A multispectral true temperature inversion method without blackbody calibration

By coupling the instrument's spectral response with the target emissivity in the original voltage domain, and combining Planck's radiation law and smoothing constraints, the dependence of multispectral thermometry on emissivity models and blackbody calibration is solved, achieving high-precision temperature inversion under calibration-free conditions, which is suitable for extreme environments.

CN123111473APending Publication Date: 2026-11-17DALIAN MARITIME UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611196854.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing multispectral temperature measurement technologies rely on emissivity models and blackbody calibration, which cannot achieve accurate temperature measurement in extreme environments. In particular, they suffer from measurement errors and ambiguities in absolute temperature scales, especially in scenarios where on-site calibration is not possible.

Method used

By coupling the instrument spectral response with the target emissivity into an equivalent spectral response function, an inversion problem is constructed in the original voltage domain. Combining Planck's radiation law and smoothing constraints, a total loss function is constructed, and a weak temperature prior is introduced to resolve global scale ambiguity, thus achieving absolute true temperature inversion without blackbody calibration.

Benefits of technology

It requires no preset emissivity model or full-band blackbody furnace calibration, and can achieve high-precision, noise-resistant absolute temperature measurement in extreme environments, making it suitable for scenarios where on-site calibration is not possible.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123111473A_ABST
    Figure CN123111473A_ABST
Patent Text Reader

Abstract

This invention relates to the interdisciplinary field of optical radiation thermometry and intelligent computing, and is a multispectral true temperature inversion method that does not require blackbody calibration. The method includes: acquiring the original voltage values ​​output by a multispectral detection system to a target at various measurement times across multiple wavelength channels; constructing an equivalent spectral response function and establishing an observation model in the original voltage domain; representing the equivalent spectral response function as an undetermined continuous function in the wavelength domain, and representing the temperature at each measurement time as an undetermined temperature sequence, thus forming an inversion model that jointly solves for temperature and spectrum; constructing a data fitting loss based on Planck radiation consistency, and introducing at least one smoothing regularization constraint, and weighted combining them to obtain the total loss function; introducing a weak temperature prior for any measurement time, constructing a scale anchoring constraint and adding it to the total loss function; iteratively solving for the minimum value of the total loss function using a numerical optimization method, and outputting the true temperature sequence and equivalent spectral response function at each measurement time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of optical radiation thermometry and intelligent computing, and in particular to a multispectral true temperature inversion method that does not require blackbody calibration. Background Technology

[0002] Multispectral radiometric thermometry is a non-contact temperature measurement technique based on Planck's blackbody radiation law. It acquires radiation signals from a target at multiple wavelengths and then retrieves the target's true temperature and spectral emissivity. Compared to contact thermometry, it offers advantages such as faster response time, higher temperature limit, and no interference with the measured temperature field, making it a core technology for temperature measurement in extreme high-temperature environments.

[0003] Existing multispectral thermometry technology has long faced two core bottlenecks that are difficult to overcome: the first is the emissivity model dependency problem. Multispectral thermometry is essentially an underdetermined inverse problem; n spectral channels can only yield n equations, but contain n emissivity and 1 temperature, for a total of n+1 unknowns. Traditional methods close the equation system by pre-setting a function model of emissivity changing with wavelength (such as linear, polynomial, or logarithmic models). However, the emissivity of actual targets is affected by multiple factors such as material composition, surface oxidation, temperature, and roughness, and its spectral distribution is often difficult to accurately describe with a simple function. Model mismatch can lead to temperature measurement errors of tens to hundreds of degrees. To address this problem, the academic community has proposed improvement schemes such as sparse representation reconstruction methods, constrained optimization iterative methods, and swarm intelligence optimization algorithms, which have reduced model dependency to some extent, but have not fundamentally eliminated the assumption constraint of the emissivity function form.

[0004] Secondly, there is the issue of reliance on blackbody calibration. Traditional methods require full-band absolute radiometric calibration in a laboratory using a standard blackbody furnace to determine the instrument response coefficients for each channel before the voltage signal can be converted into absolute radiance, and thus the temperature can be calculated. However, in practical engineering scenarios such as rocket engine testing, explosion impact testing, and high-temperature molten metallurgy, it is impossible to place the standard blackbody furnace at the location of the target for in-situ calibration. Furthermore, atmospheric attenuation during long-distance measurements and parameter drift after instrument transportation and installation can all render laboratory calibration results invalid, severely limiting the engineering applications of multispectral thermometry.

