Ultrahigh-temperature multi-band material emissivity non-contact measurement method and system

Abstract

The invention discloses a non-contact measurement method and system for emissivity of an ultrahigh-temperature multiband material. The non-contact measurement method comprises the following steps: acquiring a transient multispectral radiance image sequence of the surface of a measured material; constructing a spatio-temporal evolution perception neural network model with an embedded micro-optical calculation layer, and training by adopting a strategy based on physical information constraint; and inputting an image sequence acquired in real time into the trained model, outputting a spectral emissivity estimation value, and calculating the surface true temperature in combination with the measured radiation brightness. According to the method, the deep learning model embedded in the physical mechanism is constructed, the irregular distortion characteristic of the emissivity in the dynamic ablation process of the material is directly learned, the technical problem that a traditional temperature measurement method fails due to model mismatch is solved, and the accuracy and robustness of true temperature and emissivity decoupling are improved.

Description

Technical Field

[0001] This application relates to the field of non-contact thermophysical parameter measurement technology, specifically to a non-contact measurement method and system for emissivity of ultra-high temperature multi-band materials. Background Technology

[0002] In extreme service scenarios such as long-duration atmospheric reentry of hypersonic vehicles and advanced aero-engine combustion chambers, thermal protection structural components typically employ ultra-high temperature materials such as carbon / carbon (C / C) composites or silicon carbide-based (C / SiC) composites. These materials face stringent challenges in thermal protection performance under the combined effects of extreme temperatures exceeding 2500°C and high-speed airflow. Therefore, accurate and real-time measurement of the surface true temperature and spectral emissivity of materials under actual service conditions is crucial for evaluating their thermal protection performance, validating aerodynamic thermal environment calculation models, and ensuring the operational safety of equipment.

[0003] Currently, among non-contact measurement methods, multi-band radiation thermometry is the mainstream technique for obtaining the true temperature and spectral emissivity of material surfaces under such extreme environments. This method is based on Planck's blackbody radiation law. It measures the spectral radiance of the material surface at multiple discrete wavelengths and combines this with a pre-defined material spectral emissivity model to construct a set of equations, thereby calculating the material's surface temperature. Common emissivity models include those assuming the material is a gray body (i.e., emissivity does not change with wavelength), or assuming emissivity changes linearly with wavelength, or follows some exponential or polynomial law, etc.

[0004] However, in actual hypersonic aerodynamic heating or combustion environments, the surface state of composite materials such as C / C or C / SiC is not static or quasi-static. Under the combined effects of high-temperature oxidation and high-speed gas flow shear, complex non-equilibrium thermochemical reactions occur on the material surface. For example, the silicon carbide (SiC) component in the material matrix is ​​rapidly oxidized to form a liquid silica (SiO2) molten oxide film. Under the influence of strong gas flow shear and surface tension, this oxide film continuously undergoes a series of violent dynamic evolution processes, including formation, flow, bubble precipitation, local rupture, and vaporization. These drastic changes in the microscopic surface topology, such as uneven fluctuations in liquid film thickness, random generation and expansion of ruptured pores, and transient changes in chemical composition, directly cause the macroscopic spectral emissivity curve of the material to no longer follow any smooth, monotonous physical laws, but instead exhibit highly nonlinear, rapidly changing, and irregular distortion characteristics. In this context, any traditional inversion algorithm based on fixed physical model assumptions (such as gray body models or linear models) will produce significant fundamental errors due to the severe mismatch between the model and actual physical conditions. This could even lead to unsolvable inversion equations or the calculation of temperature values ​​that do not conform to physical laws, thus failing to meet the engineering requirements for high-precision measurements. Therefore, accurately achieving decoupled measurement of surface true temperature and spectral emissivity during this complex non-equilibrium process involving the dynamic evolution of the oxide film and the coupling of substrate ablation on the material surface is a current technical challenge in this field. Summary of the Invention

[0005] One aspect of this invention is to provide a non-contact measurement method for the emissivity of ultra-high temperature multi-band materials, which solves the technical problem that existing technologies cannot accurately achieve decoupled measurement of surface true temperature and emissivity when the surface of ultra-high temperature composite materials undergoes dynamic oxide film evolution and matrix ablation coupling, resulting in severe distortion of the spectral emissivity characteristic curve that is non-monotonic and irregular.

[0006] This invention provides a non-contact method for measuring the emissivity of ultra-high temperature multi-band materials, the method comprising the following steps: Step one involves constructing a multi-band high spatiotemporal resolution radiation image acquisition system and using this system to acquire a sequence of transient multispectral radiance images of the surface of the material under test. Specifically, the system includes a multi-band optical acquisition unit, which comprises a multi-channel spectral imaging device or an array of multiple narrowband filter high-speed cameras. Its operating bands are designed to cover the visible to near-infrared region, for example, selecting 8 to 16 discrete narrowband bands within the 0.5 to 1.7 micrometer spectral range. Furthermore, the system is equipped with a high-precision synchronous trigger controller to ensure that the image sensors of all bands expose and acquire data from the same field of view of the target at exactly the same time, thereby obtaining a transient multispectral radiance image sequence composed of N bands. This sequence is represented as L_meas(λ_i, x, y, t), where λ_i represents the center wavelength of the i-th band, (x, y) are the spatial pixel coordinates in the image, t is the sampling time, and N is the total number of bands. Preferably, the acquisition system also includes an optical path attenuation and background radiation filtering module designed for ultra-high temperature environments, in order to prevent the detector from becoming saturated due to excessive radiation signals and to effectively suppress measurement interference caused by ambient stray light.

