Optical field multispectral temperature inversion system and method based on physical information neural network

Abstract

The invention provides a light field multispectral temperature inversion system and method based on a physical information neural network, and the method comprises the steps: firstly constructing a deep learning network infrastructure, and employing a double-layer U-Net architecture to extract the radiation and spatial distribution characteristics of data; then designing a physical information embedding module, modularizing the Planck's law of radiation, and guiding the network to establish physical association between radiation information and temperature; then, light field multispectral radiation data are collected through experiments, the network is trained after data division and preprocessing are completed, and the mapping relation between radiation characteristics and temperature characteristics is established; and finally, migrating the trained network to actual test data to realize high-precision and rapid temperature inversion. The method has the advantage of processing complex data through deep learning and the constraint of a physical model, the precision and the calculation efficiency of light field multispectral temperature inversion are improved, and the applicability of the method in a high-temperature complex environment is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of temperature measurement technology, specifically relating to a light field multispectral temperature inversion system and method based on physical information neural networks. Background Technology

[0002] Multispectral radiation thermometry, with its unique ability to achieve non-contact temperature measurement in high-temperature and harsh environments, has demonstrated significant advantages in numerous fields such as combustion diagnostics and gas turbine blade monitoring. This technology, by capturing spectral radiation information closely related to temperature and combining it with imaging capabilities possessing high spatial and temporal resolution, has opened up a practical and feasible approach for real-time temperature measurement under complex operating conditions.

[0003] Optical field multispectral radiometric thermometry, as an emerging temperature measurement method, has injected new momentum into the high-precision reconstruction of two-dimensional temperature fields, powerfully promoting the further development of this field. The core innovation of this technology lies in the ingenious introduction of a microlens array into the imaging system. In this way, the system can simultaneously acquire spatial and spectral information, successfully eliminating the need for dispersive elements such as gratings and filters used in traditional methods. Therefore, the structure of the optical system is greatly simplified, and the system's robustness and compactness are significantly enhanced, greatly improving its feasibility in practical engineering applications. However, optical field multispectral radiometric thermometry is not without its limitations; its measurement accuracy and reliability are largely constrained by the performance of the inversion algorithm.

[0004] Traditional radiation thermometry methods typically rely on assumed emissivity models, employing simplified or parameterized approaches to solve the radiative transfer equation. Under specific conditions, these methods can achieve certain successes. However, their accuracy suffers significantly when dealing with objects with complex surface properties or in environments with variable characteristics. Another approach takes a different route, attempting to eliminate the dependence on emissivity by using spectral ratio methods or reference temperature models, thereby directly achieving temperature inversion. This method reduces errors caused by emissivity modeling to some extent. However, it still struggles to withstand noise interference and is prone to getting trapped in local optima, thus affecting the reliability of the measurement results.

[0005] In recent years, machine learning technology has been gradually introduced into the field of multispectral temperature measurement, using neural networks to improve the accuracy and adaptability of predictions.

[0006] For example, patent document CN113776675A discloses a multispectral radiometric thermometry inversion calculation method, computer, and storage medium based on a generalized inverse neural network. This method first simulates a multispectral radiometric thermometer to calculate the voltage value for each spectral channel; second, it constructs an equation for each spectral channel, forming a set of equations, and calculates temperature and emissivity data with similar patterns to those in the emissivity model using a generalized inverse matrix; third, it compares the data with the emissivity variation patterns in the emissivity model and classifies them; fourth, it defines a set of hyperparameters, trains the neural network, and modifies the parameters using a simulated annealing algorithm during training to train the optimal parameters; finally, it inputs a test set and a validation sample set into the neural network and outputs multispectral radiometric thermometry data. This solves the problem in existing technologies where the original data processing methods are not universally applicable to materials with different emissivities and cannot be quickly inverted. However, purely data-driven models often lack physical constraints, resulting in poor interpretability of their output results. Moreover, when training data is limited, the model's robustness is poor. To address this problem, the emergence of Physical Information Neural Networks (PINNs) provides a novel solution. These methods directly integrate prior physical knowledge into the network structure or loss function, allowing the model to learn in accordance with physical laws, thereby enhancing the model's noise resistance and generalization performance.

[0007] For example, foreign patent document US20240428063A1 discloses a neuromorphic optical computing architecture system, including: a multi-channel representation module for encoding the original input target light field signal into coherent light of different wavelengths using a multispectral laser; an attention-aware optical neural network module, including a bottom-up (BU) optical attention module and a top-down (TD) optical attention module, wherein coherent light of different wavelengths is input into the BU optical attention module and the attention-aware optical neural network is trained, and the TD optical attention module modulates the spectral and spatial transmittance of the multidimensional sparse features extracted by the BU optical attention module based on the trained attention-aware optical neural network to obtain the final spatial light output; and an output module for detecting and recognizing the final spatial light output on the output plane to obtain the position of the object in the light field and the recognition result.