[0005] In recent years, the rise of physics-guided neural networks has provided a new paradigm for solving inverse physics problems. By embedding physical laws into loss functions, these networks can solve underdetermined inverse problems from sparse and noisy data, and have been validated in fields such as computational fluid dynamics and heat transfer. However, no research has yet fundamentally solved the core theoretical problem of the ambiguity of absolute temperature scale under blackbody-free calibration conditions, nor has a calibration-free inversion scheme that can be implemented in engineering been developed.

[0006] In summary, there is a need to provide a multispectral true temperature inversion method that does not require a pre-set emissivity model or full-band blackbody furnace calibration, in order to overcome the dual dependence bottleneck of existing technologies and meet the temperature measurement requirements in extreme environments. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a multispectral true-temperature inversion method that does not require blackbody calibration. This invention couples the unknown instrument spectral response with the unknown target emissivity into a unified equivalent spectral quantity, directly constructing the inversion problem in the original voltage domain, thus bypassing absolute radiometric calibration. It closes the underdetermined equations through smoothing prior constraints in the spectral and temporal domains, and then resolves global-scale ambiguities with a single-point weak temperature prior, achieving absolute true-temperature inversion without blackbody calibration. This method overcomes the dual technical bottlenecks of traditional multispectral thermometry, which relies on emissivity model assumptions and full-band blackbody furnace calibration. It also possesses strong noise resistance and high inversion accuracy, making it suitable for extreme temperature measurement scenarios where on-site calibration is impossible.

[0008] The technical means employed in this invention are as follows: A multispectral true-temperature inversion method without blackbody calibration includes: acquiring the original voltage values ​​output by a multispectral detection system to a target at various measurement times across multiple wavelength channels; constructing an equivalent spectral response function, wherein the equivalent spectral response function includes the coupling effect between the unknown relative spectral response of the multispectral detection system and the unknown spectral emissivity of the target; establishing an observation model in the original voltage domain, such that the observed voltage values ​​of each channel are matched with the product of the equivalent spectral response function and the Planck radiation term; and expressing the equivalent spectral response function as an undetermined continuous function in the wavelength domain, and then... The temperature at each measurement moment is represented as an undetermined temperature sequence, forming an inversion model for joint solution of temperature and spectrum. A data fitting loss is constructed based on Planck radiation consistency, and at least one smoothing regularization constraint is introduced. The total loss function is obtained by weighted combination. A weak temperature prior for any measurement moment is introduced, and a scale anchoring constraint is constructed and added to the total loss function to resolve the global multiplicative scale ambiguity among voltage, equivalent spectral response, and temperature. The minimum value of the total loss function is iteratively solved through numerical optimization methods, and the true temperature sequence and equivalent spectral response function for each measurement moment are output.

[0009] Furthermore, the equivalent spectral response function The relative spectral response of the instrument With target spectral emissivity Coupled together:

[0010] The observation model is based on the complete Planck blackbody radiation law, and the Planck radiation term is used to... Represented as:

[0011] in, The first radiation constant, The second radiation constant, For wavelength, The target thermodynamic true temperature; Establish observation equations based on the complete Planck law. :

[0012] in, Indicates the time index. This indicates the wavelength channel index.

[0013] Furthermore, the undetermined continuous function in the wavelength domain is parameterized using a multilayer perceptron network with a sinusoidal activation function. The network input is the wavelength value, and the output is the equivalent spectral response value of the corresponding wavelength.

[0014] Furthermore, the total loss function Loss due to data fitting Spectral regularization term Temperature time series regularization term Single-point weak temperature prior Weighted composition:

[0015] in, These represent the weighting coefficients of the spectral regularization term. The weighting coefficients for the temperature time series regularization term are represented. This represents the weighting coefficient of a single-point weak temperature prior.

[0016] The data fitting loss is expressed as:

[0017] in, This represents the total number of measurement times. This indicates the total number of wavelength channels in a multispectral detection system.

[0018] The spectral regularization term Represented as:

[0019] in, This represents the equivalent spectral response function value corresponding to the (j+1)th wavelength channel.