[0007] Step two involves establishing a database of dynamic ablation radiation characteristics of the materials, which will be used for training subsequent neural network models. First, in a laboratory environment, using equipment such as an arc wind tunnel, a high-frequency plasma generator, or a high-power laser heating device, extreme aerodynamic thermal environments similar to those of hypersonic vehicle reentry or engine combustion chambers are simulated to test the heating of carbon / carbon (C / C), carbon / silicon carbide (C / SiC), or ultra-high temperature ceramic (UHTC) based composite thermal protection material samples containing zirconium diboride (ZrB2) or hafnium diboride (HfB2). Throughout the heating test, the multispectral radiance image sequence of the material surface is simultaneously recorded using the acquisition system described in Step one. Simultaneously, reference surface true temperature data for local areas of the material surface is obtained through auxiliary measurement methods such as pre-embedded high-temperature thermocouples in the samples, active laser absorption spectroscopy probes, or real-time comparison with a standard blackbody, denoted as T_ref(t). Based on Planck's law, its expression is L(λ,T) = ε(λ, T) × C1 × λ^(-5) × (exp(C2 / (λ × T)) - 1)^(-1), where L is radiance, ε is emissivity, T is true temperature, λ is wavelength, and C1 and C2 are Planck's radiation constants. Using the known reference surface true temperature T_ref(t) and the synchronously measured multispectral radiance L_meas(λ_i, t), the corresponding reference spectral emissivity data ε_ref(λ_i, t) can be calculated. Finally, the data pairs containing multispectral radiance image sequences, surface dynamic texture features (such as visual information like liquid film rupture and bubble formation), and corresponding surface true temperature-emissivity labels are organized and labeled to construct a training dataset covering the entire life cycle evolution of various materials from initial heating, oxide film formation, liquid film flow, rupture to vaporization and volatilization.

[0008] Step 3: Construct a spatiotemporal evolution-aware emissivity inversion neural network model. This invention constructs a deep neural network model, named Dynamic Emissivity Reconstruction Network (DER-Net). This network model is used to directly learn the nonlinear mapping relationship between the microscopic physical state of a material surface and its macroscopic spectral emissivity from multi-band radiation data. Its structure includes a spatiotemporal feature extraction module, a spectral correlation analysis module, and an emissivity generation decoder. The spatiotemporal feature extraction module uses a three-dimensional convolutional neural network (3D-CNN) or a convolutional long short-term memory network (ConvLSTM) as its backbone. This module receives several consecutive frames of multi-band radiance images as input to extract the dynamic texture features of the material surface evolving over time. These dynamic texture features include the flow velocity of the oxide film, the expansion rate of the fractured region, and the collapse frequency of bubbles. These spatiotemporal features implicitly represent the physicochemical state of the material surface at the current moment and are key evidence for determining the emissivity distortion mode. The spectral correlation analysis module can use a self-attention mechanism to process spectral data across different bands, capturing cross-band nonlinear correlations caused by oxide film interference effects or absorption peaks of specific components, thereby generating more concise spectral feature vectors.

[0009] Specifically, to improve the interpretability and physical consistency of the model, the emissivity generation decoder described in this invention is designed as a parameterized generation network based on a differentiable physical rendering layer. This network does not directly output an unconstrained numerical vector of spectral emissivity, but instead outputs a set of microscopic physical parameters that determine the radiation properties of the material surface. In one embodiment, these microscopic physical parameters include the equivalent thickness d(x,y,t) of the oxide film, the surface roughness factor σ, and the complex refractive index n-ik of the gas-liquid two-phase material (where i is the imaginary unit, n is the refractive index, and k is the extinction coefficient). Subsequently, a fixed, non-trainable differentiable optical computation layer is embedded in the network structure. This differentiable optical computation layer solidifies the optical physical model, namely the transfer matrix method (TMM) of thin-film optics or the mathematical formulas used to describe the effective medium theory for composite media. During the forward propagation of the network, this differentiable optical computing layer receives microscopic physical parameters (such as d, σ, n, k, etc.) predicted by the preceding network and calculates the multi-band spectral emissivity vector uniquely corresponding to these physical parameters, denoted as ε = [ε(λ_1), ε(λ_2), ..., ε(λ_N)], strictly following the optical physics laws embedded within it. This design directly incorporates physical mechanisms as hard constraints into the network structure, ensuring that any emissivity curve output by the network mathematically conforms to the physical laws of thin-film interference and dielectric radiation, eliminating the possibility of generating predictions that do not conform to physical laws.

[0010] Step four involves training the DER-Net described in step three using a joint training strategy based on physical information constraints. This invention also optimizes the model training process by introducing a loss function with physical information constraints to address potential physical violations that may occur when a purely data-driven model lacks sufficient training data or faces unknown operating conditions. The total loss function is defined as: Loss_total = L_data + alpha × L_physics + beta × L_temporal, where alpha and beta are weight coefficients. Here, L_data is the data-driven loss, used to calculate the mean square error between the predicted emissivity ε of the network's final output and the reference emissivity ε_ref in the database from step two, driving the model to fit the known true data. L_physics is the physical consistency loss, constrained based on the self-consistency of Planck's law. Specifically, a parallel temperature prediction branch is added to the network to predict the true surface temperature T_pred. Then, the network-predicted spectral emissivity ε and the true surface temperature T_pred are substituted into Planck's formula to calculate the theoretical multispectral radiance L_calc, and the error between L_calc and the actual measured input radiance L_meas is calculated. This loss term forces the network to predict emissivity and temperature, which must be physically explainable to the observed radiant energy, thus enhancing the model's generalization ability in unlabeled data regions. L_temporal is a spatiotemporal continuity loss, used to impose a smoothing constraint on the rate of change of the predicted emissivity field over time. This means that unless the spatiotemporal feature extraction module detects a drastic topological change event (such as a sudden large-scale rupture of the liquid film), the change in the emissivity field should be continuous, which helps suppress the interference of measurement noise on the results. Preferably, a curriculum learning strategy can be used during training. The network is first pre-trained using simple operating conditions with low temperatures and before drastic fluctuations in the oxide film, and then gradually fine-tuned by adding data from high-temperature conditions, liquid film rupture, and complex vaporization scenarios to improve the model's convergence speed and final robustness.

