Frame rate increase-based combustion temperature and component concentration rapid imaging measurement method

Abstract

The invention provides a combustion temperature and component concentration rapid imaging measurement method based on frame rate increase, a high-frame-rate laser-induced fluorescence image is obtained by using a deep learning model, and high-speed imaging of combustion temperature and component concentration is realized by fusing laser absorption spectrum data. A flame turbulence image is obtained according to fluid mechanics simulation, a fluorescence image frame interpolation model is obtained through training, and a high-frame-rate fluorescence image is generated through low-speed laser-induced fluorescence data. Obtaining temperature distribution of a combustion field by using laser absorption spectrum data, and substituting the temperature distribution into a fluorescence model satisfied by fluorescent molecules / free radical groups to obtain high-frame-rate gas molecule / free radical group concentration distribution; the concentrations of other products are solved according to a reaction kinetic equation, and the combustion field temperature and a multi-component concentration image are solved through iterative calculation. The frame rate of low-speed fluorescence data is increased according to turbulence motion estimation, the response speed of temperature and concentration distribution imaging is increased, and the method has wide application prospects in the field of combustion reaction monitoring research.

Description

Technical Field

[0001] The present invention provides a rapid imaging measurement method for combustion temperature and component concentration based on frame rate improvement, belonging to the field of combustion field parameter imaging, and involving two measurement techniques of laser absorption spectroscopy tomography imaging and planar laser induced fluorescence. Background Art

[0002] The parameter measurement of the combustion field is a very important scientific issue, and has many applications in engine design and monitoring, industrial production process control, chemical reaction mechanism analysis, environmental protection, etc. The temperature of the combustion field and the concentration of molecules are two very representative parameters. The former reflects the distribution and change trend of the thermal energy in the combustion field, and the latter reveals the distribution of the molecules participating in the combustion. The main measurement methods are divided into contact measurement methods and non-contact measurement methods.

[0003] Contact methods for measuring the temperature and concentration in a combustion field require the sensor to be in contact with the substance being measured. Representative methods include using thermocouples to measure temperature and gas analysis methods to measure the concentration of various molecules inside the combustion field. In 2023, Kim et al. described in the paper "Temperature variation measurements of an ignited energetic ionic liquid" published in Volume 84, Issue 1, pages 14 - 16 of "Science and Technology of Energetic Materials" the use of a fine thermocouple to measure the temperature change during the ignition process of an ionic liquid. The combustion process of the ionic liquid is very short, while the time constant of the thermocouple may be long, resulting in the inability to accurately capture the instantaneous high-temperature changes and the need for temperature compensation. Zhou et al. discussed the current research progress of high-temperature thermocouples in the paper "Review on High Temperature Thin Film Thermocouple for Aero-Engine" published in Volume 63, Issue 6, pages 82 - 102 of "Aeronautical Manufacturing Technology" in 2020. The article pointed out that in order to apply thermocouples to the high-temperature environment of the combustion field, a protective layer needs to be installed on them, which will reduce the response speed of the thermocouple. Kou et al. proposed a new type of air flow stagnation hood structure in the paper "Heat transfer error analysis of high-temperature wall temperature measurement using thermocouple" published in Volume 59, Document Number 104518 of "Case Studies In Thermal Engineering" in 2024, which can effectively avoid the contact between the air flow and the thermocouple, thus weakening the convection error. However, an increase in the diameter of the stagnation hood will lead to an increase in the thermocouple error and isolate the radiative heat exchange with the environment. Therefore, contact methods have problems of interfering with the combustion field and having a slow response speed.

