A wall surface spectral emissivity solving method based on a GRU network with attention mechanism

Abstract

This invention discloses a method for solving the spectral emissivity of a wall surface based on a GRU network with an attention mechanism, comprising the following steps: dual-temperature calibration of an infrared radiometer using a reference blackbody; measuring the wall surface target using the calibrated infrared radiometer and calculating the target reference spectral radiance using the measurement results; constructing a GRU network model with an attention mechanism and training the model using the target reference spectral radiance data after removing atmospheric interference-related bands; using the trained network to predict and correct the mid-wave infrared and long-wave infrared regions, and solving the wall surface emissivity using the corrected spectral radiance. This method for solving the spectral emissivity of a wall surface, by constructing a recurrent neural network model, can correct the target reference spectral radiance, effectively eliminating the influence of atmospheric absorption on the emissivity solution and improving the accuracy of the wall surface spectral emissivity solution.

Description

Technical Field

[0001] This invention relates to a method for solving wall emissivity, specifically a method for solving wall spectral emissivity based on a GRU network with attention mechanism. Background Technology

[0002] Atmospheric absorption in the infrared thermal radiation wavelength region affects the accuracy of target spectral radiance measurements, which in turn affects the accuracy of wall emissivity calculations. Current techniques often employ machine learning to correct for measured spectral radiance, but the accuracy still needs improvement.

[0003] Therefore, in order to further improve the accuracy of infrared spectral emissivity calculation, a wall infrared emissivity calculation method based on a GRU network with attention mechanism has become an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the present invention provides a method for solving the wall spectral emissivity based on a GRU network with attention mechanism, so as to achieve effective correction of the target reference spectral radiance and thus accurately solve the wall emissivity.

[0005] The technical solution provided by this invention is as follows: A method for solving wall emissivity using a GRU network with an attention mechanism, comprising the following steps:

[0006] S1: Use a reference blackbody to perform dual-temperature calibration of the infrared radiometer;

[0007] S2: Use the calibrated infrared radiometer to measure the wall target, and use the measurement results to calculate the target reference spectral radiance;

[0008] S3: Construct a GRU network model with attention mechanism, and train the GRU network model with attention mechanism using the target reference spectral radiance data after removing atmospheric interference related bands to obtain a trained GRU network with attention mechanism. The input of the GRU network with attention mechanism is the spectral radiance wavelength, and the output is the calculated target reference spectral radiance value.

[0009] S4: Use the trained network to predict and correct the mid-infrared and long-infrared regions, and use the corrected spectral radiance to solve for the wall emissivity.

[0010] Preferably, the GRU network model with attention mechanism created in S3 includes four layers. The first layer is the sequence input layer. The second layer, after the sequence input layer, is a feedforward GRU network containing multiple GRU units. The third layer is a multi-head attention mechanism layer, which passes the output of the GRU to the multi-head attention mechanism layer for self-attention operation to capture long-range dependencies. The fourth layer is a fully connected layer containing only one positive linear unit to generate the final radiance output.

[0011] Further optimization involves using two bands in S3 for training and testing: the 3–5 μm band and the 8–14 μm band.

[0012] Further optimization reveals that in S3, the atmospheric interference-related bands are the 4.1–4.5 μm band and the 13–14.5 μm band.

[0013] In a further preferred embodiment, in S3, before the data is input into the GRU network model with attention mechanism, the input data is standardized by performing deviation standardization based on the maximum and minimum values ​​of the data, so that the processed data is concentrated between 0 and 1.

[0014] The wall spectral emissivity solution method provided by this invention, based on the GRU network with attention mechanism, can correct the target reference spectral radiance by constructing a recurrent neural network model according to the principle of infrared spectral radiance solution. The infrared emissivity of the wall surface can be solved based on the corrected data, which can effectively eliminate the influence of atmospheric absorption on the emissivity solution and improve the accuracy of wall spectral emissivity solution. Attached Figure Description

[0015] 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 of the present invention or the prior art will be briefly introduced below. Obviously, the drawings described below are only a part of the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for solving the wall spectral emissivity based on a GRU network with an attention mechanism, provided by this invention;

[0017] Figure 2 Here is a diagram of the GRU network structure;

[0018] Figure 3 A diagram of the multi-head attention structure;

[0019] Figure 4 This is a diagram of the brightness regression network structure;

[0020] Figure 5 This is a comparison chart of convergence speeds during network training.