[0020] The temperature time series regularization term Represented as:

[0021] in, This represents the true temperature value at the (i+1)th measurement time.

[0022] The single-point weak temperature prior term Represented as:

[0023] in, This represents the true temperature value at the first measurement moment.

[0024] Furthermore, the weak temperature prior adopts a constraint form that combines a tolerance interval with soft penalties. A reasonable range of temperature at the target time is preset. When the inversion temperature is within the range, the penalty value is zero. When the inversion temperature exceeds the range, a penalty term is applied.

[0025] Furthermore, the numerical optimization method is a gradient descent-type optimization algorithm, using any one of Adam, SGD, and L-BFGS; during the iteration process, a learning rate decay strategy is used to improve convergence accuracy.

[0026] Compared with the prior art, the present invention has the following advantages: This invention achieves inversion directly in the voltage domain through equivalent spectral coupling modeling, eliminating the need for absolute radiation calibration in a full-band blackbody furnace. It only requires a single-point weak temperature prior that is easily obtained in engineering to obtain the absolute true temperature, completely solving the industry pain point of being unable to perform on-site calibration in extreme scenarios. It can be directly applied to temperature measurement of inaccessible targets such as rocket plumes and explosion fields.

[0027] This invention uses a continuous function to model the equivalent spectral response, which can adaptively fit spectral distributions of arbitrary shapes without pre-assuming the functional form of emissivity. This fundamentally avoids the model mismatch error of traditional methods and significantly improves the adaptability to complex emissivity targets.

[0028] This invention combines spectral smoothing and temporal smoothing regularization terms, constrained by Planck's physical laws, to construct a highly self-consistent loss function. Spectral smoothing regularization suppresses high-frequency spectral noise, while temporal smoothing regularization filters out random temperature jumps; the two work synergistically to improve inversion stability under low signal-to-noise ratio (SNR). Simulation results show that, under a 20dB SNR condition, the inversion accuracy of this method is still an order of magnitude better than traditional methods, demonstrating strong robustness to noisy industrial environments. Unlike purely data-driven black-box methods, the inversion process of this invention is entirely constrained by Planck's radiation law, and the output spectral curves have clear physical meaning. The inversion results can be cross-validated using spectral features (such as characteristic emission peaks of combustion products), resulting in higher reliability. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of the multispectral true temperature inversion method without blackbody calibration in this invention.

[0031] Figure 2 This is a schematic diagram illustrating the principle of equivalent spectral coupling modeling in this invention.

[0032] Figure 3 This is a block diagram illustrating the composition of the total loss function in this invention.

[0033] Figure 4 This is a schematic diagram of the global multiplicative scale ambiguity principle in this invention.

[0034] Figure 5 This is a comparison chart of the temperature inversion error between the present invention and the traditional method under different signal-to-noise ratio conditions in the embodiments of the present invention.

[0035] Figure 6 This is a temperature inversion result and spectral characteristic verification diagram of the measured rocket plume data in this embodiment of the invention.

[0036] Figure 7 This is a module architecture diagram of the multispectral true temperature inversion system in an embodiment of the present invention. Detailed Implementation

[0037] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0040] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0041] like Figure 1 As shown, this invention provides a multispectral true temperature inversion method without blackbody calibration, comprising: acquiring the raw voltage values ​​output by a multispectral detection system to the target at each measurement time on multiple wavelength channels; and acquiring the radiation signal of the target using an 8-channel multispectral pyrometer, with the center wavelength of each channel being 0.57 nm. 0.62 0.68 0.75 0.82 0.89 0.95 1.02 It covers the visible to near-infrared wavelength range. A raw voltage data matrix was acquired at N time points. After performing dark current subtraction and outlier removal on the original data, it is input into the inversion algorithm.

[0042] The following symbols and units are used consistently in this instruction manual: wavelength Unit is (micrometer); Temperature T: The unit is K (Kelvin), which refers to the thermodynamic true temperature; Channel observation voltage V: The unit is V (volts), which is the output voltage of the detector after signal conditioning; First radiation constant: ; Second radiation constant: K; Equivalent spectral response function Composite physical quantity, dimensionless This method integrates the instrument's spectral gain and the target's emissivity, and solves them as a whole without forcibly separating them. Planck radiation term Blackbody spectral radiant exitance, measured in units of . The solid angle, transmittance, and other coefficients of the instrument's optical system have all been incorporated into the equivalent spectral response function.