[0004] The non-contact measurement method based on optical technology does not come into contact with the measured medium and does not disrupt the measured flow field. Representative technologies include Coherent anti-stokes Raman scattering spectrometry (CARS), Planar laser induced fluorescence (PLIF), Tunable diode laser absorption spectroscopy (TDLAS), etc. The CARS method measures the anti-Stokes Raman spectrum of components, establishes the relationship between the spectrum and temperature and concentration, and then calculates the combustion field parameters at the corresponding positions. In 2023, the paper "Coherent Anti-Stokes Raman Scattering Measurements of Time and Length Scales of Temperature Fluctuations in a Turbulent Flame" by Kobtsev et al. published on pages 482-490 of Volume 77, Issue 5 of "Applied Spectroscopy" proposed using two CARS spectrometers with synchronized nanosecond pulse repetition lasers to measure the time and spatial scales of "instantaneous" local temperature in a turbulent flame. The focal size and beam overlap limit the spatial resolution of the measurement, which may lead to the inability to accurately capture local temperature changes in turbulent flows. In 2024, the paper "Advances in Femtosecond Coherent Anti-Stokes Raman Scattering for Thermometry" by Song et al. published in "Photonics", Volume 11, Issue 7, Document No. 622 described the application progress of CARS spectral technology in high-temperature combustion temperature measurement. The CARS spectral system is relatively complex and expensive, limiting its popularization in practical applications. The CARS technology is suitable for measuring the temperature and component concentration at a single point in the combustion field. Its optical path and calculation process are both very complex, and the equipment is expensive, making it not suitable for the reconstruction of high-resolution images.In 2023, the paper "Dual-Pump Vibrational Coherent Anti-Stokes Raman Scattering System Developed for Simultaneous Temperature and Relative Nitrogen-Water Vapor Concentration Measurements" by Too et al. published in Optics, Volume 4, Issue 4, pages 613 - 624 proposed a method for simultaneously measuring temperature and relative water vapor-nitrogen concentration using two narrowband pump laser beams and one broadband Stokes laser beam. However, at high temperatures, the Raman linewidth of water vapor is affected by various factors, resulting in inaccurate results. Moreover, the local pulsation of turbulence will change the refractive index of the gas, which will in turn lead to the weakening or disappearance of the CARS signal. Therefore, CARS technology is difficult to apply to a combustion field with intense turbulence.

[0005] The planar laser-induced fluorescence method uses a laser beam to excite atoms or molecules to an excited state, collects the fluorescence spontaneously emitted when the excited atoms or molecules transition from the unstable excited state to a lower energy level in the direction perpendicular to the laser beam, and then uses the relationship between the energy spectrum information and temperature to achieve the measurement of temperature and concentration. In 2024, the paper "Nitric oxide laser-induced fluorescence using the fifth harmonic of a broad-band Nd:YAG laser" by Bolderman, et al. published in Applied Physics B-lasers And Optics, Volume 130, Issue 8, Article Number 148 proposed a method for NO laser-induced fluorescence imaging using a broadband Nd:YAG laser with non-tuned fifth harmonic radiation. In high-temperature combustion exhaust gases, the attenuation of the laser beam increases significantly, and spectral interference needs to be considered. The paper "266nm Laser-Induced Fluorescence Reference Spectra of Ketones and Aromatic Compounds" by Brunzendorf et al. published in Combustion Science and Technology, Volume 196, Issue 13, pages 2128-2137 in 2024 demonstrated high-precision laser-induced fluorescence spectra of acetone and other tracers under ambient conditions by using a large ultraviolet-sensitive SCMOS camera. In 2023, the paper "LIF-imaging of temperature and iron-atom concentration in iron nitrate doped low-pressure aerosol flat flames" by Apazeller et al. published in Applications In Energy And Combustion Science, Volume 16, Document Number 110199 described imaging the temperature and iron-atom concentration in an iron nitrate doped matrix burner low-pressure aerosol flat flame through laser-induced fluorescence imaging technology. Due to the complex fluorescence mechanism in the complex combustion field, the absolute concentration of iron could not be accurately determined.In 2022, the paper "Simultaneous Measurement of Concentration and Temperature in Methane-Hydrogen Flames by Laser-Induced Fluorescence" published by Moroshkina et al. in the Journal of Russian Laser Research, Volume 43, Issue 2, pages 181-185, described the analysis of the structure of stable methane-air and methane-hydrogen-air flames on a perforated-plate burner using the method of laser-induced fluorescence. The absorption effect made the fluorescence image stronger on the left side (the side where the laser was incident) and weaker on the right side. Therefore, there are difficulties in quantitative measurement in the measurement of the combustion field by planar laser-induced fluorescence technology, and the incident laser will be absorbed by the combustion field, causing the problem of uneven fluorescence images.

[0006] To achieve continuous high-speed PLIF diagnostics, a complex laser imaging system is required. The delicate operation and high experimental costs limit the penetration of this high-speed imaging technology. In addition, the frequency limitations of lasers and cameras also restrict PLIF applications. Generating high-speed imaging from low frequencies requires a time series prediction model. With the continuous development of deep learning technology, the effectiveness of laser induction in generating high-spatial-resolution image sequences has been proven. In 2023, Han et al. published a paper titled "Deep Neural Network-Based Generation of Planar CH Distribution through Flame Chemiluminescence in Premixed Turbulent Flame" in Volume 12, Article No. 100221 of "Energy and AI", proposing to use conditional generative adversarial networks to train mappings to extract high-temporal-resolution CH-PLIF images from low-temporal-resolution chemiluminescence images. In 2021, Zhang et al. published a paper titled "Generating planar distributions of soot particles from luminosity images in turbulent fames using deep learning" in Volume 127, Article No. 18 of "Applied Physics B", using conditional generative adversarial networks to predict planar laser-induced incandescence images from flame luminosity images to achieve the prediction of soot distribution in turbulent flames. In 2022, Guo et al. published a paper titled "High-speed planar imaging of OH radicals in turbulent flames assisted by deep learning" in Volume 128, Issue 3, Article No. 52 of "Applied Physics B-lasers And Optics", establishing a computational imaging model based on convolutional long short-term memory networks, successfully improving the diagnostic frequency of OH-PLIF images in turbulent flames. However, the PLIF images have all been binarized and cannot be used for further concentration calculations. Therefore, the current neural network training methods contain less flame physical information, and the generated PLIF images are difficult to meet the requirements for further temperature and concentration calculations.