[0021] Figure 6 The images show the test results of the brightness regression network on the original training set, (a) the processing results of the GRU network, and (b) the processing results of the GRU network with multi-head attention mechanism.

[0022] Figure 7 A schematic diagram of the brightness and spectral emissivity of a wall surface in the 3-5 μm band at a temperature controlled at 80℃, (a) is the radiance, and (b) is the spectral emissivity.

[0023] Figure 8 A schematic diagram of the brightness and spectral emissivity of a wall surface in the 8–14 μm band at a temperature controlled at 80℃, (a) is the radiance, and (b) is the spectral emissivity.

[0024] Figure 9 A schematic diagram of the brightness and spectral emissivity of a wall surface in the 3-5 μm band at a temperature controlled at 100℃, (a) is the radiance, and (b) is the spectral emissivity.

[0025] Figure 10 A schematic diagram of the brightness and spectral emissivity of a wall surface in the 8–14 μm band at a temperature controlled at 100℃, (a) is the radiance, and (b) is the spectral emissivity.

[0026] Figure 11 A schematic diagram of the brightness and spectral emissivity of a wall surface in the 3-5 μm band at a temperature controlled at 120℃, (a) is the radiance, and (b) is the spectral emissivity.

[0027] Figure 12 The diagram shows the brightness and spectral emissivity of a wall surface in the 8–14 μm band at a temperature controlled at 120℃. (a) represents the radiance, and (b) represents the spectral emissivity. Detailed Implementation

[0028] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems consistent with some aspects of the invention as detailed in the appended claims.

[0029] To improve the accuracy of spectral emissivity calculations, such as Figure 1 As shown, this implementation scheme provides a method for solving the wall spectral emissivity based on a GRU network with an attention mechanism, including the following steps:

[0030] S1: Use a reference blackbody to perform dual-temperature calibration of the infrared radiometer;

[0031] To accurately measure the radiation information of a target using a radiometer, it is necessary to calibrate the radiometer using a standard blackbody before the formal experiment. Since spectral radiance, relative to spectral radiant intensity, does not require consideration of the area of ​​the object being measured, calibration using spectral radiance is more convenient; simply filling the field of view with the blackbody ensures consistent measurement conditions. This invention employs a two-temperature calibration method for radiometer calibration, which is currently the most commonly used and effective method. The standard blackbody is heated to two different temperatures, filling the radiometer's lens field of view with the blackbody. By analyzing the radiance data, radiative responsivity, and offset at the two temperatures, a response relationship between the measured and theoretical values ​​of the infrared radiometer is established, thus performing two-temperature blackbody radiometric calibration of the infrared radiometer.

[0032]

[0033] Among them, S C (λ) and S H (λ) represents the response values ​​of the infrared radiometer at lower and higher temperatures, respectively; K(λ) is the instrument response curve of the infrared radiometer, which does not change with ambient or instrument temperature; N(λ) is the bias of thermal radiation generated by the instrument itself. To ensure that the thermal radiation bias remains constant, the instrument generally uses liquid nitrogen cooling or electric cooling to ensure the stability of this value; L C (λ) and L H (λ) represents the spectral radiance at lower and higher temperatures.

[0034] The spectral radiance of an ideal blackbody at temperature T can be calculated using Planck's formula:

[0035]

[0036] In the formula: L(λ) is the spectral radiance, with units of W·cm. -2 ·sr -1 ·μm -1 λ is the wavelength, in μm; T is calculated using absolute temperature, in K; c1 is the first radiation constant, c2 is the second radiation constant, where c1 = 3.742 × 10⁻⁶. - 12 W·cm -2 c2 = 1.439 × 10 4 K·μm.