[0011] Step 5: Decouple online measurement from true temperature measurement. Deploy the trained DER-Net model on a real-time processing terminal, such as an embedded graphics processing unit (GPU) module. In actual measurement tasks, such as hypersonic wind tunnel tests or engine ground tests, acquire multi-band radiation image sequences of the surface of the material under test in real time. Input the acquired image sequences into the deployed DER-Net model in real time. The model first identifies the dynamic state of the current material surface through its internal spatiotemporal feature extraction module, such as whether it is in the stage of complete oxide film coverage or the stage of partial liquid film rupture. Based on the identified state, the model then outputs the nonlinear spectral emissivity estimate ε(λ_i, x, y, t) for each pixel at the current time and the preliminary true temperature prediction T_pred. Finally, using the actual radiance L_meas(λ_i) measured at this time and the high-precision emissivity ε(λ_i) predicted by the network, the Planck equation is solved by least squares optimization to obtain the final true temperature T_true of the material surface. The mathematical expression for this optimization process is: T_true = argmin_T ∑_{i=1}^{N} || L_meas(λ_i) - ε(λ_i) × L_BB(λ_i, T) ||^2, where L_BB(λ_i, T) is the Planck blackbody radiation function. In this optimization process, using the temperature T_pred directly predicted by the network as the initial value of the optimization algorithm can accelerate the iterative convergence process and effectively avoid getting trapped in local optima.

[0012] Another aspect of the present invention provides a non-contact measurement system for emissivity of ultra-high temperature multi-band materials. This system includes: a multi-band optical acquisition unit, corresponding to the multi-band radiation image acquisition system described above, used for spectral imaging of the target on the same optical axis to acquire narrowband radiation images with different center wavelengths. This unit preferably employs an aperture-segmented or amplitude-segmented multi-aperture camera array, or a high-speed spectroscopic camera based on an acousto-optic tunable filter (AOTF); and an intelligent computing processing unit, which has a pre-trained dynamic emissivity reconstruction network (DER-Net). This unit possesses high-performance tensor computation capabilities, used to input the transient multispectral radiance image sequence into the DER-Net to obtain the spectral emissivity estimate output by the DER-Net, and using the spectral emissivity estimate and the transient multispectral radiance image sequence, to solve the Planck equation to obtain the true temperature of the material surface.

[0013] The beneficial effects of this invention lie in that, by constructing a spatiotemporal evolution-aware deep learning model, it no longer relies on idealized emissivity physical models such as gray volumes or linear models. Instead, it directly learns the complex mapping relationship between the microscopic surface state of materials and macroscopic spectral emissivity from data. This effectively solves the technical problem of traditional temperature measurement methods failing due to irregular distortions in emissivity during the dynamic oxidation and ablation of composite materials such as C / SiC, C / C, and ultra-high temperature ceramics. By introducing a spatiotemporal feature extraction mechanism, the method of this invention can perceive the dynamic evolution history information of the material surface, thereby accurately distinguishing between radiation fluctuations caused by temperature changes and radiation mutations caused by abrupt changes in physical states such as surface oxide film rupture, improving the accuracy of decoupling surface true temperature and emissivity. Furthermore, by introducing Planck's law constraint into the loss function and embedding a differentiable optical computation layer into the network structure, the output results of the model are ensured to strictly follow the fundamental laws of thermal radiation physics, avoiding non-physical predictions that may arise from purely data-driven methods. This gives the model clear physical interpretability and improves the generalization ability and robustness of the algorithm under extreme high-temperature unknown conditions. The method and system provided by this invention are applicable to extreme environments with intense aerodynamic-thermal-chemical reaction coupling, such as the reentry phase of hypersonic vehicles and the combustion chamber of ramjet engines, providing a reliable measurement means for the performance evaluation and structural safety assessment of thermal protection materials. Attached Figure Description

[0014] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0015] Figure 1 This is a schematic diagram of the structure of a non-contact measurement system for emissivity of ultra-high temperature multi-band materials provided in an embodiment of the present invention.

[0016] Figure 2 This is a flowchart of a non-contact measurement method for emissivity of ultra-high temperature multi-band materials provided in an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of the structure of the Dynamic Emittivity Reconstruction Network (DER-Net) in an embodiment of the present invention.

[0018] Figure 4 This is a schematic diagram illustrating the principle of the joint loss function based on physical information constraints in an embodiment of the present invention.

[0019] Figure 5 This is a schematic diagram of the multi-band radiance image acquisition and synchronization timing in an embodiment of the present invention.

[0020] Figure 6 This is a schematic diagram of the differentiable physical rendering layer calculation process of the emissivity generation decoder in an embodiment of the present invention.