[0008] However, the PINNs method has not yet been extensively studied in the field of multispectral radiometric thermometry, and PINNs are mostly used for numerical calculations, which presents challenges in applying it to calculations of multidimensional composite images such as optical field multispectral radiometrics. How to combine the spatial resolution advantages of optical field multispectral radiometric thermometry with the numerical accuracy of PINNs is a major problem that needs to be solved in current optical field multispectral radiometric thermometry. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for optical field multispectral temperature inversion based on a physical information neural network, comprising: Step S1: Establish a network architecture using a two-layer U-Net network; Step S2: Embed custom physical information; Step S3: Assign network training strategies; Step S4: Conduct network training based on the data, and balance the loss coefficient and adjust the training parameters through multiple rounds of optimization iterations; Step S5: Predict the current temperature in real time.

[0010] Preferably, the dual-layer U-Net network in step S1 includes: U-Net1 and U-Net2, where U-Net1 is the first layer U-Net and U-Net2 is the second layer U-Net; each U-Net structure consists of an encoding layer and a decoding layer, with an input tensor dimension of [1, 16n, H, W], where n is the number of spectra to be calculated, and H and W are the image length and width, respectively; furthermore, U-Net1 can extract image radiation features, specifically referring to extracting the spatial and spectral features of the ratio image; U-Net2 can process image temperature features, specifically referring to convolving and fusing the physically processed features to output the final temperature prediction result.

[0011] Preferably, the data flow of the network architecture in step S1 includes: firstly, decomposing the spectral channels of the multispectral images of the light field obtained by the target and the blackbody reference at a known temperature to obtain multiple sub-images; then dividing the target sub-image pixel by pixel with the corresponding channel's blackbody sub-image to obtain a ratio image based on the reference temperature, thereby achieving normalization of the radiation information; after the above processing, the dataset is then randomly divided and preprocessed to form the input for network training.

[0012] Preferably, in the optical field multispectral temperature inversion method based on physical information neural network, steps S1 and S2 include: U-Net1 takes the calculated number of spectra as input, uses convolution kernels to extract features, and generates channel radiation feature maps. Subsequently, the channel radiation feature maps are input to a processing stage that embeds custom physical information. In this processing stage, information fusion and transformation are performed in combination with specific physical rules and constraints. After this processing, channel temperature features are generated through cascaded operations. Finally, these channel temperature features are sent to U-Net2 for further feature mining and processing. The multi-channel temperature features are fused through convolution operations to output a temperature prediction map.

[0013] Preferably, the cascaded operation specifically refers to the step-by-step optimization of the predicted temperature and radiation characteristics, incorporating knowledge from the physical domain. The physical information embedding step consists of multiple sequentially connected sub-steps. Each sub-step receives the ratio feature map and temperature feature map output from the previous stage and processes them using a multi-branch convolutional structure. Furthermore, within each sub-step, physical information is explicitly embedded to establish the physical relationship between radiance, emissivity, and temperature, and the radiation and thermal feature representations are updated through multi-level iterations.

[0014] Preferably, step S3 includes: co-training with a custom multivariate loss function based on the principle of optical field multispectral radiation thermometry; further, the custom multivariate loss function includes: true temperature loss, prior temperature loss, channel consistency loss, and emissivity loss.

[0015] Preferably, steps S3 and S4 specifically refer to constructing a composite loss function for the network architecture to balance data fidelity and physical constraints during the training process. As shown in the following formula:

[0016] in, This is the total composite loss function; The loss between the network's predicted temperature and the actual temperature is the main loss term used to measure model performance. The loss is the difference between the network-predicted temperature and the prior numerical temperature. The temperature difference loss calculated for different spectral bands is used to ensure the spectral consistency of temperature prediction; The emissivity loss for network autonomous learning is to normalize the learned emissivity value to a physically valid boundary range and add a penalty term to suppress unreasonable predictions that exceed the range. to Let be the weight coefficients corresponding to different loss terms, and satisfy . The values ​​of the aforementioned weight coefficients depend on the training requirements of the model.

[0017] Preferably, step S5 specifically refers to directly transferring the trained network model to test or validation data outside the training set. The network model can achieve temperature prediction output based on the built-in data processing flow, without the need for model fine-tuning and secondary training.