[0043] Subscript rules: Indicates the time index. Indicates the wavelength channel index, such as Indicates the first Time, Number The voltage value of the channel.

[0044] In practice, the center wavelength of each channel is used for approximate calculations. This approximation has sufficient accuracy under conditions of narrowband filters and gradually varying spectra. For broadband channels or scenarios with drastic spectral changes, an integral form can be used to establish the observation equations. ,in Let be the spectral response function of the j-th channel.

[0045] An equivalent spectral response function is constructed, which incorporates the coupling effect between the unknown relative spectral response of the multispectral detection system and the unknown spectral emissivity of the target. An observation model is established in the original voltage domain, ensuring that the observed voltage values ​​of each channel match the product of the equivalent spectral response function and the Planck radiation term. In a preferred embodiment of this invention, the equivalent spectral response function... The relative spectral response of the instrument With target spectral emissivity Coupled together:

[0046] The observation model is based on the complete Planck blackbody radiation law, and the Planck radiation term is... Represented as:

[0047] in, The first radiation constant, The second radiation constant, For wavelength, The target thermodynamic true temperature; Establish observation equations based on the complete Planck law. :

[0048] in, Indicates the time index. This indicates the wavelength channel index.

[0049] During implementation, the temperature measurement range is above 1500K, and the wavelength is less than 1.1. ,satisfy Therefore, the Wien approximation is used to simplify the calculation, with an error of less than 0.5%. The simplified form is as follows:

[0050] The equivalent spectral response function is represented as an undetermined continuous function in the wavelength domain, and the temperature at each measurement moment is represented as an undetermined temperature sequence, thus forming an inversion model for joint solution of temperature and spectrum. In a specific implementation, as a preferred embodiment of the present invention, the undetermined continuous function in the wavelength domain is parameterized using a multilayer perceptron network with a sinusoidal activation function. The network input is the wavelength value, and the output is the equivalent spectral response value of the corresponding wavelength.

[0051] In implementation, a hidden equivalent spectral field is constructed using a SIREN network with two hidden layers. The network input is a single-wavelength scalar, each layer contains 32 neurons, the activation function is a sine function, and the output is the equivalent spectral response value for the corresponding wavelength. The network parameters are as follows: This continuous representation enables spectral interpolation at arbitrary wavelengths, overcoming the limitation of the number of discrete channels.

[0052] Depending on the complexity of the target spectrum, the number of hidden layers can be 3-4, and the number of neurons can be 64-128, in order to improve the ability to fit complex spectral curves.

[0053] Simultaneously, the temperatures at N time points are set as a one-dimensional learnable parameter vector. The initial values ​​are uniformly set to the median of the prior temperature range. The final parameters to be optimized include the spectral field network parameters. With temperature vector .

[0054] Explanation of numerical stability implementation: Since the Planck term involves exponential operations, the numerical range is large under high temperature and shortwave conditions, which can easily lead to gradient vanishing or overflow problems. This embodiment constructs a fitting loss in the logarithmic voltage domain, that is, calculating the mean square error after simultaneously taking the logarithm of both the observed voltage and the model-predicted voltage. This can significantly reduce the numerical condition number and improve optimization stability; alternatively, a temperature logarithmic parameterization method can also be used. Further optimize numerical characteristics.

[0055] A data fitting loss is constructed based on Planck radiation consistency, and at least one smoothing regularization constraint is introduced. A weighted combination yields the total loss function. Specifically, in a preferred embodiment of this invention, the total loss function... Loss due to data fitting Spectral regularization term Temperature time series regularization term Single-point weak temperature prior Weighted composition:

[0056] in, These represent the weighting coefficients of the spectral regularization term. The weighting coefficients for the temperature time series regularization term are represented. This represents the weighting coefficient of a single-point weak temperature prior.

[0057] The data fitting loss is expressed as:

[0058] in, This represents the total number of measurement times. This represents the total number of wavelength channels in the multispectral detection system. This term forces all measurement points to satisfy Planck's radiation law and is the core constraint for the physical validity of the inversion results. To ensure the dimensionless nature of the weighting coefficients, all loss terms are normalized during construction by dividing them by their initial iteration values ​​to maintain a consistent initial magnitude for each loss term.