[0007] The laser absorption spectroscopy method utilizes the relationship between laser absorption spectroscopy and combustion field parameters for measurement, featuring advantages such as fast measurement response speed, accurate quantitative measurement, and the ability to detect multiple parameters simultaneously, and is suitable for the quantitative measurement of combustion field parameters. Combining with tomography algorithms, TDLAS realizes the imaging of parameter distributions, but is usually limited by the optical path layout and it is difficult to obtain results with high spatial resolution. The article "Simultaneous measurement of temperature, concentration, and velocity in the combustion field using TDLAS" published by Hong et al. in the Journal of Mechanical Science and Technology, Volume 37, Issue 10, pages 5199–5206 in 2023 describes a measurement system for temperature, concentration, and velocity in the combustion field that uses a grid structure containing 35 laser sources. Each laser source is inserted into the measurement unit and transmits through the combustion field, measuring two-dimensional temperature and concentration images with an 8×8 resolution. In 2023, Pang et al. explored the influence of the beam arrangement on the reconstruction results of various distribution models in the paper "Optimization of beam arrangement for tunable diode laser absorption tomography reconstruction based on fractional Tikhonov regularization" published in Acta Physica Sinica, Volume 72, Issue 3, Article No. 037801, and performed two-dimensional reconstruction of the combustion field for a 10×10 discretized grid area using absorption spectral lines in the near-infrared band combined with 20 test beams. For example, the ammonia concentration measurement system described in the paper "Laser absorption tomography for ammonia measurement in diesel engine exhaust" published by Nasir et al. in Applied Physics B, Volume 126, Issue 11, pages 1–7 in 2020 measured ammonia concentration images with a 71×71 resolution through a TDLAS imaging system with 2130 beams at 6 angles.For example, the scanning temperature measurement device described in the paper "Towards a two-dimensional concentration and temperature laser absorption tomography sensor system" published by Kasyutich et al. in Applied Physics B, Vol. 102, No. 1, pp. 149-162 in 2011 measured a temperature distribution image with a resolution of 18×18 through an imaging system with 55 optical paths at 5 angles. Therefore, the laser absorption spectroscopy method can provide quantitative measurement results, but it is not easy to obtain high-resolution images due to the limitation of the optical path arrangement density in space.

[0008] Combining the measurement methods of PLIF and TDLAS, such as the measurement scheme proposed in the patent "A quantitative measurement method and device for transient combustion field" (authorization announcement number: CN110823849B), achieves high-resolution parameter distribution measurement. However, this measurement scheme does not utilize the known reaction mechanism when the fuel is single and clear and the combustion process is known, which limits the improvement of concentration accuracy. In 2024, Yudong Guo et al. published a paper titled "Using LAS to fuse Nd-YAG laser-induced fluorescence to measure temperature and H" in IEEE Transactions on Instrumentation and Measurement, Volume 73, Document No. 4503810, Pages 1-10. 2 O / CH 2 Simultaneous Quantitative Imaging of Temperature and H O Concentration 2 O / CH 2 A quantitative imaging method based on a reliable chemical reaction kinetic model combined with PLIF and LAS technology is proposed to simultaneously obtain the temperature and H concentration in the flame cross section. 2 O / CH 2 O concentration distribution. However, the current fusion imaging measurement method only realizes the fusion calculation of low-speed PLIF images and high-speed TDLAS signal synchronization images. The imaging frame rate is low and it is difficult to capture the details of dynamic flames.