[0018] This invention proposes a light field multispectral temperature inversion system based on a physical information neural network, employing a light field multispectral temperature inversion method based on a physical information neural network, including: Module M1: Establishes a network architecture using a two-layer U-Net network; Module M2: Embeds custom physical information; Module M3: Assigns network training strategies; Module M4: Conducts network training based on data, and balances the loss coefficient and adjusts the training parameters through multiple rounds of optimization iteration; Module M5: Real-time temperature prediction.

[0019] Preferably, the dual-layer U-Net network in module M1 includes: U-Net1 and U-Net2, where U-Net1 is the first layer U-Net and U-Net2 is the second layer U-Net; each U-Net structure consists of an encoding layer and a decoding layer, with an input tensor dimension of [1, 16n, H, W], where n is the number of spectra to be calculated, and H and W are the image length and width, respectively; furthermore, U-Net1 can extract image radiation features, specifically referring to extracting the spatial and spectral features of the ratio image; U-Net2 can process image temperature features, specifically referring to convolving and fusing the physically processed features to output the final temperature prediction result; The modules M1 and M2 include: U-Net1 takes the calculated spectral number as input, uses convolution kernels to extract features, and generates channel radiation feature maps. Subsequently, the channel radiation feature maps are input to the processing stage that embeds custom physical information. In this stage, information fusion and transformation are performed in combination with specific physical rules and constraints. After this processing, channel temperature features are generated through cascade operations. Finally, these channel temperature features are sent to U-Net2 for further feature mining and processing. The multi-channel temperature features are fused through convolution operations, and the output is a single temperature prediction map. The cascaded operation specifically refers to the step-by-step optimization of the predicted temperature and radiation characteristics, incorporating knowledge from the physical domain. The physical information embedding module consists of multiple sequentially connected sub-modules. Each sub-module receives the ratio feature map and temperature feature map output from the previous stage and processes them through a multi-branch convolutional structure. Furthermore, within each sub-module, physical information is explicitly embedded to establish the physical relationship between radiance, emissivity, and temperature, and the radiation and thermal feature representations are updated through multi-level iterations.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces physical constraints such as Planck's law and reference temperature into the deep learning framework, without relying on an explicit emissivity model, to achieve high-precision, spatially resolved inversion of the temperature field.

[0021] 2. This invention constructs a network structure combining a two-layer U-Net with a physical information embedding module. It adopts a series sandwich structure, which takes into account the advantages of deep learning feature extraction and the constraints of physical models. It can simultaneously extract the radiation and spatial features of multispectral light field data and achieve pixel-level temperature inversion.

[0022] 3. This invention uses a physical information embedding step to achieve physical constraint temperature prediction across spectral channels, which significantly improves the efficiency of inversion calculation while ensuring high accuracy and high signal-to-noise ratio, and is suitable for temperature measurement in high-temperature and complex environments.

[0023] 4. This invention can not only adapt to complex film cooling conditions, but also significantly improve the accuracy and reliability of temperature inversion. The absolute error of blade temperature inversion under different conditions is less than 10K, which can accurately reconstruct the temperature distribution of the region of interest of the blade and reflect the temperature difference at the thermocouple position.

[0024] 5. The method proposed in this invention has high accuracy in blackbody temperature prediction and performs even better under high temperature conditions. It has advantages in stability and accuracy compared with the Chameleon Swarm Algorithm (CSA) and the Minimum Parameter Optimization Algorithm (Fmincon).

[0025] 6. This invention improves the physical consistency and prediction accuracy of the inversion results, enhances its applicability in complex environments with high temperature and strong radiation, and demonstrates significant engineering application value and innovation. Attached Figure Description

[0026] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the process and network model structure of the optical field multispectral temperature inversion method based on physical information neural network proposed in this invention; Figure 2 This is a schematic diagram of the physical information embedding module network structure and data transmission process of the optical field multispectral temperature inversion method based on physical information neural network proposed in this invention. Figure 3 This is a schematic diagram showing the blackbody furnace temperature calibration experiment and the comparison results of the temperature calculation accuracy with different algorithms using the optical field multispectral temperature inversion method based on physical information neural network proposed in this invention. Figure 4 This is a schematic diagram showing the experimental temperature calculation results of turbine blade heating using the optical field multispectral temperature inversion method based on physical information neural network proposed in this invention, and the comparison results with the accuracy of thermocouples. Figure 5 This diagram shows the experimental temperature calculation results of turbine blade cooling effect using the optical field multispectral temperature inversion method based on physical information neural network proposed in this invention, and the comparison results with the temperature measurement accuracy of different algorithms and thermocouples. in, LFMP stands for Optical Field Multispectral Physical Information Neural Network; n represents the number of spectra to be calculated; Conv2d 3×3 represents a 3×3 two-dimensional convolutional layer; BatchNorm2d represents a two-dimensional batch normalization layer; ReLU represents the activation function; MaxPool 2d represents a two-dimensional max pooling layer; Conv2d 1×1&Dropout0.2 means a 1×1 two-dimensional convolutional layer combined with a Dropout layer with a dropout rate of 0.2; ConvTranspose2d represents a two-dimensional transposed convolutional layer; PAW stands for Physical Information Embedding Module, which is used to establish the relationship between radiation intensity R and temperature output T in a neural network. This represents the radiation intensity predicted by the model. This represents the radiation temperature predicted by the model. ( i-1 , i-1 ) represents the previous module in the i-th cascaded correction modules; ( i , i ) represents i cascaded correction modules; R represents the radiation parameter; ε represents the emissivity parameter. Detailed Implementation