[0021] Figure 7This is a schematic diagram illustrating the model training using a course learning strategy in an embodiment of the present invention.

[0022] Figure 8 This is a schematic diagram of the optical path structure of the multi-band optical acquisition unit in an embodiment of the present invention.

[0023] Figure 9 This is a schematic diagram illustrating the dynamic evolution of the oxide film on the surface of the C / SiC composite material in an embodiment of the present invention.

[0024] Figure 10 This is a schematic diagram of the spatial distribution of the temperature field and emissivity field on the surface of the material under test in an embodiment of the present invention, wherein (a) is the temperature field distribution and (b) is the emissivity field distribution. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0026] Example 1 This embodiment provides a non-contact method for measuring the emissivity of ultra-high temperature multi-band materials, addressing the technical challenge of irregular distortion of spectral emissivity caused by dynamic oxidation and ablation of the material surface, which renders traditional temperature measurement methods ineffective. (Refer to...) Figure 2 The method includes the following steps: Step S1: Construct a multi-band high spatiotemporal resolution radiation image acquisition system and use this system to acquire a sequence of transient multispectral radiance images of the surface of the material under test. (Refer to...) Figure 1 The system 100 includes a multi-band optical acquisition unit 110. Specifically, the multi-band optical acquisition unit 110 includes a multi-channel spectral imaging device or an array of multiple high-speed cameras. Each camera is equipped with a narrowband filter with a different center wavelength, and the same field of view is acquired through a beam splitter or beam splitter array. Preferably, to achieve a higher degree of optical path coaxiality, a single sensor combined with a high-speed switching filter wheel or an acousto-optic tunable filter (AOTF) can be used. The spectral operating band of the system is designed to cover the visible to near-infrared region, for example, selecting 8 to 16 discrete narrowband bands in the spectral range of 0.5 micrometers to 1.7 micrometers. This band range selection takes into account the advantages of high radiation signal intensity from ultra-high temperature objects and high quantum efficiency of silicon-based detectors.

[0027] like Figure 5As shown, this embodiment of the invention employs a high-precision synchronous trigger controller to achieve synchronous acquisition of multi-band images. The system is equipped with a synchronous trigger controller located on the left side, which sends synchronous trigger pulses to all band channels at preset times (t0, t1, t2, t3, t4, etc.). The multiple band channels include bands λ1 (0.6 μm) to λN (1.6 μm), covering 8 to 16 discrete narrowband bands selected within a spectral range of 0.5 to 1.7 micrometers. At each trigger moment, the image sensors of all bands simultaneously respond to the trigger signal, exposing and acquiring the same field of view on the surface of the material under test, obtaining radiation images at the same time. The gray rectangles in the figure represent the exposure intervals of each band; all exposure intervals are strictly aligned on the time axis, demonstrating that the system's time synchronization accuracy can reach the microsecond level (<1 μs). Through this high-precision synchronization mechanism, the system can acquire transient multispectral radiance image sequences L_meas(λ_i, x, y, t), ensuring the spatiotemporal correspondence between images of different bands and providing a reliable data foundation for subsequent multispectral data processing and temperature field inversion. The red vertical alignment line marks the strict alignment state of all bands at the trigger time t1, and the "simultaneous exposure" characteristic described in the specification is fully reflected in the timing.

[0028] Furthermore, the system is equipped with a high-precision synchronous trigger controller to ensure that image sensors of all bands expose and acquire data from the same field of view on the surface 101 of the measured material at exactly the same time. The time synchronization accuracy can reach the microsecond level, thereby obtaining a transient multispectral radiance image sequence composed of N bands. This sequence is represented as L_meas(λ_i, x, y, t), where λ_i represents the center wavelength of the i-th band, the value of i ranges from 1 to N, (x, y) are the spatial pixel coordinates in the image, t is the sampling time, and N is the total number of bands. Preferably, the acquisition system also includes an optical path attenuation and background radiation filtering module designed for ultra-high temperature environments. For example, multiple switchable neutral density attenuators are set in the optical path to adapt to a wide dynamic range of radiation signals from room temperature to over 2500°C, avoiding detector saturation due to excessively strong radiation signals. At the same time, appropriate bandpass filters are used to effectively suppress measurement interference caused by stray light such as plasma glow and laser radiation that may exist in the environment.

[0029] Step S2: Establish a database of dynamic ablation radiation characteristics of the material. This database forms the basis for subsequent neural network model training. First, in a laboratory environment, using equipment such as an electric arc wind tunnel, a high-frequency plasma generator, or a high-power laser heating device, the extreme aerodynamic thermal environment of hypersonic vehicle reentry or engine combustion chamber is simulated to test the heating of ultra-high temperature ceramic (UHTC) based thermal protection material samples, including carbon / carbon (C / C), carbon / silicon carbide (C / SiC), or those containing zirconium diboride (ZrB2) or hafnium diboride (HfB2). Throughout the heating test, the acquisition system described in step S1 is used to simultaneously record a sequence of multispectral radiance images of the material surface.

