Target three-dimensional temperature field prediction method based on diffusion model generation

Abstract

The invention discloses a target three-dimensional temperature field prediction method based on a generated diffusion model. The method comprises the following steps: establishing a simplified geometric model of a target, and determining key parameters such as a target material, a power system and weather; establishing a target outer boundary heat balance equation, and respectively calculating surface temperature fields under different working conditions as basic data; establishing a multi-modal feature extraction module, and performing deep feature extraction and fusion on the three-dimensional coordinates, the node temperature and the key parameters; establishing a generation diffusion model to train fused features, and learning mapping from feature noise to real temperature features; according to the input parameters and the three-dimensional shape constraint, a predicted node temperature value is output through a reverse denoising process, diffusion loss is calculated, and finally a complete three-dimensional temperature field is generated. According to the invention, the problem of low efficiency of traditional simulation is effectively solved, and rapid and accurate prediction of the target temperature field is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of infrared scene simulation, and particularly relates to a target three-dimensional temperature field prediction method based on a generated diffusion model. BACKGROUND

[0002] With the continuous development of infrared detection technology, accurate acquisition of target infrared radiation characteristics has become the core prerequisite for breakthroughs and application landing in related fields. Whether it is technical research and development, scheme optimization, simulation testing or actual application scenarios, there is an urgent need for accurate characterization and data support of target infrared characteristics, and the technical value directly affects the performance upper limit and application effectiveness of related systems.

[0003] The traditional prediction method of target infrared radiation characteristics is to use a numerical calculation model, and the core idea is to model the heat and radiation exchange between the target and the environment to predict the infrared radiation characteristics of the target. This method considers the global thermal environment and material parameter conditions of the target and the environment, and solves the numerical solution of the conservation equation to complete the infrared radiation characteristic simulation.

[0004] Target infrared characteristic modeling usually needs to be based on known detailed structure parameters, with dynamic characteristic parameters, material parameters, surface optical parameters and external environment parameters as inputs. The calculation of target infrared radiation characteristics depends on the accurate calculation of the target temperature field, but the traditional simulation method is time-consuming and laborious, and it is difficult to meet the real-time requirements in complex environments.

[0005] Traditional deep learning methods face great challenges in target parameter inversion and three-dimensional temperature field prediction due to the problems of sample scarcity, low utilization and weak model universality caused by changing environments. Existing researches are mostly limited to two-dimensional visual tasks such as conversion or fusion between infrared and visible light images, and lack of research on direct mapping from parameter space to three-dimensional physical field. SUMMARY

[0006] The application provides a target three-dimensional temperature field prediction method based on a generated diffusion model.

[0007] The technical scheme for achieving the application is as follows: A target three-dimensional temperature field prediction method based on a generated diffusion model, comprising the following steps:

[0008] Step 1: Establish a simplified geometric model of the target and determine the target parameters;

[0009] Step 2: Establish a target outer boundary heat balance equation and calculate the target surface temperature field under different working conditions as basic data;

[0010] Step 3, the parameter physical characteristics, node temperature characteristics, geometric characteristics and time step characteristics are extracted and fused by using a multi-modal feature extraction module to obtain multi-modal fusion features;

[0011] Step 4, training the multi-modal fusion features by using a generative diffusion module, learning the mapping from feature noise to real temperature features, and saving the model weight of the trained generative diffusion module;

[0012] Step 5, inputting the material parameters, dynamic system parameters and meteorological parameters of the target and the three-dimensional coordinates of the target into the generative diffusion model, and calling the model weight of the generative diffusion module of step 4, gradually denoising the pure Gaussian noise temperature field, and finally generating a predicted three-dimensional temperature field;

[0013] Step 6, using a differentiable renderer to render the predicted three-dimensional temperature field in multiple directions to generate a two-dimensional temperature image.

[0014] Compared with the prior art, the present application has the following advantages: 1) the present application trains the fused features based on a generative diffusion model, learns the mapping from feature noise to real temperature features, and quickly and accurately predicts the three-dimensional temperature field of the target by only inputting parameters and three-dimensional shape constraints, meeting the complex and variable environmental conditions. 2) The present application uses a differentiable renderer to accurately and effectively render two-dimensional temperature images in different directions according to the predicted three-dimensional temperature field, and solves the problem of poor prediction accuracy of individual temperature points by using temperature and grid smoothing processing, greatly increasing flexibility.

[0015] The present application will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flow chart of a target three-dimensional temperature field prediction method based on a generative diffusion model.

[0017] Figure 2 is a target three-dimensional temperature field prediction algorithm architecture diagram based on a generative diffusion model.

[0018] Figure 3 is a schematic diagram of the forward and reverse processes of the training of the generative diffusion model. DETAILED DESCRIPTION

[0019] In combination Figure 1 A target three-dimensional temperature field prediction method based on a generative diffusion model, comprising the following steps:

[0020] Step 1, using Solidworks software, according to the actual geometric shape of the target to simplify modeling. The target is divided into calculation grid by using the ICEM software of the grid division of ANSYS platform, and the grid of the target and the background is non-structured grid. Through field test or data query, the information of different target materials, power system and meteorological parameters is determined.

[0021] The present application is aimed at the target under the thermal static working condition, and the actual geometric shape of the target is simplified to model; the target material, power system and meteorological parameters are determined through test or data. The material parameters mainly include: density, specific heat capacity, thermal conductivity, surface emissivity and solar absorption rate; the power system parameters mainly include: the surface temperature of the heat source in the target power cabin, the temperature and mass flow rate of the inlet air and the temperature and mass flow rate of the exhaust smoke; the meteorological parameters mainly include: air temperature, wind speed, humidity, direct solar radiation and solar scattering radiation.

