200 spectral channels (wavelength range: 400-900nm, bandwidth: 2.5nm) of the hyperspectral image are stitched together along the channel dimension to generate 210-channel fused input data.

[0043] A generator based on an improved VGG network is constructed. The max pooling layer of the original VGG is removed to maintain spatial resolution, and 210 input channels are set in the first convolutional kernel with a kernel size of 3×3 and a stride of 1.

[0044] A dual discriminator structure is adopted, which consists of a spatial discriminator and a spectral discriminator;

[0045] The generator and discriminator are trained using an adversarial loss function. The generator's objective is to minimize the following loss:

[0046]

[0047] in, For spatial discriminators, For spectral discriminators, For generator, For multispectral images, This is a hyperspectral image.

[0048] Preferably, the spatial discriminator consists of 5 convolutional layers, with the first 4 layers having a kernel size of 4×4 and a stride of 2, and the spectral discriminator consists of 5 convolutional layers, with the first 4 layers having a kernel size of 1×1 and a stride of 1, and the last layer having a kernel size of 1×1, outputting a spectral feature map.

[0049] Preferably, the loss function of the generative adversarial network consists of three parts:

[0050] Content loss: Calculate the spectral angular distance between the fused image and the original hyperspectral image in the key bands, using the following formula:

[0051] ;

[0052] in, and These are the spectral vectors of the fused image and the original hyperspectral image at the i-th pixel, respectively.

[0053] Spatial loss: Calculate the structural similarity index between the fused image and the original multispectral image, using the following formula:

[0054] ;

[0055] in, For the mean difference, For variance, and It is the stability constant;

[0056] Adversarial Loss: The fusion quality is improved by minimizing the cross-entropy loss between the generator and the bi-discriminator. The total generator loss is:

[0057] ;

[0058] in, =0.1, =1.0, =0.5.

[0059] Preferably, the physical constraint-based machine learning model construction step includes:

[0060] Extract the spectral features of the exhaust smoke and construct an input parameter set, including the excess air coefficient, water content and fuel density. The formula for calculating the excess air coefficient is: actual air volume / theoretical air volume.

[0061] Design a neural network architecture, including hidden layers, output layers, and physical constraint layers:

[0062] Hidden layers: consist of two fully connected layers, each containing 128 neurons, using the ReLU activation function;

[0063] Output layer: Composed of 4 neurons, the Sigmoid function is used to normalize the component prediction values ​​to the [0, 1] interval, which correspond to the mole fractions of CO2, H2O, CH4 and C2H6 respectively;

[0064] Physical constraint layer: Apply the following constraints after the output layer:

[0065] Total component constraints: This is automatically satisfied through the Softmax function;

[0066] Carbon-hydrogen ratio constraint: 1.5 ≤ CH2 / CH3 ≤ 3.0, by adding a penalty term:

[0067] accomplish;

[0068] Gas density matching constraints: ≤0.05 kg / m 3 By adding penalty items:

[0069] accomplish;

[0070] The total loss function is then:

[0071] ;

[0072] in, The mean square error between the predicted and actual values. , .

[0073] Compared with the prior art, the present invention provides an on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust, which has the following advantages:

[0074] This invention constructs a smoke spectral radiative transfer equation considering coupling effects, taking into account the coupling between gas molecules and alumina solid particles. This provides a more comprehensive description of the complex radiative transfer process in the smoke, more accurately reflecting the spectral radiation characteristics of the smoke and laying the foundation for subsequent precise analysis of smoke composition and flow field parameters. A spherical harmonic-discrete coordinate hybrid algorithm is used to solve the equation. By rationally allocating tasks between the two coordinate systems, the complex radiative transfer process in the smoke can be effectively handled while ensuring computational accuracy, obtaining accurate smoke flow field parameters, such as radiation intensity distribution, providing precise data for subsequent smoke composition analysis. Furthermore, three-dimensional mesh partitioning divides the smoke flow field into numerous small meshes, each storing detailed data such as temperature, pressure, gas mole fraction, and alumina particle size distribution. This allows for a more refined description of the physical characteristics of the smoke flow field, capturing subtle changes and features, and providing rich information for in-depth analysis of smoke composition and combustion processes. Multispectral images and hyperspectral images are combined... Superimposing and fusing images along spectral channels combines the high spatial resolution of multispectral images with the rich spectral resolution of hyperspectral images. The resulting fused image possesses both clear spatial details and detailed spectral information, enabling a more comprehensive and accurate description of the characteristics of exhaust smoke. Physically constrained machine learning models combine the advantages of physical laws and machine learning algorithms. They can utilize the rich spatial and spectral information in the fused image while adhering to the physical mechanisms of exhaust smoke composition changes. This avoids the overfitting or physical inconsistencies that may occur with purely machine learning models, thereby improving the accuracy of predicting the CO2, H2O, and unburned hydrocarbon content in exhaust smoke. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0076] 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.