[0037] Combining formulas (1) and (2), we can obtain the two unknown coefficients K(λ) and N(λ,T).

[0038] S2: Use the calibrated infrared radiometer to measure the wall target, and use the measurement results to calculate the target reference spectral radiance;

[0039] During target testing, the radiometer's field of view, gain, and scan rate are set to be consistent with those used during calibration. The test location is determined primarily by factors such as the instrument's field of view, the target size, and the personnel's safe distance. Assuming the estimated target length at a certain test angle is L (m) and the instrument's instantaneous field of view angle is θ (rad), the minimum test distance R is expressed as:

[0040] R=L / θ (3)

[0041] After the radiometer measures the target, the calibrated radiance is obtained from the calibrated K(λ) and N(λ,T), as shown in the following formula:

[0042]

[0043] S3: Construct a GRU network model with attention mechanism, and train the GRU network model with attention mechanism using the target reference spectral radiance data after removing atmospheric interference related bands to obtain a trained GRU network with attention mechanism. The input of the GRU network with attention mechanism is the spectral radiance wavelength, and the output is the calculated target reference spectral radiance value.

[0044] As a type of recurrent neural network, the GRU network replaces each traditional recurrent neural network node in its hidden layers with a gated recurrent unit. The structure of each unit node is as follows: Figure 2 As shown.

[0045] x t h represents the input information at the current moment. t-1 This represents the hidden state from the previous time step. The hidden state acts as the neural network's memory, containing information about the data seen by previous nodes. t This represents the hidden state that is passed on to the next time step. This represents the candidate hidden state. t To reset the door, z t To update the door.

[0046] The reset gate determines how new input information is combined with previous memories, and this relationship is expressed mathematically as follows:

[0047] r t =σ(W r ·[h t-1 ,x t (5)

[0048] Where σ is a logical S(Sigmoid) form function, W r It is the weight vector. t The solution will be used in the formula for the candidate hidden state:

[0049]

[0050] From equation (6), we can see that r t The larger r is, the more the previous time step is combined with the current time step. t When the value is 1, the hidden state from the previous time step is completely preserved; when r... t When the value is 0, all hidden states from the previous time step are retained but discarded. Therefore, the reset gate helps to capture short-term dependencies in a time series.

[0051] The update gate is used to update memory, and its mathematical formula is as follows:

[0052] z t =σ(W z ·[h t-1 ,x t (7)

[0053] The expression for updating memory is as follows:

[0054]

[0055] The final memory h t Forgot to pass down h t-1 Add some information from the current node's input, and then add some information from the current node's input, where z t The closer to 1, the more data is "remembered"; the closer to 0, the more data is "forgotten".

[0056] For infrared radiance sequences, noise bands generated by atmospheric absorption have a greater impact on emissivity measurements than other external conditions. When computational power is limited, attention mechanisms are a primary resource allocation scheme for addressing information overload, enabling the allocation of computational resources to more important tasks. Incorporating attention mechanisms into a model can increase model performance without changing its complexity.

[0057] Multi-head attention utilizes multiple queries to compute multiple pieces of information from the input in parallel, with each attention focusing on a different part of the input. Unlike using a single attention pooling, h different linear projections can be learned independently to transform the query, key, and value. These h transformed query, key, and value sets are then pooled in parallel. Finally, the outputs of these h attention poolings are concatenated and transformed through another learnable linear projection to produce the final output. Figure 3 This demonstrates the use of fully connected layers to implement multi-head attention that can learn linear transformations.

[0058] The inputs Q, K, and V are the query, key, and value, respectively. The attention mechanism first calculates the compatibility between the query and each key. Each compatibility is used as a weight for each value. The products of each weight and the value are added together to obtain the output.

[0059]

[0060] Multi-head attention involves applying attention to the same Q, K, and V multiple times to obtain multiple different outputs, and then connecting these different outputs to obtain the final output.