[0009] The present invention proposes a rapid imaging measurement method for combustion temperature and component concentration based on frame rate improvement, and proposes a deep learning model based on convolutional neural network to generate high frame rate laser induced fluorescence images, improve the frame rate of fused images, and shorten the step size of multi-component concentration change required by kinetics. Summary of the Invention

[0010] The present invention proposes a rapid imaging measurement method for combustion temperature and component concentration based on frame rate improvement, which realizes the fusion of laser absorption spectroscopy and laser induced fluorescence spectroscopy data. The present invention decouples the temperature through the data of two laser absorption spectral lines, substitutes the temperature into the laser induced fluorescence model to obtain the concentration of fluorescent molecules / free radical groups. The relationship between fluorescent molecules / free radical groups and absorbing molecules is established by using the chemical reaction kinetics model to obtain high-resolution images of the two kinds of molecules, and then the high-resolution image of the temperature is obtained and iteratively adjusted. The specific implementation steps are as follows:

[0011] Step 1: Use the simulated combustion data as the training set of the time super-resolution frame interpolation model based on deep learning. The model training has two parts:

[0012] The first part is the optical flow calculation network, which calculates the bidirectional input optical flow forward optical flow F 0 and reverse optical flow F 1 between two input frames based on the pixel flow between the two input frames I 0→1 and I 1→0 .

[0013] This part of the network uses the U-Net architecture based on convolutional neural network (CNN) to capture long-distance motion and is optimized through end-to-end training. Take the direction of the optical flow between two input images, whether the directions are the same or opposite, and scale its amplitude accordingly. Then the intermediate optical flow at time t is approximately the combination of the bidirectional input optical flows

[0014]

[0015]

[0016] The second part is the optical flow interpolation network, which uses another U-Net to refine the optical flow and predict the visibility map. By applying the visibility maps V t←0 and V t←1 to the warped images, the contribution of occluded pixels to the interpolated intermediate frame is excluded. The optical flow approximation performs well in smooth regions, but performs poorly near motion boundaries because the motion near motion boundaries is not locally smooth. To solve the pixel occlusion problem and reduce the artifacts around the motion boundaries that may cause poor image synthesis, the optical flow interpolation CNN predicts two visibility maps V t←0 and Vt←1 , and satisfy the constraint:

[0017] V t←0 = 1 - V t←1 (3)

[0018] Adding this constraint prevents the network training from diverging. When V t←0 (p) = 0, V t←1 (p) = 1, and the pixel p of I 0 is occluded at T = t, the intermediate frame image does not contain I 0 . It is rare for a pixel to be occluded at time t by both time 0 and time 1.

[0019] The input images of the optical flow interpolation network are six: the input images I 0 and I 1 , the approximate optical flow and and the warped images g(I 0 , F t→0 ) and g(I 1 , F t→1 ) that take the approximate optical flow image as the input. g(I 0 , F t→0 ) and g(I 1 , F t→1 ) represent the interpolation frames synthesized in two directions respectively. g(·,·) is a backward translation function implemented by bilinear interpolation. These input images are used to generate and optimize the intermediate optical flow F t→1 and F t→0 , as well as the visibility maps V t←0 and V t←1 that contain occlusion information.

[0020] Then, the second part of the neural network is used for arbitrary temporal flow interpolation to further correct the intermediate optical flow and the non - linear learnable visualization mapping V t→0 , and finally, the intermediate optical flow and the input image are warped to a specific time step using the backward warping function, and the two warped images are linearly fused to generate the intermediate frame. The predicted intermediate frame image is

[0021]

[0022] Step 2: The temporal super - resolution frame interpolation model calculates data such as the intermediate optical flow and the visibility map based on the original low - frame - rate fluorescence image, fuses and calculates the intermediate frame image between two frames, and finally generates a high - frame - rate fluorescence image F(x, y) that conforms to the flame turbulence rule. According to the absorption spectrum data, the initial temperature image T 0 (x, y) is reconstructed using the tomography algorithm, and the initial image x of the fluorescence molecule / free radical concentration is calculated by combining the fluorescence image0 (x, y).

[0023] Assume that in the absorption spectrum measurement, M measurement optical paths are adopted. Each optical path starts from the vertex of the pentagonal sensor and reaches the infrared photodetector. The light intensity data obtained on the i-th optical path is I i (v), and the absorptivity on this optical path is calculated as

[0024]

[0025] where I 0 (v) represents the light intensity distribution when the laser just outputs from the collimating mirror, and generally it is replaced by a straight line obtained by fitting the positions on the I i (v) signal that are not affected by absorption. α i (v) represents the accumulation of the absorptivity on this optical path and is the integral of the absorptivity curve on the straight line. After obtaining the absorptivity curves α i (ν), i = 1, 2, …, M, it is necessary to further calculate the integral absorptivity on the straight line, that is

[0026]

[0027] In the formula, α i (v) has a shape similar to the Gaussian distribution, and the integral only needs to cover the spectrum within a few wavenumbers to ensure the calculation accuracy. In addition, in order to suppress the influence of noise, this integral is replaced by the fitting of the absorptivity curve α i (ν). When the absorption spectrum α i (v) is only affected by one absorption spectral line, the shape of its curve should be basically the same as that of the Voigt line shape function φ V (v). Take the parameters Δv D and Δv C that determine the Voigt line shape, and a C (i) that determines the area of the Voigt line shape as undetermined parameters, and fit the spectrum α i (ν) by the optimization method. The obtained a C (i) is theoretically equivalent to the result obtained by integration.