[0077] Example 1

[0078] Please refer to Figure 1 As shown, an on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust plumes, the system 100 includes:

[0079] Module 110 is used to construct the exhaust spectral radiative transfer equation considering the coupling effect between gas molecules and alumina solid particles. The equation includes gas absorption terms, particle scattering terms, and medium extinction terms.

[0080] The solution module 120 is used to numerically solve the exhaust spectral radiative transfer equation using a spherical harmonic-discrete coordinate hybrid algorithm to obtain exhaust flow field parameters. The spherical harmonic coordinate system is used to calculate the radiation source function, and the discrete coordinate system is used to iteratively update the radiation intensity in each direction.

[0081] The partitioning module 130 is used to perform three-dimensional meshing of the exhaust flow field parameters. Each mesh point stores data on temperature, pressure, gas mole fraction, and alumina particle size distribution.

[0082] Preprocessing module 140 is used to perform atmospheric correction on multispectral images and endmember extraction on hyperspectral images;

[0083] The fusion module 150 is used to superimpose and fuse the preprocessed multispectral image and the hyperspectral image in the spectral channel dimension, and generate a fused image with both spatial resolution and spectral resolution through a generative adversarial network.

[0084] The prediction module 160 is used to predict the content of CO2, H2O and unburned hydrocarbons in the exhaust smoke based on the fused image and a machine learning model combined with physical constraints.

[0085] Those skilled in the art will understand that the exhaust spectral radiative transfer equation constructed in this invention, which considers the coupling effect between gas molecules and alumina solid particles, more comprehensively describes the complex radiative transfer process in the exhaust and more accurately reflects the spectral radiative characteristics of the exhaust, laying the foundation for subsequent precise analysis of exhaust composition and flow field parameters. The spherical harmonic-discrete coordinate hybrid algorithm is used to solve the equation. By rationally allocating the tasks of the two coordinate systems, the complex radiative transfer process in the exhaust can be effectively handled while ensuring computational accuracy, obtaining accurate exhaust flow field parameters, such as radiative intensity distribution, providing precise data for subsequent exhaust composition analysis. Furthermore, the three-dimensional mesh partitioning divides the exhaust flow field into numerous small meshes, with each mesh point storing detailed data such as temperature, pressure, gas mole fraction, and alumina particle size distribution. This allows for a more refined description of the physical characteristics of the exhaust flow field, capturing subtle changes and features within the flow field, providing rich information for in-depth analysis of exhaust composition and combustion processes. The integration of multispectral images with hyperspectral images... Superimposing and fusing images along spectral channels combines the high spatial resolution of multispectral images with the rich spectral resolution of hyperspectral images. The resulting fused image possesses both clear spatial details and detailed spectral information, enabling a more comprehensive and accurate description of the characteristics of exhaust smoke. Physically constrained machine learning models combine the advantages of physical laws and machine learning algorithms. They can utilize the rich spatial and spectral information in the fused image while adhering to the physical mechanisms of exhaust smoke composition changes. This avoids the overfitting or physical inconsistencies that may occur with purely machine learning models, thereby improving the accuracy of predicting the CO2, H2O, and unburned hydrocarbon content in exhaust smoke.

[0086] The specific steps for constructing the gas-solid coupled exhaust spectral radiative transfer equation include:

[0087] Define the coupling relationship between the gas molecule emission term and the solid particle scattering term, where the gas emission term satisfies Kirchhoff's laws and the particle scattering term adopts the Mie scattering model;

[0088] By introducing the assumption of local thermodynamic equilibrium, the temperature fields of gas and particles are unified to ensure the consistency of energy transfer;

[0089] The comprehensive transport equation, which includes the absorption coefficient, extinction coefficient, and scattering albedo, is derived as follows:

[0090] ;

[0091] in, Representative wavelength is λ Direction is Ω, position is s The radiation intensity at that location Represents the total extinction coefficient, derived from the gas absorption coefficient. With particle extinction coefficient We get the weighted sum. = + , Represents the scattering phase function, describing the scattering from the direction Scattered to direction The probability distribution, and This represents the path length of radiation propagation.