[0059] The first derivative of the equivalent spectrum is calculated at the wavelength sampling points using the central difference method. The sum of squares of the derivatives is used as a regularization term to define the spectral regularization term. Represented as:

[0060] in, This represents the equivalent spectral response function value corresponding to the (j+1)th wavelength channel. This term utilizes the physical prior that the emissivity of a real object and the instrument response are both continuously slowly varying functions of wavelength to suppress high-frequency oscillations in the spectral curve caused by noise.

[0061] Temperature time series regularization term Represented as:

[0062] in, This represents the true temperature value at the (i+1)th measurement moment. This term utilizes the temporal continuity of the physical process to suppress abnormal temperature jumps at single moments, thereby improving the stability of dynamic temperature measurement.

[0063] like Figure 4As shown, voltage domain inversion suffers from global scale ambiguity: when the equivalent spectrum is scaled by 1 / k and the temperature by k, the observed voltage remains completely unchanged, leading to an inability to uniquely determine the absolute temperature. This embodiment selects the first time step to apply a weak temperature prior, setting a reasonable temperature range of [1900K, 2100K], and using a piecewise quadratic penalty loss function. The single-point weak temperature prior term... Represented as:

[0064] in, This represents the true temperature value at the first measurement time. This prior only penalizes solutions that exceed a reasonable range, and does not constrain the inversion results within the interval, thus resolving scale ambiguity and avoiding systematic biases introduced by strong priors.

[0065] A weak temperature prior for any measurement time is introduced, and a scale anchoring constraint is constructed and added to the total loss function to resolve the global multiplicative scale ambiguity among voltage, equivalent spectral response and temperature. In a preferred embodiment of the invention, the weak temperature prior adopts a constraint form of tolerance interval combined with soft penalty. A reasonable range of temperature at the target time is preset. When the inversion temperature is within the range, the penalty value is zero. When the inversion temperature exceeds the range, a penalty term is applied.

[0066] The total loss function is iteratively minimized using a numerical optimization method, outputting the true temperature sequence and equivalent spectral response function at each measurement time. Specifically, as a preferred embodiment of this invention, the numerical optimization method is a gradient descent-type optimization algorithm, employing any one of Adam, SGD, or L-BFGS; a learning rate decay strategy is used during the iteration process to improve convergence accuracy.

[0067] During implementation, the Adam optimizer was used for end-to-end iterative optimization, with an initial learning rate set to 2×10⁻⁶. - ³, after 1500 iterations, the learning rate decays to 0.1 of its original value, and the total number of iterations is 2500. After convergence, the true temperature sequence at each time step is output. and continuous equivalent spectral response curve .

[0068] Example 1 This embodiment further illustrates the specific implementation of the method of the present invention, as detailed below: S1. Equivalent spectral coupling modeling: The relative spectral response, unknown to the instrument, is coupled with the spectral emissivity, unknown to the target, and defined as the equivalent spectral response function. The observation equation based on Planck's law is established directly in the original voltage signal domain. It is important to note that this method does not require the equivalent spectral response function to be decomposed into two independent quantities: instrument response and emissivity. These are treated as a composite physical quantity in the overall solution. If the relative distribution of emissivity needs to be obtained separately later, the relative spectral curve can be obtained through normalization, without requiring additional calibration information. This modeling method, in principle, bypasses the absolute radiometric calibration process.

[0069] S2. Construction of the joint inversion model: The equivalent spectral response function is represented as an undetermined continuous function in the wavelength domain, and the temperature at each time step is set as a discrete unknown, thus forming an inversion model for joint solution of temperature and spectrum.

[0070] S3. Construction of self-consistent constraint loss function: A data fitting loss function based on Planck's radiation law is constructed to ensure the physical validity of the solution. Simultaneously, smoothing regularization constraints are introduced, including spectral and temporal smoothing regularization. Spectral smoothing regularization utilizes the physical property that both the true spectral emissivity and instrument response are slowly varying functions of wavelength, effectively suppressing high-frequency oscillations in the spectral curve introduced by measurement noise. Temporal smoothing regularization utilizes the temporal continuity of the target's true temperature field, filtering out random jump noise from single-time temperature inversion. The weighted combination of these two regularizations yields the basic total loss function, achieving self-consistent temperature and spectral constraints.