[0028] In the absorption spectrum tomography technology, obtaining the temperature distribution is the main task. In order to obtain the temperature distribution, first divide the space into grids. The size of the grids is the same as that of the fluorescence image. There are L r rows and L c columns for the grids, and there are N grids in total, that is, L r ×L c= N. By analyzing the relationship between the pentagonal sensor and the grid area, the lengths of the M measurement optical paths in the N grids are calculated, and thus the sensitivity matrix L of the pentagonal sensor is formed. The matrix L is an m×n matrix, where the element in the i-th row and j-th column represents the length of the i-th optical path in the j-th grid. If the i-th optical path does not pass through the j-th grid, the matrix element is zero. In order to solve the unit absorption rate a V (i), i = 1, 2, …, N, the equation

[0029] La v = A C . (7)

[0030] In the formula, the vector a v is a column vector containing N elements, and the i-th element a C (i) is the unit integrated absorption rate in the i-th grid; the vector A C is a column vector containing M elements, and the i-th element is the integral value of the absorption spectrum measured on the i-th optical path. To solve this equation, a simultaneous iterative reconstruction algorithm needs to be used. After solving the equation, the unit absorption rates a V (i), i = 1, 2, …, N in the N grids can be obtained.

[0031] The absorption rate b(j) per unit length in each grid is obtained by solving. To calculate the temperature, the tunable laser scans two absorption spectra. Let the absorption rate per unit length in the grid of the other spectral line be d(j), then the temperature T(j) in each small grid is calculated by the following formula.

[0032]

[0033] In the formula, h is the Planck constant, c is the speed of light, k is the Boltzmann constant, E″ d and E″ b are the low-level energies of the spectral lines corresponding to b(j) and d(j) respectively, and T 0 = 296K is the reference temperature. In this way, the initial temperature distribution image T(j) is obtained. Based on the temperature distribution image T(j), the resolution is expanded using the interpolation algorithm to obtain the initial temperature T 0 (x, y) with the same resolution as the fluorescence image F(x, y). According to the position of the grid in space, the spatial coordinates of the grid center point are obtained. Based on the two sets of variables of the grid position and the grid center coordinates, the temperature distribution T 0(x, y). This temperature image has quantitative temperature data sourced from LAS tomography, and the position of each pixel in space is given by the PLIF fluorescence image F(x, y). In space, each pixel in the formaldehyde fluorescence image F(x, y) corresponds to a position in the flame cross-section, and the corresponding pixel of T 0 (x, y) is spatially consistent with it.

[0034] Using a nanosecond-level ultra-narrow shutter and precise delay control technology to eliminate the influence of fluctuations in the quenching coefficient. At this time, from the temperature T 0 (x, y) and fluorescence intensity F 0 (x, y), the concentration X of the corresponding fluorescent molecules / free radical groups is calculated as 0 (x, y) is

[0035]

[0036] Step 3: Using the initial temperature distribution T 0 (x, y) and concentration distribution X 0 (x, y), H 0 (x, y) to calculate the net generation rate of the chemical reaction of water molecules and the mass fluxes of the conventional diffusion and thermal diffusion of water molecules Based on this, update the concentration distribution of water molecules to

[0037]

[0038] In the formula, ξ is the projected area along the diffusion direction of the region centered at (x, y) in the combustion field; H′(i, j) and T′ in the formula (i,j) represent the water molecule concentration and temperature results of the previous LAS imaging. Since the LAS system has a higher frame rate, the time interval Δt is not too high, generally less than 1 ms, and even lower than 1 μs.

[0039] The water molecule concentration image H obtained according to the chemical reaction kinetics model 1 (x, y) contains spectral data from both LAS and PLIF, as well as information on chemical reaction kinetics. After obtaining H 1 (x, y), the spectral information in it is fused into the temperature image, and the fusion method is to use the absorption model. The unit integral absorption rate distribution image A obtained by LAS tomography V1 (x, y) and A V2 (x, y), these two images provide the relationship between the water molecule concentration and temperature. The new temperature T 1 (x, y) needs to be obtained using an optimization method, and the optimization model is

[0040]

[0041] In this optimization problem, the variable to be optimized is T 1 (x, y). For each region T 1 (x, y), this is a single-variable optimization problem within an interval. The temperature T 1 (x, y) of the combustion field has clear upper and lower bounds, and the single-variable optimization problem is solved by methods such as the steepest descent method.