[0092] The numerical solution of the exhaust spectral radiative transfer equation using a hybrid spherical harmonic-discrete coordinate algorithm specifically includes:

[0093] Expand the radiation source function in spherically harmonic coordinates, retaining it up to the Lth order spherically harmonic function to ensure computational accuracy. satisfy:

[0094] ;

[0095] in, The absorption coefficient is... The scattering coefficient is... The blackbody radiation intensity. The average radiation intensity in the direction;

[0096] The radiation intensity of each grid point is iteratively updated along N detection directions in the discrete coordinate system. The Gauss-Seidel method is used to accelerate convergence in each iteration. The iterative formula is as follows:

[0097] ;

[0098] in, This is the nth discrete direction;

[0099] The convergence criterion is set as the difference in spectral radiance between two adjacent iterations being less than 10. −5 Or the maximum number of iterations reaches 10 4 Second-rate.

[0100] The specific steps for three-dimensional mesh generation of the exhaust flow field parameters include:

[0101] The computational domain size was determined based on the flow field data. A grid of 200 layers was created along the flight direction, and 50 layers each were created laterally and vertically. The grid size was [not specified]. = =0.1m, =0.05m;

[0102] Each grid point stores the following parameters: temperature T, measured and interpolated by a thermocouple; pressure P, measured and interpolated by a pressure sensor; mole fraction of CO2, measured and interpolated by a gas chromatograph; mole fraction of H2O, measured and interpolated by an infrared spectrometer; and density of alumina particles, measured and interpolated by a laser particle size analyzer.

[0103] The alumina particle size distribution was fitted using a log-normal distribution, with distribution parameters including the geometric mean diameter. Geometric standard deviation The fitting formula is:

[0104] .

[0105] Performing atmospheric correction on multispectral images specifically includes:

[0106] Atmospheric correction was performed on the multispectral images to eliminate the effects of solar radiation and atmospheric scattering. The correction model adopted the MODTRAN atmospheric radiative transfer model, and the input parameters included observation geometry, atmospheric model, aerosol type, and visibility.

[0107] Endmember extraction for hyperspectral images specifically includes:

[0108] Minimum Noise Fraction (MNF) transform is performed on hyperspectral images to reduce the impact of noise;

[0109] In the data after MNF transformation, pure endmembers are found by iterative projection method;

[0110] The extracted endmembers were spectrally unmixed to obtain the abundance maps of each endmember.

[0111] The process of overlaying and fusing the preprocessed multispectral image with the hyperspectral image along the spectral channel dimension specifically includes:

[0112] Ten spectral channels (wavelength range: 400-900nm, bandwidth: 50nm) of the multispectral image and 200 spectral channels (wavelength range: 400-900nm, bandwidth: 2.5nm) of the hyperspectral image are stitched together along the channel dimension to generate 210-channel fused input data.

[0113] A generator based on an improved VGG network is constructed. The max pooling layer of the original VGG is removed to maintain spatial resolution, and 210 input channels are set in the first convolutional kernel with a kernel size of 3×3 and a stride of 1.

[0114] A dual-discriminator structure is adopted, consisting of a spatial discriminator and a spectral discriminator.

[0115] The generator and discriminator are trained using an adversarial loss function. The generator's objective is to minimize the following loss:

[0116]

[0117] in, For spatial discriminators, For spectral discriminators, For generator, For multispectral images, This is a hyperspectral image.

[0118] The spatial discriminator consists of 5 convolutional layers, with the first 4 layers having a kernel size of 4×4 and a stride of 2. The spectral discriminator consists of 5 convolutional layers, with the first 4 layers having a kernel size of 1×1 and a stride of 1. The last layer has a kernel size of 1×1 and outputs a spectral feature map.

[0119] The loss function of a generative adversarial network consists of three parts:

[0120] Content loss: Calculate the spectral angular distance between the fused image and the original hyperspectral image in the key bands, using the following formula:

[0121] ;

[0122] in, and These are the spectral vectors of the fused image and the original hyperspectral image at the i-th pixel, respectively.

[0123] Spatial loss: Calculate the structural similarity index between the fused image and the original multispectral image, using the following formula:

[0124] ;

[0125] in, For the mean difference, For variance, and It is the stability constant;

[0126] Adversarial Loss: The fusion quality is improved by minimizing the cross-entropy loss between the generator and the bi-discriminator. The total generator loss is:

[0127] ;

[0128] in, =0.1, =1.0, =0.5.