[0027] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0028] This invention proposes a method for optical field multispectral temperature inversion based on a physical information neural network, comprising: Step S1: Establish a network architecture using a two-layer U-Net network; Step S1 includes: establishing a basic architecture using a two-layer U-Net network; further, the two-layer U-Net network includes U-Net1 and U-Net2, where U-Net1 is the first layer U-Net and U-Net2 is the second layer U-Net; U-Net1 can extract image radiation features, specifically referring to extracting the spatial and spectral features of the ratio image; U-Net2 can process image temperature features, specifically referring to convolving and fusing the physically processed features to output the final temperature prediction result; each U-Net structure consists of an encoding layer and a decoding layer, with an input tensor dimension of [1, 16n, H, W], where n is the number of spectra to be calculated, and H and W are the image length and width, respectively; The basic data flow of the network architecture includes: First, the multispectral images of the light field of the target and the blackbody reference at a known temperature are decomposed into multiple sub-images. For example, in a typical case, the four spectral channels correspond to 1535 nm, 1550 nm, 1654 nm and 1710 nm, respectively. Then, the target sub-image is divided pixel by pixel with the blackbody sub-image of the corresponding channel to obtain a ratio image based on the reference temperature, thereby realizing the normalization of radiation information. After the above processing, the dataset is randomly divided and preprocessed to form the input for network training.

[0029] Step S2: Embed custom physical information; Specifically, step S2 refers to modularly embedding Planck's law into the neural network, using a cascaded correction framework to establish a physical mapping between the radiation features and temperature features of the original light field multispectral image; Preferably, steps S1 and S2 include: U-Net1 generates an n-channel radiation feature map based on the calculated number of spectra n using a 3x3 convolution kernel, for example, with an output dimension of [1,n,H,W]. Then, after embedding custom physical information, it generates an n-channel temperature feature map through cascading. Finally, it is processed by U-Net2 and fused into a single temperature prediction map through convolution, for example, with an output dimension of [1,1,H,W]. Preferably, embedding custom physical information in step S2 specifically refers to refining the radiation features and predicted temperature features generated by U-Net1 by integrating knowledge from specific physical domains and using a cascaded correction framework. Furthermore, the cascaded correction framework includes: each sub-module receives the intermediate radiation and temperature feature maps of the previous stage after being connected in series, and iterates through multiple convolutional branches; each sub-module embeds Planck's law, and realizes physical sensing temperature prediction across spectral channels by modeling the relationship between radiance, emissivity and temperature. Preferably, the present invention adopts a sandwich structure in which a two-layer U-Net and a physical information embedding module are connected in series; the physical information embedding module is based on a customized formula of Planck's law, which explicitly introduces the radiation physics process into the network, and generates a multi-channel temperature feature map by combining reference temperature information. In this embodiment, as shown in the appendix Figure 2 As shown, the Physical Information Embedding Module (PLM) used in this invention employs a cascaded correction framework to progressively optimize the predicted temperature and radiation characteristics, incorporating knowledge from the physical domain, such as Planck's law. The PLM consists of multiple sequentially connected sub-modules. Each sub-module receives the ratio feature map and temperature feature map output from the previous stage and processes them through a multi-branch convolutional structure. Furthermore, Planck's law is explicitly embedded within each sub-module to establish the physical relationship between radiance, emissivity, and temperature. Through multi-level iterative updates of the radiation and thermal feature representations, the PLM achieves physically constrained temperature prediction across spectral channels, thereby further improving the accuracy and physical consistency of temperature inversion. Preferably, in this embodiment, a single PAW module first processes the radiation information input from the previous cascaded correction module through the radiation layer, and then integrates the emissivity and radiation intensity through the feature sharing layer. Finally, it outputs the corresponding temperature T̃ according to Planck's law. The module output at this time is then used as the input of the next cascaded module for calculation.