[0061] Based on the above, a GRU network model with a multi-head attention mechanism (luminance regression network) is designed as an infrared spectral radiance correction model. The luminance regression network is built upon the GRU network model, and a multi-head attention mechanism is added to the multiple hidden layer information output by the GRU network model. The output is then output through a fully connected layer. The input sequence data is the spectral radiance wavelength, and the output is the calculated target reference spectral radiance value. Therefore, the network has one unit in the input sequence layer and one unit in the output layer. The constructed luminance regression network structure is as follows: Figure 4 As shown.

[0062] The second layer after the sequence input layer is a feedforward GRU network containing multiple GRU units. The third layer is a multi-head attention mechanism layer, which passes the output of the GRU to the multi-head attention mechanism layer for self-attention operation to capture long-range dependencies. The fourth layer is a fully connected layer containing only one positive linear unit to produce the final radiance output.

[0063] The dataset used was a sequence of infrared radiance measurements obtained from an infrared spectroradiometer, containing measurement data from both mercury cadmium telluride (MCT) and indium antimonide (InSb) detectors in the 2–15 μm band. To obtain the best regression performance within two atmospheric windows (3–5 μm and 8–14 μm), a plan was adopted to train and test the network model separately for each band to determine the training and testing sets. First, the spectral radiance and its corresponding wavelengths in the 7.5–14.5 μm band collected by the MCT channel of the radiometer and the spectral radiance and its corresponding wavelengths in the 2.5–5.5 μm band collected by the InSb channel were extracted from the calibration experiment of a blackbody at a certain temperature T, forming two sets of sequences. Based on the characteristics of atmospheric transmittance, it is known that the influence of atmospheric absorption on the infrared spectrum is concentrated in the 4.1–4.5 μm and 13–14.5 μm bands. Therefore, the data at the corresponding wavelength positions were removed and used as the original input set for the regression network of each respective window band. The original test data were smoothed and mean-processed to obtain the original output set. Because radiance data has a relatively small numerical magnitude, directly using radiance data to train a network carries the risk of gradient vanishing. To address this issue, the input data is standardized before being fed into the network. This standardization is performed based on the maximum and minimum values ​​of the data, ensuring that the processed data is concentrated between 0 and 1, which helps in training an effective model.

[0064] S4: Use the trained network to predict and correct the mid-infrared and long-infrared regions, and use the corrected spectral radiance to solve for the wall emissivity.

[0065] After training, the radiation data of the wall surface measured by the infrared radiometer is fed into the trained network for output. The output result is the radiance data after removing atmospheric interference.

[0066] Since the temperature-controlled wall sample has reached thermal equilibrium with its surroundings, the wall radiance collected by the radiometer includes background radiance at the same temperature as the environment. In macroscopic observation, the radiant energy affected by the wall's emissivity is only the portion emitted from the wall that exceeds the ambient temperature.

[0067] L λm (λ)-L λbg (λ)=ε λ (λ)·(L λbb (λ)-L λbg (λ)) (10)

[0068] Among them, L λm It is the spectral radiance of the wall measured by the radiometer, L λbg It is the background spectral radiance at the experimental ambient temperature, ε λ It is the spectral emissivity of the wall, Lλbb It is a blackbody reference brightness calculated based on Planck's law.

[0069] The wall-emitted radiance L measured by the radiometer λm After processing with a luminance regression network, the wall-out emission luminance regression curve L is obtained. λmr Based on the measured regression luminance data L λmr (Corrected spectral radiance) and theoretically calculated blackbody reference radiance L λbb Calculating the spectral emissivity of the wall sample requires excluding the overlap brightness of the wall sample and the environment at the experimental temperature; that is, the wall spectral emissivity ε... λ Calculation formula

[0070]

[0071] and the average emissivity ε of the inner wall surface of band λ1-λ2 m Calculation formula

[0072]

[0073] Example

[0074] The network training process uses measurement data from a blackbody calibration radiation source with an emissivity of 0.98 at 80℃ in the 2.5–5.5 μm band as follows: Figure 5 As shown.