[0071] S4. Weak prior scale anchoring: A weak temperature prior for any measurement moment is introduced, and a scale anchoring constraint is incorporated into the total loss function to resolve the global multiplicative scale ambiguity. It should be noted that voltage domain blind inversion inherently suffers from the problem of global multiplicative scale ambiguity: if a set of solutions is found... Then apply a scaling transformation to it. (where k is any positive constant), the observation equation always holds true. This means that data fitting alone cannot uniquely determine the absolute temperature scale. This invention introduces a single-point weak temperature prior that is readily available in engineering, requiring only a rough temperature range to anchor the absolute temperature scale, thus restoring the identifiability of the problem with extremely low engineering cost. The methods for obtaining the weak temperature prior include, but are not limited to: coarse temperature estimation using two-color colorimetry, material melting point, phase transition point reference, historical measurement data under similar operating conditions, and thermal simulation calculation envelope, etc.

[0072] S5. Iterative optimization solution: The total loss function is iteratively solved using numerical optimization methods, and the spectral function and temperature nodes are optimized simultaneously. Finally, the true temperature sequence and equivalent spectral response curve at each time point are output.

[0073] Furthermore, the observation model can be established based on the complete Planck blackbody radiation law, which is applicable to general temperature measurement scenarios across the entire temperature and wavelength range. Under high temperature and shortwave conditions, the Wien approximation can also be used to simplify the calculation and reduce the computational load.

[0074] Furthermore, the parameterization methods for the equivalent spectral response function include, but are not limited to, continuous representation by neural networks, polynomial fitting, spline curve fitting, and other forms.

[0075] Furthermore, the weak temperature prior adopts a tolerance interval combined with soft penalty, which only applies penalties to solutions that exceed a reasonable range, and does not constrain the inversion results within the interval. This solves the scale ambiguity and avoids the systematic bias introduced by the strong prior.

[0076] It should be further explained that the role of smoothing regularization constraints is to mathematically close the underdetermined system of equations, transforming the underdetermined problem, which originally had infinitely many solutions, into a well-posed problem with a unique solution (in the sense of regularization); however, this unique solution still has global multiplicative scaling ambiguity—that is, the spectrum and temperature can be scaled simultaneously without changing the observed values. The role of weak temperature prior is precisely to anchor this absolute scale; the two have a clear division of labor and are indispensable.

[0077] To verify the performance of this method, simulation experiments were constructed for testing, and the results were compared with traditional polynomial fitting, ratio smoothing, and constrained optimization iterative methods. Among them, polynomial fitting is a typical method of traditional multispectral temperature measurement, which closes the system of equations and solves for temperature by assuming that the spectral emissivity changes with wavelength in a first- or second-order polynomial manner.

[0078] For easy comparison, Figure 5 The meanings of the algorithm labels are as follows: WC-PISF(ideal) and WC-PISF(non-ideal) represent the inversion results of this method under ideal spectral response and actual non-ideal spectral response, respectively; Poly represents polynomial fitting method; Ratio-Smooth represents ratio smoothing method; PCA-BI represents principal component analysis-double exponential method.

[0079] Simulation settings: A triangular wave dynamic process with a temperature range of 1900K~2100K, spanning 50 time points; the emissivity is a complex multi-peak curve containing two characteristic emission peaks, simulating the spectral characteristics of alumina combustion products in a rocket plume; Gaussian white noise of different intensities is superimposed on the voltage signal, with signal-to-noise ratios of no noise, 40dB, 30dB, and 20dB, respectively. The error evaluation metric is the mean absolute error (MAE), based on the actual temperature set in the simulation.

[0080] Inversion results as follows Figure 5As shown, under ideal conditions with no noise, the mean absolute errors of temperature inversion by this method under ideal and non-ideal conditions are 4.8K and 3.6K, respectively; under a medium signal-to-noise ratio of 40dB, the mean absolute errors of temperature inversion by this method under ideal and non-ideal conditions are 4.0K and 4.8K, respectively; under a signal-to-noise ratio of 30dB, the mean absolute errors of temperature inversion by this method under ideal and non-ideal conditions are 4.6K and 6.3K, respectively; under a low signal-to-noise ratio of 20dB, the mean absolute errors of this method under ideal and non-ideal conditions are 19.8K and 22.8K, respectively; the equivalent spectral curves obtained by inversion can accurately reproduce the bimodal characteristics of emissivity, with a peak position error of less than 10nm.