[0030] Step S3: Assign network training strategies; Step S3 includes: based on the principle of optical field multispectral radiation thermometry, a custom multivariate loss function is used for collaborative training; further, the custom multivariate loss function includes: true temperature loss, prior temperature loss, channel consistency loss, and emissivity loss. Step S4: Conduct network training based on the data, and adjust the training parameters by balancing the loss coefficient through multiple rounds of optimization iterations; Furthermore, steps S3 and S4 specifically refer to constructing a composite loss function for the network architecture to balance data fidelity and physical constraints during the training process, as shown in the following formula:

[0031] in, The total composite loss function consists of four sub-loss function terms; The loss between the network's predicted temperature and the actual temperature is the main loss term used to measure model performance. The loss is the difference between the network-predicted temperature and the prior numerical temperature. The temperature difference loss calculated for different spectral bands is used to ensure the spectral consistency of temperature prediction; To address the emissivity loss of network autonomous learning, the learned emissivity values ​​are normalized to the physically valid boundary range, and a penalty term is added to suppress unreasonable predictions that exceed the range. These are the weight coefficients corresponding to different loss terms, and The values ​​of the weight coefficients depend on the training requirements of the model.

[0032] Step S5: Predict the current temperature in real time; Specifically, step S5 refers to transferring the trained deep learning model to test data to achieve temperature prediction. Further, step S5 specifically refers to directly transferring the trained network model to test or validation data outside the training set. The network model can then achieve temperature prediction output based on its built-in data processing flow, without requiring model fine-tuning or secondary training. This invention proposes a light field multispectral temperature inversion system based on a physical information neural network, employing a light field multispectral temperature inversion method based on a physical information neural network, including: Module M1: Establishes a network architecture using a two-layer U-Net network; Module M2: Embeds custom physical information; Module M3: Assigns network training strategies; Module M4: Conducts network training based on data, and balances the loss coefficient and adjusts the training parameters through multiple rounds of optimization iteration; Module M5: Real-time temperature prediction.

[0033] Preferably, the dual-layer U-Net network in module M1 includes: U-Net1 and U-Net2, where U-Net1 is the first layer U-Net and U-Net2 is the second layer U-Net; each U-Net structure consists of an encoding layer and a decoding layer, with an input tensor dimension of [1, 16n, H, W], where n is the number of spectra to be calculated, and H and W are the image length and width, respectively; furthermore, U-Net1 can extract image radiation features, specifically referring to extracting the spatial and spectral features of the ratio image; U-Net2 can process image temperature features, specifically referring to convolving and fusing the physically processed features to output the final temperature prediction result; The modules M1 and M2 include: U-Net1 takes the calculated spectral number as input, uses convolution kernels to extract features, and generates channel radiation feature maps. Subsequently, the channel radiation feature maps are input to the processing stage that embeds custom physical information. In this stage, information fusion and transformation are performed in combination with specific physical rules and constraints. After this processing, channel temperature features are generated through cascade operations. Finally, these channel temperature features are sent to U-Net2 for further feature mining and processing. The multi-channel temperature features are fused through convolution operations, and the output is a single temperature prediction map. The cascaded operation specifically refers to the step-by-step optimization of the predicted temperature and radiation characteristics, incorporating knowledge from the physical domain. The physical information embedding module consists of multiple sequentially connected sub-modules. Each sub-module receives the ratio feature map and temperature feature map output from the previous stage and processes them through a multi-branch convolutional structure. Furthermore, within each sub-module, physical information is explicitly embedded to establish the physical relationship between radiance, emissivity, and temperature, and the radiation and thermal feature representations are updated through multi-level iterations.

[0034] To better understand the present invention, the following detailed description is provided in conjunction with examples and accompanying drawings: Example 1: This invention proposes a multispectral temperature inversion method for optical fields based on a physical information neural network. The method was used in a blackbody furnace calibration experiment. Specifically, blackbody images in the range of 673K ​​to 823K were acquired in the blackbody furnace calibration experiment, with a temperature interval of 25K. 100 frames of images were continuously acquired at each temperature point, resulting in a total of 1100 samples. After the acquired data was normalized by the calibration curve, it was randomly divided into a training set (approximately 70%), a test set (approximately 25%), and a validation set (approximately 5%). A fixed random seed was used in the division process to ensure repeatability. During model training, the blackbody emissivity was set to 0.999, the batch size to 32, the learning rate to 0.001, and the total number of training epochs to 150. The AdamW optimizer was used in conjunction with a checkpointing strategy to improve convergence and stability. The training platform consisted of four GTX 1080 Ti graphics cards and CUDA version 12.6. Convergence occurred after approximately 80 iterations, with a total training time of approximately 6 hours and a single image inference time of approximately 2 seconds.