[0042] After completing the optimization to solve for the temperature T 1 (x, y), the imaging algorithm does not end. The fluorescence model includes constraints on temperature and formaldehyde concentration, and the new temperature T 1 (x, y) will bring a new concentration image, that is

[0043]

[0044] After obtaining the new formaldehyde concentration, the temperature T 1 (x, y), the concentration image H 1 (x, y) and X 1 (x, y) are basically self-consistent under the constraints of spectral data. However, under the new temperature and concentration distributions, it is necessary to recalculate the data in the chemical reaction kinetics model to form a new water molecule concentration image, and then obtain new temperature and formaldehyde concentration images.

[0045] This progressive image iteration enables the final image to meet the requirements of the three models. The following describes the complete process of the p-th iteration of the imaging algorithm. In the previous iteration, the obtained water molecule image was H p-1 (x, y), the temperature image was T p-1 (x, y), and the formaldehyde concentration image was X p-1 (x, y). Substitute the images of the previous iteration as the temperature and combustion field component concentration data into the chemical reaction kinetics model to calculate the net generation rate of the chemical reaction of water molecules and the mass fluxes of the conventional diffusion and thermal diffusion of water molecules Based on the water molecule imaging of the previous frame in the LAS system, calculate the high-resolution concentration imaging of water molecules as

[0046]

[0047] After obtaining the water molecule concentration image of this iteration, using the constraints on temperature and water molecule concentration in the laser absorption spectroscopy model, solving the optimization problem can obtain the new temperature distribution T p (x, y) as

[0048]

[0049] Finally, substituting the new temperature image into the laser-induced fluorescence model, the new formaldehyde concentration distribution can be obtained as

[0050]

[0051] In this way, the image iteration of this round is completed. For the termination condition of the iteration, it is completed by means of the relative distribution of two concentration images, and the iteration termination condition is set as

[0052]

[0053] In the formula, ε represents the threshold for terminating the iteration. When the relative change in the concentration images of water and formaldehyde is small enough, the image converges, and the obtained image is used as the final imaging result.

[0054] The effect of the present invention: By means of a deep learning model based on a convolutional neural network, a high-frame-rate laser-induced fluorescence image is generated, the frame rate of the fused image is increased, and high-frame-rate imaging measurement of the temperature in the combustion field and the concentration distributions of multiple components is realized. Description of the Drawings

[0055] Figure 1 is the flowchart for reconstructing the high-resolution temperature and concentration distributions of the present invention.

[0056] Figure 2 is the fluorescence image with a high frame rate obtained in the example, with a resolution of 1000×1000.

[0057] Figure 3 is the temperature distribution measured by laser absorption spectroscopy in the example, with a resolution of 50×50.

[0058] Figure 4 is the high-resolution temperature distribution obtained by iteration in the example.

[0059] Figure 5 is the high-resolution OH molecular concentration distribution obtained by iteration in the example. Detailed Embodiments

[0060] The present invention will be further described below in conjunction with examples. In the examples, the objects of laser absorption spectroscopy measurement are two spectral lines of water molecules at 7185.58 cm -1 and 7444.43 cm -1 The absorption spectroscopy sensor is a regular pentagon with a side length of 10 cm, and 11 photoelectric sensors are arranged on each side. A collimator is arranged at each vertex, and the laser emitted by the collimator irradiates 11 sensors on the opposite side and 6 sensors on each of the left and right sides. The laser-induced fluorescence system measures the Q of the OH group in the (A-X) transition branch 1The fluorescence spectrum of (5), the resolution of the laser-induced fluorescence image is 1000×1000, and 10 frames of data are obtained per second. The resolution of the reconstructed image of the laser absorption spectrum is 50×50, and 200 frames of data are obtained per second. There is a circular combustion area with a diameter of 3 cm in the center of the pentagonal sensor. The fuel is methane, and a diffusion flame is formed in the air.

[0061] Step 1: First, use the simulated combustion data as the training set for the time super-resolution frame interpolation model based on deep learning. During model training, first run the optical flow calculation network. Based on the pixel flow between two input frames I 0 and I 1 , calculate the bidirectional input optical flow, the forward optical flow F 0→1 and the backward optical flow F 1→0 . Use the second part of the neural network for arbitrary temporal flow interpolation to further correct the intermediate optical flow and the non-linear learnable visualization mapping V t→0 . Finally, use the inverse warping function to warp the intermediate optical flow and the input image to a specific time step to generate a high-frame-rate fluorescence image as shown in the appendix Figure 2 .