[0030] Meanwhile, to obtain the true label for supervised learning, it is necessary to acquire the reference surface true temperature data of a local area on the material surface through auxiliary measurement methods, denoted as T_ref(t). Specifically, a high-temperature thermocouple can be pre-embedded inside the sample or on the back side near the measurement area, and the surface temperature can be obtained by correcting it using a heat transfer model. Alternatively, advanced non-contact measurement methods such as active laser absorption spectroscopy probes can be used to directly acquire surface temperature information. Based on Planck's law, its expression is L(λ, T) = ε(λ, T) × C1 × λ^(-5) × (exp(C2 / (λ × T)) - 1)^(-1), where L is radiance, ε is emissivity, T is true temperature, λ is wavelength, and C1 and C2 are Planck's radiation constants. Using the known reference surface true temperature T_ref(t) and the simultaneously measured multispectral radiance L_meas(λ_i, t), the corresponding reference spectral emissivity data ε_ref(λ_i, t) can be calculated. Finally, the collected data were organized and labeled to form data pairs containing multispectral radiance image sequences, surface dynamic texture features (such as visual information like liquid film rupture areas and bubble generation locations labeled by image segmentation algorithms), and corresponding surface true temperature-emissivity labels. This constructed a training dataset covering the entire life cycle evolution of various materials from initial heating, oxide film formation, liquid film flow, rupture to vaporization and volatilization.

[0031] Step S3: Construct a spatiotemporal evolution-aware emissivity inversion neural network model. This invention constructs a deep neural network model, named Dynamic Emissivity Reconstruction Network (DER-Net). (Refer to...) Figure 3 This dynamic emissivity reconstruction network is used to directly learn the nonlinear mapping relationship between the microscopic physical state of a material surface and its macroscopic spectral emissivity from multi-band radiation data. Its structure includes a spatiotemporal feature extraction module 310, a spectral correlation analysis module 320, and an emissivity generation decoder 330.

[0032] Specifically, the spatiotemporal feature extraction module 310 employs a three-dimensional convolutional neural network (3D-CNN) or a convolutional long short-term memory network (ConvLSTM) as its backbone network. This module receives a series of consecutive frames (e.g., a time window of 5 frames) of multi-band radiance images as input to extract the dynamic texture features of the material surface evolving over time. For example, the flow pattern of the oxide film can be captured by a group of convolutional kernels similar to optical flow, and higher-order spatiotemporal information such as the expansion rate of the fractured region and the collapse frequency of bubbles can also be characterized by the deep network. These spatiotemporal features implicitly represent the physicochemical state of the material surface at the current moment and are key criteria for determining the emissivity distortion mode.

[0033] The spectral correlation analysis module 320 employs a self-attention mechanism to process spectral data from different bands. This module can dynamically assign different weights to different bands, capturing cross-band nonlinear correlations caused by oxide film interference effects or absorption peaks of specific components. For example, when the oxide film thickness varies, the radiance at different wavelengths will exhibit periodic constructive or destructive interference. The self-attention mechanism can effectively learn this complex dependency, thereby generating a more condensed spectral feature vector.

[0034] To enhance the interpretability and physical consistency of the model, the emissivity generation decoder 330 described in this invention is designed as a parameterized generation network based on a differentiable physical rendering layer. This network does not directly output an unconstrained numerical vector of spectral emissivity, but instead outputs a set of microscopic physical parameters 331 that determine the radiation properties of the material surface. In one embodiment, these microscopic physical parameters include the equivalent thickness d(x,y,t) of the oxide film, the surface roughness factor σ, and the complex refractive index n-ik of the gas-liquid two-phase material (where i is the imaginary unit, n is the refractive index, and k is the extinction coefficient). Subsequently, a fixed, non-trainable differentiable optical computation layer 332 is embedded within the network structure. This layer solidifies optical physical models, such as the Transfer Matrix Method (TMM) for calculating thin-film interference effects or mathematical formulas from the Effective Medium Theory for describing the optical constants of composite media. During the forward propagation of the network, the differentiable optical computing layer 332 receives the microscopic physical parameters (such as d, σ, n, k, etc.) predicted by the preceding network and calculates the multi-band spectral emissivity vector uniquely corresponding to these physical parameters, denoted as ε = [ε(λ_1), ε(λ_2), ..., ε(λ_N)], strictly following the optical physical laws embedded within it. This design directly incorporates the physical mechanism as a hard constraint into the network structure, ensuring that any emissivity curve output by the network conforms mathematically to the physical laws of thin-film interference and dielectric radiation, thus avoiding the possibility of predictions that do not conform to physical laws.

[0035] like Figure 6As shown, the emissivity generation decoder 330 adopts a parameterized generation architecture based on a differentiable physical rendering layer. This architecture is divided into three functional modules: First, the feature vector output by the front layer of the neural network is fed into the microphysical parameter prediction layer. This layer is a trainable layer and is responsible for predicting four key microphysical parameters that determine the radiation properties of the material surface, including the equivalent thickness d(x,y,t) of the oxide liquid film, the surface roughness factor σ, and the complex refractive index n and extinction coefficient k of the gas-liquid two-phase material. Second, the predicted microphysical parameters are fed into the differentiable optical computation layer. This layer is a non-trainable layer and has embedded optical physical models inside, including the transfer matrix method (TMM) for calculating the optical properties of thin films and the effective medium theory (EMT) for calculating the properties of effective media. These physical constraints are hard-coded into the network structure. Finally, the differentiable optical computation layer strictly calculates according to the physical formulas of TMM and EMT and outputs a multi-band spectral emissivity vector ε=[ε(λ1),ε(λ2),...,ε(λN)]. Because the entire computation process is differentiable, gradients can propagate backward from the output to the microscopic physical parameters, enabling the network to learn parameter distributions that conform to physical laws during training. This ensures that the generated emissivity data is physically plausible and interpretable. This design significantly differs from traditional end-to-end black-box neural networks, effectively improving the model's generalization ability and prediction accuracy by incorporating prior physical knowledge into the network architecture.