[0035] Multiple independent training results show that the mean loss function on the validation set is approximately 0.01, indicating stable and repeatable performance. The model's absolute prediction error is less than 5K and the relative error is less than 0.5% across the entire temperature range, demonstrating high prediction accuracy.

[0036] For example, taking some results as examples, when the actual temperature is 648K, 698K, 748K, and 798K, the corresponding prediction errors are 3.0K, 4.3K, -3.0K, and 2.7K, respectively, with relative errors of 0.46%, 0.62%, 0.40%, and 0.34%. The results show that the method proposed in this invention has high accuracy in blackbody temperature prediction, and performs even better under high-temperature conditions.

[0037] Example 2: This invention proposes a method for optical field multispectral temperature inversion based on physical information neural networks, according to the attached... Figure 3 As shown, the method proposed in this invention has advantages in stability and accuracy compared to the Chameleon Swarm Optimization (CSA) algorithm and the Fmincon algorithm. Specifically, the comparison results across the entire temperature range of 673K ​​to 823K show that the method proposed in this invention significantly outperforms the two commonly used methods in terms of both absolute and relative error. In particular, the absolute error of the method proposed in this invention is consistently less than 5K, and the relative error is stable below 0.9%; while CSA and Fmincon exhibit larger error fluctuations, with Fmincon's deviation exceeding 30K at some temperature points and its relative error exceeding 4%. In comparison, although CSA is superior to Fmincon, it is still significantly inferior to the optical field multispectral temperature inversion method based on physical information neural networks proposed in this invention in the high-temperature region.

[0038] Example 3: This invention proposes a multispectral temperature inversion method for optical fields based on physical information neural networks. To verify the applicability of the proposed method under actual complex working conditions, a turbine blade heating experiment was conducted.

[0039] In this embodiment, the main airflow is electrically heated before entering the flow channel. The test blade is installed at the corner of the flow channel, and a cooling gas pipe and a temperature-measuring thermocouple are arranged on it. To ensure optical measurement accuracy, three sapphire optical windows are set around the blade, and the optical field multispectral measurement system is arranged at the lower window to obtain the full-field radiation information of the blade. The spectral transmittance of the window is calibrated and used for image correction to compensate for transmission loss.

[0040] As attached Figure 4 As shown, the numbers marked on the blade surface represent the absolute error between the temperature results calculated by LFMP and the thermocouple measurements, corresponding to measurement conditions with gas temperatures ranging from 673 to 823 K. Specifically, measurements were taken under conditions where the mainstream temperatures were 673 K to 823 K, without cooling air, to allow the blade surface to reach thermal equilibrium. 100 frames of multispectral optical field images were acquired at each temperature point, and after normalization and correction using calibration curves, a complete dataset was constructed, divided according to the proportions of training, testing, and validation sets.

[0041] During model training, a 723K blackbody image was selected as the reference input temperature. The emissivity of the blade computation region, ranging from 0 to 1, was estimated pixel-by-pixel by the network. Four GTX 2080 Ti graphics cards were used for parallel computation during training, with a dropout rate of 0.2, weight decay of 0.01, and a learning rate of 0.0001. Gradient accumulation was also employed to prevent overfitting. The model converged after 5000 training epochs, with an average validation loss of 0.01 and an inference time of approximately 3 seconds per frame.

[0042] The experimental results of this embodiment show that the absolute error of the blade temperature inversion method proposed in this invention is less than 10K under different operating conditions. It can accurately reconstruct the temperature distribution of the region of interest of the blade and reflect the temperature difference at the thermocouple position, showing strong accuracy and robustness.

[0043] Example 4: Building upon Example 3, to further verify the applicability of the proposed method under complex cooling conditions, the trained model was applied to a turbine blade film cooling experiment. In this example, background radiation intensity was obtained by briefly cooling the blade surface and used to correct the collected data. Subsequently, multispectral images of the optical field were acquired at different cooling gas mass flow rates (17 g / s, 14 g / s, 11 g / s, and 10 g / s) and compared with the inversion results of the two optimization methods (Fmincon and CSA).

[0044] The results of this embodiment show that all three methods can reflect the influence of cooling air flow on blade temperature distribution, but the method proposed in this invention performs better in terms of accuracy and spatial continuity. Specifically, as the cooling air flow decreases, the blade surface temperature increases, and the low-temperature region gradually shrinks. The method proposed in this invention can clearly depict this trend and obtain a smoother and more continuous temperature field distribution.