[0062] Step 2: According to the absorption data of the absorption spectrum ring sensor, reconstruct the low-resolution temperature distribution image. Calculate the absorptivity of two spectral lines from the light intensity signal received by the detector. Integrate the absorptivity to obtain the spectral line area. After solving the system of equations, obtain the absorptivity per unit length of the two spectral lines at 7185.58 cm-1 and 7444.43 cm-1 within the grid, and then calculate the temperature distribution within the grid. The result is shown in the appendix Figure 3 . Substitute the temperature image T 0 (x, y) and the concentration image H 0 (x, y) of the OH group into the chemical reaction kinetics model to obtain the high-resolution images of the water molecule concentration and temperature.

[0063] Step 3: According to the high-resolution images of temperature and fluorescence particle concentration, in accordance with the chemical reaction kinetics model, correct the water molecule concentration distribution. According to the corrected absorption molecule concentration, solve for the new temperature distribution and the concentration distribution of the OH group. Iterate the temperature distribution, the concentration distribution of the OH group, and the fluorescence intensity distribution until the optimization termination condition is reached. The temperature distribution at this time is shown in the appendix Figure 4 , and the concentration distribution of the OH group is shown in the appendix Figure 5 , completing the solution of the high-resolution temperature and concentration images.

[0064] The above description of the present invention and its implementation manners is not limited thereto. What is shown in the drawings is only one of the implementation manners of the present invention. Without departing from the gist of the present invention, structures or embodiments similar to this technical solution designed without creativity all fall within the protection scope of the present invention.

Claims

1. A method for rapid imaging measurement of combustion temperature and component concentration based on frame rate improvement, characterized in that A deep learning model is used to obtain high-frame-rate fluorescence images from low-frame-rate laser-induced fluorescence data, and the combustion field temperature distribution is obtained using laser absorption spectrum data. The fluorescence model satisfied by the fluorescent molecules / free radicals is substituted to obtain the high-frame-rate gas molecule / free radical concentration distribution. The concentrations of other products are calculated based on the reaction kinetics equation, and the combustion field temperature and multi-component concentration images are solved by iterative calculation. The steps of this method are as follows: Step 1: Establish a fluorescence image frame interpolation model; establish a combustion model based on the size of the burner, the type of fuel used, and the amount of fuel consumed, the model includes different grids extending from the center of the burner to the periphery, and generates temperature concentration distribution data of the burner at different combustion moments; use the data as a training set to establish a time super-resolution frame interpolation model based on deep learning, based on the original fluorescence image with a low frame rate, deform the front and back images to a specific time step, adaptively fuse the two distorted images to generate an intermediate image, and obtain a high frame rate fluorescence image F(x,y) that conforms to the flame turbulence rules; Step 2: Solve the temperature and concentration distribution; input the low frame rate fluorescence image into the frame interpolation model to obtain a high frame rate fluorescence image, use the tomography algorithm to reconstruct the initial temperature image T0(x,y) and the molecular concentration image H0(x,y) corresponding to the absorption spectrum according to the laser absorption spectrum data, and substitute the high frame rate fluorescence image and the initial temperature image into the fluorescence model satisfied by the initial image X0(x,y) of the fluorescent molecule / free radical concentration: In the formula, is the calibration factor related to system performance and experimental arrangement, η P hν represents the signal size generated by collecting one photon of light from the camera that captures the fluorescence. f is the photon energy, N A is Avogadro's constant, R is the ideal gas constant, Ω represents the solid angle of the fluorescence received by the camera in the area with the center coordinate (x, y), B 12 and B 21 The Einstein coefficient for stimulated emission, A 21 represents the spontaneous emission rate of particles excited to high energy levels, l represents the length of the region along the laser path, Λ represents the cross-sectional area of ​​the laser illuminated region, P represents the standard atmospheric pressure, these parameters are calibrated in real time in the same experiment using a standard flame formed by a standard burner with known temperature and OH group distribution, P(x,y) represents the pressure distribution, and f(T0(x,y)) represents the Boltzmann distribution; Step 3: Iteratively reconstruct the temperature and concentration distribution; obtain the concentration relationship between molecules according to the chemical reaction kinetics equation, substitute the temperature image T0(x, y) and the fluorescent molecule / free radical concentration image X0(x, y) into the chemical reaction kinetics model, and calculate the net generation rate R of the chemical reaction of the absorbing molecule L (x,y) and mass flux of conventional diffusion and thermal diffusion Thus, the absorption spectrum corresponding to the molecular concentration H1(x,y) is updated. Where ξ is the projection area of ​​the area centered at (x, y) in the combustion field along the diffusion direction, Δt is the time interval, and W is H The molar mass of water molecules is represented by substituting the new temperature distribution into the fluorescence model to obtain a new distribution of fluorescent molecules / free radicals. This round of image iteration is completed. The termination condition of the iteration is completed with the help of the relative distribution of the two concentration images. The termination condition of the iteration is set to: Where ε represents the threshold for terminating the iteration. When the relative change of the concentration images of water and formaldehyde is less than the threshold, the final iterative imaging result is obtained.