[0036] Step S4: Train the DER-Net described in step S3 using a joint training strategy based on physical information constraints. (Refer to...) Figure 4 This invention optimizes the model training process by introducing a loss function constrained by physical information. The total loss function is defined as: Loss_total = L_data + alpha × L_physics + beta × L_temporal, where alpha and beta are weighting coefficients used to balance the contributions of different loss terms.

[0037] Here, L_data is the data-driven loss, used to calculate the mean squared error or other distance metric between the predicted emissivity ε of the network's final output and the reference emissivity ε_ref in the database of step S2. This loss term drives the model to fit the known ground truth data and is the foundation of supervised learning.

[0038] L_physics represents the physical consistency loss. It is constrained by the self-consistency of Planck's law. Specifically, a parallel temperature prediction branch is added to the network to predict the true surface temperature T_pred. Then, the network-predicted spectral emissivity ε and the true surface temperature T_pred are substituted into Planck's formula to calculate the theoretical multispectral radiance L_calc, and the error between L_calc and the actual measured input radiance L_meas, such as the mean square error, is calculated. This loss term forces the network to predict emissivity and temperature, which must be physically explainable to the observed radiant energy. This allows the model to self-supervise and optimize even on unlabeled data, enhancing its generalization ability in unlabeled data regions.

[0039] L_temporal is the spatiotemporal continuity loss, used to impose a smoothing constraint on the rate of change of the predicted emissivity field over time. For example, it can calculate the L1 or L2 norm of the difference between the predicted emissivity fields of two consecutive frames. The physical basis of this constraint is that, unless the spatiotemporal feature extraction module detects a drastic topological abrupt event (such as a sudden large-area rupture of a liquid film), the physical state and emissivity of the material surface should change continuously and smoothly. This loss term helps suppress the interference of measurement noise on the results, making the output more stable.

[0040] Preferably, a curriculum learning strategy can be adopted during the training process. The network is first pre-trained using simple working conditions with low temperature and no drastic fluctuations in the oxide film. After the model initially converges, data from complex scenarios such as high temperature, liquid film rupture, and vaporization are gradually added for fine-tuning to improve the convergence speed and final robustness of the model.

[0041] like Figure 7 As shown, this invention employs a curriculum learning strategy for model training, dividing the training process into two stages to improve convergence speed and robustness. In stage one (pre-training stage), simple operating conditions with stable oxide films at low temperatures (below 1500°C) are used for initial model training. This stage has a low task difficulty, allowing the model to quickly learn basic temperature field characteristics. After the model initially converges under simple conditions, it enters stage two (fine-tuning stage). In this stage, complex scenario data such as liquid film rupture and vaporization at high temperatures (above 1500-2500°C) are introduced, significantly increasing the task difficulty. The stepped curve in the figure clearly illustrates the gradual increase in task difficulty during training, with two black dots marking key nodes in the stage transition. Through this progressive training strategy from easy to difficult, the model can gradually improve its adaptability to complex operating conditions while ensuring convergence stability, ultimately achieving an effective improvement in temperature measurement accuracy across the entire temperature range.

[0042] Step S5 involves decoupling online measurement from true temperature measurement. The trained DER-Net model is deployed on a real-time processing terminal, such as an embedded graphics processing unit (GPU) module. In actual measurement tasks, such as hypersonic wind tunnel tests or engine ground tests, multi-band radiation image sequences of the surface of the material under test are acquired in real time. The acquired image sequences are input into the deployed DER-Net model in real time. The model first identifies the dynamic state of the current material surface through its internal spatiotemporal feature extraction module, such as whether it is in the stage of complete oxide film coverage or the stage of local liquid film rupture. Based on the identified state, the model then outputs the current time, the nonlinear spectral emissivity estimate ε(λ_i, x, y, t) corresponding to each pixel, and the preliminary surface true temperature prediction T_pred.

[0043] Finally, using the actual radiance L_meas(λ_i) measured at that moment and the high-precision emissivity ε(λ_i) predicted by the network, the Planck equation is solved using nonlinear least squares optimization to obtain the final true temperature T_true of the material surface. The mathematical expression for this optimization process is: T_true = argmin_T ∑_{i=1}^{N} || L_meas(λ_i) - ε(λ_i) × L_BB(λ_i, T) ||^2, where L_BB(λ_i, T) is the Planck blackbody radiation function. In this optimization process, using the temperature T_pred directly predicted by the network as the initial value for the optimization algorithm's iteration can accelerate the iterative convergence process and effectively avoid getting trapped in local optima, ensuring the real-time performance and accuracy of the solution.

[0044] Example 2 This embodiment provides a non-contact measurement system for the emissivity of ultra-high temperature multi-band materials, which is used to perform the method described in Embodiment 1. (Refer to...) Figure 1 The system 100 includes: The multi-band optical acquisition unit 110, corresponding to the multi-band radiation image acquisition system described in step S1 of Embodiment 1, is used to perform spectral imaging on the surface 101 of the material under test along the same optical axis to acquire narrowband radiation images with different center wavelengths. This unit preferably employs an aperture-segmented or amplitude-segmented multi-aperture camera array, or a high-speed spectroscopic camera based on an acousto-optic tunable filter (AOTF), to ensure high spatiotemporal resolution and data acquisition synchronization.