[0045] Furthermore, as shown in the appendix Figure 5 As shown in the figure, numbers 1-4 correspond to the thermocouple measuring points. The second to fourth columns of images show the temperature results calculated by the Fmincon and CSA algorithms and the present invention, respectively. The measurement conditions are cooling airflow rates Qc = 17, 14, 11, and 10 g / s. The values ​​in the figures indicate the comparison of the absolute errors of the three algorithms and the thermocouple measurements at the measuring points. Clearly, the method proposed in this invention maintains the highest accuracy under all four conditions and at all measuring points, with a maximum absolute error of less than 15K and a relative error of less than 2%. Specifically, under the conditions of 11 g / s and 10 g / s, the absolute error is controlled within 10K, and the relative error is less than 1%. In contrast, the errors of existing technologies are significantly larger, with a maximum absolute error exceeding 35K and a relative error approaching or exceeding 5%.

[0046] Example 5: As attached Figure 1 As shown, this invention proposes a multispectral temperature inversion method for optical fields based on physical information neural networks, including: Step S101: Obtain the blackbody light field image and the target light field image at the reference temperature as the original images, perform preprocessing on them by wavelength decoupling and division, and divide the preprocessed data into test set, training set and validation set; Step S102: Construct a light field multispectral physical information neural network (LFMP). This network adopts a two-layer U-Net architecture. First, image features are extracted through U-Net1. Then, the physical relationship between image features and temperature features is established using the embedded Planck's law module (PLM). Finally, the temperature field is output through U-Net2. Step S103: Train the optical field multispectral physical information neural network using the training set and validation set, and optimize the network parameters; Step S104: Transfer the trained network to the actual test data to realize the multispectral temperature inversion of the light field; Furthermore, the aforementioned optical field multispectral physical information neural network (LFMP) includes: Step S1041: Input multi-channel data into U-Net1, and through multi-layer operations, extract and downsample step by step to obtain multi-scale image features; Step S1042: The image features output by U-Net1 are input to the Planck's Law module (PLM), and pass through PLM1, PLM2, PLM3 and PLM4 units in sequence. With the help of the physical constraints of Planck's Law, the image features are converted into temperature features. Step S1043: The temperature features output by PLM are input into U-Net2, and through multi-layer operations, they are gradually upsampled and reconstructed to finally output the temperature field; The present invention also provides a method and system for optical field multispectral temperature inversion based on physical information neural network. The optical field multispectral temperature inversion system based on physical information neural network can be implemented by executing the process steps of the optical field multispectral temperature inversion method based on physical information neural network. That is, those skilled in the art can understand the optical field multispectral temperature inversion system based on physical information neural network as a preferred embodiment of the optical field multispectral temperature inversion method based on physical information neural network.

[0047] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0048] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for optical field multispectral temperature inversion based on physical information neural networks, characterized in that, include: Step S1: Establish a network architecture using a two-layer U-Net network; Step S2: Embed custom physical information; Step S3: Assign network training strategies; Step S4: Conduct network training based on the data, and balance the loss coefficient and adjust the training parameters through multiple rounds of optimization iterations; Step S5: Predict the current temperature in real time.

2. The optical field multispectral temperature inversion method based on physical information neural network according to claim 1, characterized in that, The dual-layer U-Net network in step S1 includes: U-Net1 and U-Net2; U-Net1 is the first layer U-Net, and U-Net2 is the second layer U-Net; each U-Net structure consists of an encoding layer and a decoding layer, with an input tensor dimension of [1, 16n, H, W], where n is the number of spectra to be calculated, and H and W are the image length and width, respectively; furthermore, U-Net1 can extract image radiation features, specifically referring to extracting the spatial and spectral features of the ratio image; U-Net2 can process image temperature features, specifically referring to convolving and fusing the physically processed features to output the final temperature prediction result.

3. The optical field multispectral temperature inversion method based on physical information neural network according to claim 1, characterized in that, The data flow of the network architecture in step S1 includes: firstly, the multispectral images of the light field obtained by the target and the blackbody reference at a known temperature are decomposed into multiple sub-images by spectral channels; then, the target sub-image is divided pixel by pixel by the blackbody sub-image of the corresponding channel to obtain a ratio image based on the reference temperature, thereby realizing the normalization of radiation information; after the above processing, the dataset is randomly divided and preprocessed to form the input for network training.