[0081] The results show that the proposed method is significantly superior to the traditional method under different noise levels, and its advantages are particularly prominent in low signal-to-noise ratio scenarios.

[0082] It is worth noting that Poly and Ratio-Smooth utilize the known instrument response R(λ) (with calibration methods) and are slightly better than WC-PISF at extremely low noise levels. This is a natural manifestation of the advantage of "perfect calibration". As a blind estimation method, WC-PISF only requires a single-point temperature prior to achieve performance close to that of the calibration method.

[0083] Example 2 The verification was performed using a multispectral measurement dataset of rocket engine nozzles from publicly available literature. This dataset contains voltage measurement data at 12 time points and 8 channels during steady-state testing, and the operating parameters are reproducible.

[0084] Set the prior temperature range as follows: The average temperature obtained from the inversion is The temperature is highly consistent with the engine's design rated temperature; the equivalent spectral curve obtained from the inversion clearly shows the alumina particles in... and The characteristic emission peak at that location is consistent with the known spectral characteristics of the combustion products, thus verifying the correctness of the inversion results from a spectral perspective. Figure 6 As shown.

[0085] Example 3 This embodiment provides a method for determining the regularization weight coefficients, ensuring that the parameter selection is based on scientific evidence and adapts to different measurement scenarios.

[0086] Since each loss term has been pre-normalized, and the weight coefficients are dimensionless relative weights, the optimization steps are as follows: First, fix the time-series regularization weights. Only the spectral smoothing regularization term is retained, and it is determined using the L-curve method. Optimal range: Plot an L-curve with the logarithm of the data fitting loss on the x-axis and the logarithm of the spectral regularization loss on the y-axis, and select the point corresponding to the maximum curvature of the curve. This is considered the optimal value. This point balances the accuracy of data fitting with the degree of spectral smoothness, avoiding overfitting or oversmoothing.

[0087] To obtain the optimal Then determine using the same L-curve method. Optimal range: Plot an L-curve with the logarithm of the data fitting loss on the x-axis and the logarithm of the time-series regularization loss on the y-axis, and select the point corresponding to the maximum curvature. As the final value.

[0088] The regularization parameters determined by the above method can maximize noise suppression while ensuring inversion accuracy, without relying on manual parameter tuning based on experience, thus improving the method's engineering applicability. In this embodiment, the preferred parameters are: .

[0089] Example 4 This embodiment provides another multispectral true temperature inversion scheme that does not require blackbody calibration. It adopts a discrete channel gain model, does not require neural network parameterization, and is more suitable for traditional embedded computing scenarios.

[0090] The specific steps are as follows: S1, Discrete Channel Modeling The observation voltage of each spectral channel is expressed as:

[0091] In the formula, The unknown proportional gain of the j-th channel includes factors such as instrument response and optical transmittance. Let be the equivalent emissivity factor of the j-th channel.

[0092] S2, Constraint Construction Spectral smoothing constraint: The equivalent emissivity factors of adjacent channels satisfy the smoothing variation constraint, i.e. ; Temporal smoothing constraint: The temperature at adjacent time points satisfies the constraint of continuous change; Scale anchoring constraint: Select the first time step and apply a weak temperature prior interval constraint to resolve the multiplicative ambiguity between gain and temperature.

[0093] S3, Optimization Solution The above problem is transformed into a constrained nonlinear optimization problem, which is solved by the Sequential Quadratic Programming (SQP) algorithm, and the output is the temperature at each time point and the equivalent emissivity and channel gain of each channel.

[0094] The solution in this embodiment has a small computational load and can be deployed on low-computing-power embedded devices. It is suitable for engineering scenarios with a small number of channels and high real-time requirements.

[0095] Example 5 like Figure 7 As shown, this embodiment provides a multispectral true temperature inversion system that does not require blackbody calibration, comprising a multispectral detection device and a data processing device.

[0096] In this embodiment, the multispectral detection device is an 8-channel filter array type multispectral pyrometer, consisting of an optical lens, a narrowband filter array, a silicon photodetector, a signal conditioning circuit, and a data acquisition card. It is used to collect the multispectral radiation of the target and convert it into a raw voltage signal. In other embodiments, a hyperspectral imager, a multichannel fiber optic spectrometer, or other similar equipment can be used instead, without affecting the execution of the method of this invention.