[0045] like Figure 8As shown, the multi-band optical acquisition unit 110 includes a front lens, an optical path attenuation module, a beam splitter, a narrowband filter array, and a detector array. Thermal radiation emitted from the surface 101 of the material under test is focused by the front lens and then attenuated by an ND attenuator to prevent detector saturation. The attenuated beam enters the beam splitter and is decomposed into four beams with different propagation directions. Each beam passes through a corresponding narrowband filter (λ1, λ2, λ3, λ4), filtering out narrowband spectra with different center wavelengths. Finally, these are received by the detector arrays of each channel and converted into image signals. Through this optical path design, the system can simultaneously acquire narrowband radiation images of multiple bands on the same optical axis, providing a data foundation for subsequent multispectral temperature field inversion. The entire optical path structure is compact, with each optical element arranged sequentially on the optical axis, ensuring the stability of the optical system and the imaging quality.

[0046] The data synchronization and preprocessing unit 120 may include a synchronization trigger controller and a data acquisition card in hardware, and a preprocessing module in software. This unit is used to ensure precise temporal synchronization of image data across different bands and to perform preprocessing operations such as radiometric calibration, non-uniformity correction, and image registration. Radiometric calibration converts the digital grayscale values ​​output by the camera into physically meaningful spectral radiance values; non-uniformity correction eliminates the effects of inconsistent detector pixel responses; and image registration precisely aligns images from different optical paths or cameras to eliminate geometric distortions that may be caused by differences in optical paths, ensuring that each pixel corresponds to the same point on the material surface across all bands.

[0047] The intelligent computing processing unit 130 has a high-performance graphics processing unit (GPU) or field-programmable gate array (FPGA) as its hardware core. This unit incorporates a pre-trained dynamic emissivity reconstruction network (DER-Net), which possesses high-performance tensor computation capabilities for real-time execution of deep learning inference tasks. It receives image sequences from the data synchronization and preprocessing unit 120, executes the calculation process described in step S5 of Embodiment 1, reconstructs the dynamic emissivity field from the input radiation image sequence, and calculates the true surface temperature field.

[0048] The system also includes a results visualization and storage unit 140, typically a workstation or industrial computer equipped with a monitor and data storage devices. This unit displays in real-time the temperature distribution cloud map, emissivity distribution cloud map, and temperature-time history curves of key monitoring points on the surface of the tested material, calculated by the intelligent computing processing unit 130. Furthermore, this unit can set temperature thresholds, provide audible or visual alerts for abnormally high temperatures, and store all raw data and calculation results for subsequent analysis.

[0049] Example 3 This embodiment uses a thermal evaluation test of a silicon carbide-based (C / SiC) composite material conducted in a plasma wind tunnel as an application background to specifically illustrate the implementation process of the method of this invention. In this test, when the surface temperature of the material exceeds 2000 degrees Celsius, the silica oxide film formed on its surface enters a molten flow state. Under the shearing action of high-speed airflow, the thickness distribution of this liquid film layer is uneven, and dynamic phenomena such as local rupture, bubble formation and collapse occur. This process results in a highly non-uniform spatial spectral emissivity of the material surface, and rapid and irregular temporal changes. Any traditional multi-band temperature measurement method based on a fixed emissivity model (such as a gray body or linear model) cannot obtain an accurate temperature field because the model is seriously inconsistent with physical reality. Therefore, it is necessary to adopt the method described in this application.

[0050] like Figure 9 As shown, under hypersonic aerodynamic heating or combustion conditions, the surface of the C / SiC composite material undergoes a dramatic dynamic evolution process. The bottom layer in the figure represents the C / SiC matrix material 801, upon which a SiO2 liquid oxide film 806 of uneven thickness is formed. Under the shearing action of high-speed airflow, the oxide film exhibits complex morphological changes: in the stable liquid film region 802, the oxide film is relatively intact with a wavy surface profile; in the fractured region 803, cracks appear in the oxide film, partially exposing the matrix material; in the bubble region 804, bubbles precipitated inside the oxide film can be observed; and in the vaporization region 805, located at the edge of the oxide film, vaporization and volatilization occur. The liquid film thickness d dynamically changes due to the combined effects of surface tension and airflow shearing. This dramatic change in the microscopic surface topology results in a highly nonlinear, rapidly changing, and irregularly distorted macroscopic spectral emissivity curve, making it difficult for traditional spectrophotometric thermometry methods to obtain accurate temperature measurements in such scenarios.

[0051] In this embodiment, a multi-band optical acquisition unit with eight channels (center wavelength covering the range of 0.6 micrometers to 1.6 micrometers) is first used to synchronously acquire a sequence of multispectral radiance images L_meas(λ_i, x, y, t) of the specimen surface during the ablation process at a frame rate of 100 Hz. The acquired image sequence is fed in real time into an intelligent computing processing unit that has deployed a pre-trained DER-Net model. The DER-Net's spatiotemporal feature extraction module receives multiple consecutive frames of images and identifies the physical state of different regions in the image by analyzing the spatiotemporal evolution of image texture and brightness, such as stable liquid film regions, liquid film flow regions, and substrate exposure regions caused by liquid film rupture.

[0052] Based on these spatiotemporal dynamic characteristics, the model's emissivity generator outputs a pixel-by-pixel spectral emissivity vector ε(λ_i, x, y, t) that matches the physical state of each region. Specifically, in the stable liquid film region, the emissivity curve output by the network exhibits oscillating characteristics caused by thin-film interference effects; while in the exposed substrate region, the output emissivity closely approximates the characteristics of the C / SiC material itself. Subsequently, the system utilizes this high-precision emissivity field ε and the radiance field L_meas measured at that moment to optimize and solve the Planck equation using the least squares method, ultimately calculating the true temperature distribution T_true(x, y, t) of the material surface. The temperature field obtained by this method can clearly distinguish the hotspot regions formed by localized oxide film peeling, where the temperature is higher than the surrounding liquid film-covered area, providing data support for evaluating the uniformity and reliability of the material's thermal protection performance.