4. The optical field multispectral temperature inversion method based on physical information neural network according to claim 1, characterized in that, Steps S1 and S2 include: U-Net1 uses the calculated number of spectra as input, performs feature extraction using convolution kernels, and generates channel radiation feature maps. Subsequently, the channel radiation feature maps are input to a processing stage that embeds custom physical information. In this processing stage, information fusion and transformation are performed in combination with specific physical rules and constraints. After this processing, channel temperature features are generated through cascaded operations. Finally, these channel temperature features are sent to U-Net2 for further feature mining and processing. The temperature features of multiple channels are fused through convolution operations to output a temperature prediction map.

5. The optical field multispectral temperature inversion method based on physical information neural network according to claim 4, characterized in that, The cascaded operation specifically refers to the step-by-step optimization of the predicted temperature and radiation characteristics, incorporating knowledge from the physical domain. The physical information embedding step consists of multiple sequentially connected sub-steps. Each sub-step receives the ratio feature map and temperature feature map output from the previous stage and processes them through a multi-branch convolutional structure. Furthermore, within each sub-step, physical information is explicitly embedded to establish the physical relationship between radiance, emissivity, and temperature, and the radiation and thermal feature representations are updated through multi-level iterations.

6. The optical field multispectral temperature inversion method based on physical information neural network according to claim 1, characterized in that, Step S3 includes: co-training a custom multivariate loss function based on the principle of optical field multispectral radiation thermometry; further, the custom multivariate loss function includes: true temperature loss, prior temperature loss, channel consistency loss, and emissivity loss.

7. The optical field multispectral temperature inversion method based on physical information neural network according to claim 1, characterized in that, Specifically, steps S3 and S4 refer to constructing a composite loss function for the network architecture to balance data fidelity and physical constraints during training, as shown in the following formula: in, This is the total composite loss function; The loss between the network's predicted temperature and the actual temperature is the main loss term used to measure model performance. The loss is the difference between the network-predicted temperature and the prior numerical temperature. The temperature difference loss calculated for different spectral bands is used to ensure the spectral consistency of temperature prediction; The emissivity loss for network autonomous learning is to normalize the learned emissivity value to a physically valid boundary range and add a penalty term to suppress unreasonable predictions that exceed the range. to Let be the weight coefficients corresponding to different loss terms, and satisfy . The values ​​of the aforementioned weight coefficients depend on the training requirements of the model.

8. The optical field multispectral temperature inversion method based on physical information neural network according to claim 1, characterized in that, Specifically, step S5 refers to directly transferring the trained network model to test or validation data outside the training set. The network model can then achieve temperature prediction output based on the built-in data processing flow, without the need for model fine-tuning or secondary training.

9. A light field multispectral temperature inversion system based on a physical information neural network, characterized in that, The optical field multispectral temperature inversion method based on physical information neural networks as described in claims 1 to 8 includes: Module M1: Establishes a network architecture using a two-layer U-Net network; Module M2: Embeds custom physical information; Module M3: Assigns network training strategies; Module M4: Conducts network training based on data, and balances the loss coefficient and adjusts the training parameters through multiple rounds of optimization iteration; Module M5: Real-time temperature prediction.

10. The optical field multispectral temperature inversion system based on a physical information neural network according to claim 9, characterized in that, The dual-layer U-Net network in module M1 includes U-Net1 and U-Net2, where U-Net1 is the first layer U-Net and U-Net2 is the second layer U-Net. Each U-Net structure consists of an encoding layer and a decoding layer, with an input tensor dimension of [1, 16n, H, W], where n is the number of spectra to be calculated, and H and W are the image length and width, respectively. Furthermore, U-Net1 can extract image radiation features, specifically by extracting the spatial and spectral features of the ratio image. U-Net2 can process image temperature features, specifically by convolving and fusing the physically processed features to output the final temperature prediction result. The modules M1 and M2 include: U-Net1 takes the calculated spectral number as input, uses convolution kernels to extract features, and generates channel radiation feature maps. Subsequently, the channel radiation feature maps are input to the processing stage that embeds custom physical information. In this stage, information fusion and transformation are performed in combination with specific physical rules and constraints. After this processing, channel temperature features are generated through cascade operations. Finally, these channel temperature features are sent to U-Net2 for further feature mining and processing. The multi-channel temperature features are fused through convolution operations, and the output is a single temperature prediction map. The cascaded operation specifically refers to the step-by-step optimization of the predicted temperature and radiation characteristics, incorporating knowledge from the physical domain. The physical information embedding module consists of multiple sequentially connected sub-modules. Each sub-module receives the ratio feature map and temperature feature map output from the previous stage and processes them through a multi-branch convolutional structure. Furthermore, within each sub-module, physical information is explicitly embedded to establish the physical relationship between radiance, emissivity, and temperature, and the radiation and thermal feature representations are updated through multi-level iterations.

Images