2. A method for rapid imaging measurement of combustion temperature and component concentration based on frame rate improvement, characterized in that The optical flow neural network is used to estimate the distribution of fluid concentration over time, thereby realizing super-time resolution imaging and improving the estimation accuracy of flame turbulence. The fluid motion estimation algorithm of the neural network is used to make up for the defects of the optical flow algorithm. This method includes three steps for frame interpolation of low frame rate fluorescence images. The specific method is as follows: Step 1: Use the simulation data of the temperature and concentration distribution of the experimental working conditions as the training set to generate a frame interpolation model based on deep learning; in order to solve the problem that the training input data of the fluid mechanics simulation distribution changes greatly due to turbulence, a batch normalization step is added after each layer of the convolutional neural network output data to divide the training set into multiple small batches of data B. Each batch of data contains several training set images, and all training images contain m pixel values ​​p1…p m ; Calculate the average pixel value of this batch of data respectively and variance Normalize the values ​​in the data. ε is a very small amount to prevent the denominator from being 0, and is set to 10 -10 : Then scale and translate each pixel value to obtain a distribution close to the original image and reduce the loss of feature information: In the formula, h i is the normalized pixel value, γ and β are learnable scaling parameters and translation parameters. After adding γ and β, the weights of all flame turbulence features are more balanced to prevent some of the effects from being too large. The network adaptively learns scaling and translation parameters for each neuron to improve the pixel value distribution while retaining the learning results of each neuron. The addition of multiple learning parameters requires a sufficient number of training sets, and the fluid mechanics simulation data set meets the requirements. The normalized pixel values ​​are trained batch by batch to obtain the frame interpolation model; Step 2: Calculate the velocity vector diagram of the fluorescent molecules / free radicals in the two fluorescence images in the low frame rate fluorescence image sequence, including the eddy diffusion coefficient; the velocity vector network calculates the velocity vector of the input images I0 and I1 at time t = 0 or 1, extracts the small-scale vortex structure in the image that has an important influence on the energy and mass transfer, and optimizes it through end-to-end training to obtain the predicted forward velocity vector F 0→1 and the reverse velocity vector F 1→0 , the intermediate velocity vector at time t takes the direction between the two input images, the same or opposite direction, and scales its amplitude accordingly, then the intermediate velocity vector distribution is approximately In the formula, To predict the forward intermediate velocity vector, To predict the reverse intermediate velocity vector; Step 3: Realize the interpolation of fluorescence images at any time; first calculate the deformed image using the approximate fluorescence velocity vector image as input and and represent the interpolated frames synthesized in two directions, respectively, and g(·,·) is a backward translation function implemented by bilinear interpolation; in order to solve the pixel occlusion problem and reduce artifacts around the moving boundaries that may cause poor image synthesis, the interpolation model predicts two fluorescent molecule / free radical occlusion relationship graphs V t←0 and V t←1 Used to distort images and satisfy: V t←0 =1-V t←1 , (8) Where V t←0 (p) = 0 means V t←1 (p) = 1, pixel p from I0 is occluded at T = t. It is rare for a pixel to be occluded at both time 0 and time 1 at time t. The fluorescence image frame interpolation network approximates the intermediate velocity vector by inputting images I0 and I1 and And the deformed image g(I0,F t→0 ) and g(I1,F t→1 ), generate and further correct the intermediate velocity vector F t→0 、F t→1 and fluorescent molecule / free radical occlusion relationship image V t→0 、V t→1 , the velocity gradient loss is added to the training loss function to satisfy the energy conservation constraint of the uniform flow field, the intermediate velocity vector and the input image are deformed to a specific time step using the reverse warping function, and the two images are linearly fused to generate an intermediate frame, and the predicted intermediate frame fluorescence image is calculated. for In the image-based interpolation method, there is a potential temporal consistency principle, that is, the closer the interpolation time t is to T = 1, the greater the contribution of frame I1 to the interpolated frame; due to the calibration coefficient C e The calibration coefficient C corresponding to the intermediate frame fluorescence image changes with each fluorescence image acquisition. e,t According to the C corresponding to the two frames of fluorescence images I0 and I1 e,0 With C e,1 To update: C e,t =tC e,1 +(1-t)C e,0 (10)。

Images