[0053] like Figure 10 As shown, the surface of the C / SiC composite material exhibits non-uniform spatial distribution characteristics of temperature field and emissivity field under high-temperature oxidation environment. Figure 10 (a) shows the temperature field distribution T(x,y) on the material surface. It can be observed that the central region has a higher temperature, appearing as a dark gray area, with a temperature of approximately 1400K to 1600K; while the edge region has a relatively lower temperature, appearing as a light gray area, with a temperature of approximately 1200K to 1300K. The three locations marked with dashed circles in the temperature field are hot spots where the substrate is exposed. These areas have significantly higher temperatures than the surrounding liquid film-covered areas due to the exposure of the substrate caused by oxide film rupture.

[0054] Figure 10 (b) illustrates the corresponding emissivity field distribution ε(x,y). The distribution of the emissivity field corresponds to the temperature field: in the stable liquid film region completely covered by the oxide film, the material surface emissivity is low, approximately 0.75 to 0.81; while in the substrate-exposed area (hot spot) marked by the dashed circle, the emissivity of the SiC substrate is higher than that of the oxide film, reaching 0.87 to 0.93. The liquid film flow region, as a transitional area, exhibits an intermediate emissivity value. The DER-Net deep learning network in this invention can accurately identify these regions with different characteristics by learning the spatial correspondence between the temperature field and the emissivity field, and output an accurate emissivity distribution map, thus providing a reliable basis for subsequent real temperature inversion. This emissivity identification method based on spatial distribution characteristics overcomes the limitation of traditional single-point measurements in failing to capture the non-uniformity of the material surface.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A non-contact method for measuring the emissivity of ultra-high temperature multi-band materials, characterized in that, Includes the following steps: Step 1: Obtain a sequence of transient multispectral radiance images of the surface of the material under test; Step 2: Establish a database of dynamic ablation radiation characteristics of materials, which includes multispectral radiance image sequences and corresponding surface true temperature-emissivity labels. Step 3: Construct a spatiotemporal evolution-aware emissivity inversion neural network model. The model includes a spatiotemporal feature extraction module, a spectral correlation analysis module, and an emissivity generation decoder. The emissivity generation decoder is a parameterized generation network based on a differentiable physical rendering layer, configured to output a set of microscopic physical parameters that determine the radiation characteristics of the material surface. Furthermore, a fixed, non-trainable differentiable optical computation layer is embedded within the emissivity generation decoder. This layer embeds an optical physical model, receives the microscopic physical parameters, and calculates a multi-band spectral emissivity vector uniquely corresponding to each microscopic physical parameter based on the optical physical model. Step four: Train the spatiotemporal evolution sensing emissivity inversion neural network model; Step 5: The trained spatiotemporal evolution-aware emissivity inversion neural network model is used for online measurement. Specifically, the real-time acquired transient multispectral radiance image sequence is input into the spatiotemporal evolution-aware emissivity inversion neural network model to obtain the spectral emissivity estimate output by the spatiotemporal evolution-aware emissivity inversion neural network model. The true temperature of the material surface is obtained by solving the Planck equation using the spectral emissivity estimate and the transient multispectral radiance image sequence.

2. The method according to claim 1, characterized in that, Step one, which involves acquiring the transient multispectral radiance image sequence of the surface of the material under test, specifically includes: The surface of the material under test is synchronously exposed and acquired using a multi-band optical acquisition unit. The multi-band optical acquisition unit includes a multi-channel spectral imaging device or an array of cameras configured with narrowband filters of different center wavelengths, and its working band covers the visible light to near-infrared region.

3. The method according to claim 1, characterized in that, The spatiotemporal feature extraction module uses a three-dimensional convolutional neural network or a convolutional long short-term memory network as the backbone network to receive several consecutive frames of images from the transient multispectral radiance image sequence as input, and extracts the dynamic texture features of the material surface as it evolves over time. The dynamic texture features include the flow velocity of the oxide film, the expansion rate of the fractured area, or the collapse frequency of bubbles.

4. The method according to claim 1, characterized in that, The spectral correlation analysis module uses a self-attention mechanism to process spectral data across different bands, in order to capture cross-band nonlinear correlations caused by oxide film interference effects or absorption peaks of specific components, and generate spectral feature vectors.

5. The method according to claim 1, characterized in that, The microscopic physical parameters include the equivalent thickness of the oxide liquid film, the surface roughness factor, and the complex refractive index of the gas-liquid two-phase material; the optical physical model includes the transfer matrix method of thin film optics or the effective medium theory used to describe composite media.

6. A non-contact measurement system for the emissivity of ultra-high temperature multi-band materials, characterized in that, include: A multi-band optical acquisition unit is used to acquire transient multispectral radiance image sequences of the surface of the material under test; The intelligent computing processing unit is equipped with a spatiotemporal evolution-aware emissivity inversion neural network model constructed and trained according to any one of claims 1-5. The intelligent computing processing unit is used to input the transient multispectral radiance image sequence into the spatiotemporal evolution-aware emissivity inversion neural network model to obtain the spectral emissivity estimate output by the spatiotemporal evolution-aware emissivity inversion neural network model, and use the spectral emissivity estimate and the transient multispectral radiance image sequence to obtain the true temperature of the material surface by solving the Planck equation.

Images