[0129] The steps for building a machine learning model based on physical constraints include:

[0130] Extract the spectral features of the exhaust smoke and construct an input parameter set, which includes the excess air coefficient, water content and fuel density. The formula for calculating the excess air coefficient is: actual air volume / theoretical air volume.

[0131] Design a neural network architecture, including hidden layers, output layers, and physical constraint layers:

[0132] Hidden layers: consist of two fully connected layers, each containing 128 neurons, using the ReLU activation function;

[0133] Output layer: Composed of 4 neurons, the Sigmoid function is used to normalize the component prediction values ​​to the [0, 1] interval, which correspond to the mole fractions of CO2, H2O, CH4 and C2H6 respectively;

[0134] Physical constraint layer: Apply the following constraints after the output layer:

[0135] Total component constraints: This is automatically satisfied through the Softmax function;

[0136] Carbon-hydrogen ratio constraint: 1.5 ≤ CH2 / CH3 ≤ 3.0, by adding a penalty term:

[0137] accomplish;

[0138] Gas density matching constraints: ≤0.05 kg / m 3 By adding penalty items:

[0139] accomplish;

[0140] The total loss function is then:

[0141] ;

[0142] in, The mean square error between the predicted and actual values. , .

[0143] 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. An on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust plumes, characterized in that, include: The construction module is used to construct the exhaust spectral radiative transfer equation considering the coupling effect between gas molecules and alumina solid particles. The equation simultaneously includes gas absorption terms, particle scattering terms, and medium extinction terms. The solution module is used to numerically solve the exhaust spectral radiative transfer equation using a spherical harmonic-discrete coordinate hybrid algorithm to obtain exhaust flow field parameters, wherein the spherical harmonic coordinate system is used to calculate the radiation source function, and the discrete coordinate system is used to iteratively update the radiation intensity in each direction; The partitioning module is used to perform three-dimensional meshing of the exhaust flow field parameters. Each mesh point stores data on temperature, pressure, gas mole fraction, and alumina particle size distribution. The preprocessing module is used to perform atmospheric correction on multispectral images and endmember extraction on hyperspectral images; The fusion module is used to superimpose and fuse the preprocessed multispectral image and the hyperspectral image in the spectral channel dimension, and generate a fused image with both spatial resolution and spectral resolution through a generative adversarial network. The prediction module is used to predict the content of CO2, H2O and unburned hydrocarbons in the exhaust smoke based on the fused image and a machine learning model combined with physical constraints. The specific steps for constructing the gas-solid coupled exhaust spectral radiative transfer equation include: Define the coupling relationship between the gas molecule emission term and the solid particle scattering term, where the gas emission term satisfies Kirchhoff's laws and the particle scattering term adopts the Mie scattering model; By introducing the assumption of local thermodynamic equilibrium, the temperature fields of gas and particles are unified to ensure the consistency of energy transfer; The comprehensive transport equation, which includes the absorption coefficient, extinction coefficient, and scattering albedo, is derived as follows: ； in, Representative wavelength is λ Direction is Ω, position is s The radiation intensity at that location Represents the total extinction coefficient, derived from the gas absorption coefficient. With particle extinction coefficient We get the weighted sum. = + , Represents the scattering phase function, describing the scattering from the direction Scattered to direction The probability distribution, and The path length represents the radiation propagation. The numerical solution of the exhaust spectral radiative transfer equation using a hybrid spherical harmonic-discrete coordinate algorithm specifically includes: Expand the radiation source function in spherically harmonic coordinates, retaining it up to the Lth order spherically harmonic function to ensure computational accuracy. satisfy: ； in, The absorption coefficient is... The scattering coefficient is... The blackbody radiation intensity. The average radiation intensity in the direction; The radiation intensity of each grid point is iteratively updated along N detection directions in the discrete coordinate system. The Gauss-Seidel method is used to accelerate convergence in each iteration. The iterative formula is as follows: ； in, This is the nth discrete direction; The convergence criterion is set as the difference in spectral radiance between two adjacent iterations being less than 10. -5 Or the maximum number of iterations reaches 10 4 Second-rate; The specific steps of performing three-dimensional mesh generation on the exhaust flow field parameters include: The computational domain size was determined based on the flow field data. A grid of 200 layers was created along the flight direction, and 50 layers each were created laterally and vertically. The grid size was [not specified]. = =0.1m, =0.05m; Each grid point stores the following parameters: temperature T, measured and interpolated by a thermocouple; pressure P, measured and interpolated by a pressure sensor; mole fraction of CO2, measured and interpolated by a gas chromatograph; mole fraction of H2O, measured and interpolated by an infrared spectrometer; and density of alumina particles, measured and interpolated by a laser particle size analyzer. The alumina particle size distribution was fitted using a log-normal distribution, with distribution parameters including the geometric mean diameter. Geometric standard deviation The fitting formula is: 。 2. The on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust plumes according to claim 1, characterized in that, The atmospheric correction of the multispectral image specifically includes: Atmospheric correction was performed on the multispectral images to eliminate the effects of solar radiation and atmospheric scattering. The correction model adopted the MODTRAN atmospheric radiative transfer model, and the input parameters included observation geometry, atmospheric model, aerosol type, and visibility.