[0075] The mean-squared error (MSE) was used to evaluate the model's convergence speed. Adding the attention mechanism significantly improved the convergence speed. After changing the learning rate at the 100th iteration, the loss function value remained largely unchanged. On the current training set, the brightness regression network parameters converged, and training was complete. The newly constructed network converged approximately 1.5 times faster than the original network.

[0076] The performance of the brightness regression network in the 3–5 μm band on the original training set is as follows: Figure 6 As shown, the regressed brightness curve fits the theoretical brightness curve well, and it can also achieve smooth output of outlier input values ​​in the 4.1–4.4 μm band. Comparing the two algorithms, the GRU network's prediction result has an MSE of 1.45e-10 and a coefficient of determination (R²) of 0.9985; after adding the multi-head attention mechanism, the prediction result has an MSE of 8.07e-11 and an R² of 0.9991. This demonstrates that adding multi-head attention allows the GRU network to converge faster and improves the accuracy of the final prediction result.

[0077] The average spectral emissivity of the calibrated blackbody, obtained from the regression and theoretical brightness values, is 0.9705 (3–5 μm) and 0.9748 (8–14 μm), respectively, which is consistent with the actual emissivity of the calibrated blackbody. Table 1 shows the emissivity results for other temperature models. Therefore, the network model trained in this study demonstrates accuracy in regressing the infrared spectral radiance sequence of the target.

[0078] Table 1 Results of Blackbody Emissivity Calculation

[0079]

[0080] Based on the experimentally obtained wall spectral radiance, the obtained regression radiance data L λmr Blackbody reference brightness L λbb and ambient background spectral radiance L λbg Substituting into equation (12), the spectral radiance and emissivity of the wall sample at the three temperatures are calculated as follows: Figures 7-12 As shown.

[0081] The obtained regression brightness data L λmr Blackbody reference brightness L λbb and ambient background spectral radiance L λbg Substituting into equation (12), the average emissivity of the wall surface can be obtained as shown in Table 2.

[0082] Table 2 Emissivity calculation results of wall samples

[0083]

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended 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 therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for solving the wall spectral emissivity based on a GRU network with attention mechanism, characterized in that, Includes the following steps: S1: Use a reference blackbody to perform dual-temperature calibration of the infrared radiometer; S2: Use the calibrated infrared radiometer to measure the wall target, and use the measurement results to calculate the target reference spectral radiance; S3: Construct a GRU network model with an attention mechanism, and train the GRU network model using target reference spectral radiance data after removing atmospheric interference-related bands to obtain a trained GRU network. The input of the GRU network is the spectral radiance wavelength, and the output is the calculated target reference spectral radiance value; the atmospheric interference-related band is 4.1~4.

5. Bands and 13~14.5 Band; Before inputting data into the GRU network model with attention mechanism, the input data is standardized by performing deviation standardization based on the maximum and minimum values ​​of the data, so that the processed data is concentrated between 0 and 1. S4: Use the trained network to predict and correct the mid-infrared and long-infrared regions, and use the corrected spectral radiance to solve for the wall emissivity, where the wall spectral emissivity... Calculation formula: ； In the formula, It is the wall-emitted radiance measured by a radiometer. The wall emitted light after processing by a luminance regression network. It is the background spectral radiance at the experimental ambient temperature. It is a blackbody reference brightness calculated based on Planck's law.

2. The method for solving the wall spectral emissivity based on a GRU network with attention mechanism according to claim 1, characterized in that: The GRU network model with attention mechanism created in S3 consists of four layers. The first layer is the sequence input layer. The second layer, after the sequence input layer, is a feedforward GRU network containing multiple GRU units. The third layer is a multi-head attention mechanism layer, which passes the output of the GRU to the multi-head attention mechanism layer for self-attention operation to capture long-range dependencies. The fourth layer is a fully connected layer containing only one positive linear unit to generate the final radiance output.

3. The method for solving the wall spectral emissivity based on a GRU network with attention mechanism according to claim 1, characterized in that: In S3, the GRU network with attention mechanism is trained and tested separately in two bands: the 3~5μm band and the 8~14μm band.

Images