[0097] The data processing device is implemented using a general-purpose industrial host computer, including a processor and a memory. The memory stores a computer program that implements the inversion method of this invention. When the processor executes this program, it completes the following process: (1) Receive raw voltage data and channel wavelength parameters transmitted by the multispectral detection device; (2) Perform the complete algorithm steps of equivalent spectral modeling, inversion model construction, loss function calculation, and iterative optimization; (3) Output three types of results: true temperature sequence, equivalent spectral curve, and convergence curve. The device is also equipped with a display and storage unit for the visualization of inversion results and data archiving.

[0098] This system can complete absolute true temperature measurement without the need for a blackbody calibration device, making it easy to deploy on-site and conduct mobile testing, and suitable for various extreme temperature measurement scenarios.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multispectral true-temperature inversion method without blackbody calibration, characterized in that, include: Obtain the raw voltage values ​​of the target output by the multispectral detection system at each measurement time on multiple wavelength channels; An equivalent spectral response function is constructed, which includes the coupling effect between the unknown relative spectral response of the multispectral detection system and the unknown spectral emissivity of the target; an observation model is established in the original voltage domain to match the observed voltage values ​​of each channel with the product of the equivalent spectral response function and the Planck radiation term; The equivalent spectral response function is expressed as an undetermined continuous function in the wavelength domain, and the temperature at each measurement moment is expressed as an undetermined temperature sequence, thus forming an inversion model for joint solution of temperature and spectrum. The data fitting loss is constructed based on Planck radiation consistency, and at least one smoothing regularization constraint is introduced. The total loss function is obtained by weighted combination. A weak temperature prior for any measurement moment is introduced, and a scale anchoring constraint is constructed and added to the total loss function to resolve the global multiplicative scale ambiguity among voltage, equivalent spectral response and temperature. The minimum value of the total loss function is solved iteratively by numerical optimization method, and the true temperature sequence and equivalent spectral response function at each measurement time are output.

2. The multispectral true temperature inversion method without blackbody calibration according to claim 1, characterized in that, The equivalent spectral response function The relative spectral response of the instrument With target spectral emissivity Coupled together: The observation model is based on the complete Planck blackbody radiation law, and the Planck radiation term is used to... Represented as: in, The first radiation constant, The second radiation constant, For wavelength, The target thermodynamic true temperature; Establish observation equations based on the complete Planck law. : in, Indicates the time index. This indicates the wavelength channel index.

3. The multispectral true-temperature inversion method without blackbody calibration according to claim 1, characterized in that, The undetermined continuous function in the wavelength domain is parameterized using a multilayer perceptron network with a sinusoidal activation function. The network input is the wavelength value, and the output is the equivalent spectral response value for the corresponding wavelength.

4. The multispectral true temperature inversion method without blackbody calibration according to claim 2, characterized in that, The total loss function Loss due to data fitting Spectral regularization term Temperature time series regularization term Single-point weak temperature prior Weighted composition: in, These represent the weighting coefficients of the spectral regularization term. The weighting coefficients for the temperature time series regularization term are represented. This represents the weighting coefficient of a single-point weak temperature prior. The data fitting loss is expressed as: in, This represents the total number of measurement times. This indicates the total number of wavelength channels in a multispectral detection system; The spectral regularization term Represented as: in, This represents the equivalent spectral response function value corresponding to the (j+1)th wavelength channel; The temperature time series regularization term Represented as: in, This represents the true temperature value at the (i+1)th measurement time. The single-point weak temperature prior term Represented as: in, This represents the true temperature value at the first measurement moment.

5. The multispectral true-temperature inversion method without blackbody calibration according to claim 1, characterized in that, The weak temperature prior adopts a constraint form that combines a tolerance interval with soft penalties. A reasonable range of temperature at the target time is preset. When the inversion temperature is within the range, the penalty value is zero. When the inversion temperature exceeds the range, a penalty term is applied.

6. The multispectral true temperature inversion method without blackbody calibration according to claim 1, characterized in that, The numerical optimization method is a gradient descent-type optimization algorithm, using any one of Adam, SGD, or L-BFGS; a learning rate decay strategy is used during the iteration process to improve convergence accuracy.