3. The on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust plumes according to claim 2, characterized in that, The specific steps of performing endmember extraction on the hyperspectral image include: Minimum Noise Fraction (MNF) transform is performed on hyperspectral images to reduce the impact of noise; In the data after MNF transformation, pure endmembers are found by iterative projection method; The extracted endmembers were spectrally unmixed to obtain the abundance maps of each endmember.

4. The on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust plumes according to claim 3, characterized in that, The process of superimposing and fusing the preprocessed multispectral image and the hyperspectral image along the spectral channel dimension specifically includes: Ten spectral channels of a multispectral image (wavelength range: 400-900nm, bandwidth: 50nm) and 200 spectral channels of a hyperspectral image (wavelength range: 400-900nm, bandwidth: 2.5nm) are stitched together along the channel dimension to generate 210-channel fused input data. A generator based on an improved VGG network is constructed. The max pooling layer of the original VGG is removed to maintain spatial resolution, and 210 input channels are set in the first convolutional kernel with a kernel size of 3×3 and a stride of 1. A dual discriminator structure is adopted, which consists of a spatial discriminator and a spectral discriminator; The generator and discriminator are trained using an adversarial loss function. The generator's objective is to minimize the following loss: ； in, For spatial discriminators, For spectral discriminators, For generator, For multispectral images, This is a hyperspectral image.

5. The on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust plumes according to claim 4, characterized in that, The spatial discriminator consists of 5 convolutional layers, with the first 4 layers having a kernel size of 4×4 and a stride of 2. The spectral discriminator consists of 5 convolutional layers, with the first 4 layers having a kernel size of 1×1 and a stride of 1. The last layer has a kernel size of 1×1 and outputs a spectral feature map.

6. The on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust smoke according to claim 5, characterized in that, The loss function of the generative adversarial network consists of three parts: Content loss: Calculate the spectral angular distance between the fused image and the original hyperspectral image in the key bands, using the following formula: ； in, and These are the spectral vectors of the fused image and the original hyperspectral image at the i-th pixel, respectively. Spatial loss: Calculate the structural similarity index between the fused image and the original multispectral image, using the following formula: ； in, For the mean difference, For variance, and It is the stability constant; Adversarial Loss: The fusion quality is improved by minimizing the cross-entropy loss between the generator and the bi-discriminator. The total generator loss is: ； in, =0.1, =1.0, =0.

5.

7. The on-chip integrated system for multispectral and hyperspectral composition analysis of aircraft exhaust plumes according to claim 6, characterized in that, The steps for building the physical constraint-based machine learning model include: Extract the spectral features of the exhaust smoke and construct an input parameter set, including the excess air coefficient, water content and fuel density. The formula for calculating the excess air coefficient is: actual air volume / theoretical air volume. Design a neural network architecture, including hidden layers, output layers, and physical constraint layers: Hidden layers: consist of two fully connected layers, each containing 128 neurons, using the ReLU activation function; Output layer: Composed of 4 neurons, the Sigmoid function is used to normalize the component prediction values ​​to the [0, 1] interval, which correspond to the mole fractions of CO2, H2O, CH4 and C2H6 respectively; Physical constraint layer: Apply the following constraints after the output layer: Total component constraints: This is automatically satisfied through the Softmax function; Carbon-hydrogen ratio constraint: 1.5 ≤ CH2 / CH3 ≤ 3.0, by adding a penalty term: accomplish; Gas density matching constraints: ≤0.05kg / m 3 By adding penalty items: accomplish; The total loss function is then: ； in, The mean square error between the predicted and actual values. , .

Images