[0022] Step 2, the target boundary heat balance equation is established, and the surface temperature field under different working conditions is calculated as basic data; specifically including:

[0023] Step 21, the target boundary heat balance equation is established, and the factors affecting the outer surface temperature of the target mainly include solar radiation, sky background radiation, ground background radiation and air convection heat transfer, etc.:

[0024]

[0025] In the formula, The net radiation heat flux is: The solar radiation heat flow is: The absorption rate of the target surface to the solar radiation is: The absorption rate of the target to the ground and sky background long wave radiation is: The sky background radiation is: The ground background radiation is: The Stefan-Boltzmann constant is 5.67x10 -8 W / (m 2 ·K 4 ); The emissivity of the target surface is: The convection heat transfer coefficient is: The temperature of the target surface is: The air temperature is:

[0026] Step 22, based on the Fluent simulation calculation software, according to the established target outer boundary heat balance equation and the key parameters of thermal physical properties, dynamic system and meteorology determined in step 1, the orthogonal experiment method is used to calculate the target surface temperature field under different working conditions. The results of the target three-dimensional temperature field mainly include the three-dimensional coordinates of the target and the corresponding node temperature values. The first 20 examples are classified as a training set, and the model is trained to master the nonlinear mapping rule between multiple physical parameters and temperature field through supervised learning. The remaining 5 examples are used as a test set that does not participate in training, to verify the effect of the model on new working conditions.

[0027] Step 3, establish a multi-modal feature extraction module, and perform deep feature extraction and fusion on the three-dimensional coordinates of the target temperature field, the node temperature values, and the target material, dynamic system and meteorological parameters in step 1. The specific steps are as follows:

[0028] Step 31, construct a convolution feature encoding module to process material parameters, dynamic system parameters and meteorological parameters. Through multi-layer convolution operation and nonlinear activation function mapping, the low-dimensional original physical parameters are projected into a high-dimensional feature space to obtain parameter physical features C;

[0029] Step 32, construct a multi-layer perceptron (MLP) feedforward neural network to extract features of node temperature values. Use fully connected layers and nonlinear activation functions to project low-dimensional scalar temperature data into a high-dimensional latent feature space to obtain node temperature features P;

[0030] Step 33, construct a PointNet encoder for point cloud feature extraction, taking the three-dimensional coordinates of the target temperature field obtained in step 2 as input. Perform uniform nonlinear high-dimensional mapping on each spatial point through a multi-layer perceptron with shared weights, and introduce a max-pooling function to globally aggregate point-by-point high-dimensional features, preserving the key topological structure in the target three-dimensional space, to obtain the geometric features G of the target;

[0031] Step 34, construct a time encoder for extracting time steps. Through sinusoidal positional encoding, use different sine and cosine functions in the Transformer to convert discrete time steps t into high-dimensional vectors to obtain time step features T;

[0032] Step 35, use a fusion module to extract parameter physical features, node temperature features, target geometric features and time step features in the previous four steps, and use a feature concatenation strategy to stack dimensions, effectively fusing multi-modal features and providing complete feature input for subsequent model reasoning. The specific model structure is shown in Figure 2 .

[0033] Step 4, training the multi-modal fusion features with the generative diffusion module to learn the mapping from feature noise to real temperature features, saving the model weights of the trained generative diffusion module. The generative diffusion module includes an alignment module and a diffusion module.

[0034] Step 41, introducing a contrastive language-image pre-training (CLIP) alignment mechanism to build a dual-tower mapping structure containing temperature branches and physical branches. This module uses two independent fully connected mapping layers as projection heads to process the features extracted in step 3: compressing the high-dimensional node temperature features P and mapping the low-dimensional physical parameter features C to the same dimension alignment feature space. On this basis, L2 normalization is used to eliminate the dimensional influence, and a learnable scaling factor is used for dot product interaction operation. This process effectively completes the measurement and calibration of multi-modal features, ensuring the semantic consistency of parameter physical features P and node temperature features C, and finally obtaining the alignment features extracted by the CLIP module, providing high-quality alignment features for the subsequent generation model.

[0035] Step 42, the diffusion module takes the multi-modal features (including parameter physical features, node temperature features, target geometric features, and time step features) fused in step 3 as input features, and starts the training process of the diffusion module based on the feature input. Through iterative optimization of model parameters, the mapping relationship between features and target output is learned. The diffusion module uses a U-Net architecture based on conditional point clouds. The network consists of an encoder for downsampling and a decoder for upsampling, and uses skip connections to fuse deep semantic and shallow details. In the middle layer of the network, the alignment features extracted by the CLIP module in step 41 are injected as conditions through cross-attention mechanisms. Figure 3 As shown in the forward noise addition process, the diffusion module training gradually adds noise to the original real temperature field to obtain a pure Gaussian noise temperature field The total time step T of the diffusion process is 1000 steps, and the [0, 1000] time step interval of the diffusion process is set. After training, the model weights are saved.

[0036] During the training of the generative diffusion module, a diffusion loss function is added to calculate the difference between the real added noise and the network predicted noise. In the construction of the diffusion loss function, the node temperature features P and the parameter physical features C are included in the conditional constraint system, and through this constraint mechanism, the model optimization target is focused on the conditional distribution learning strictly limited by the physical law and node temperature related constraints, instead of the unbiased learning of general data distribution.

[0037] The calculation formula of the diffusion loss function is:

[0038]

[0039] wherein: is the time step; is the real noise; is the model prediction noise; is the node temperature feature; is the parameter physical feature; is the noisy temperature field of the time step; is the mathematical expectation, i.e., the average, of the weighted average level of the noise prediction error square.

[0040] The CLIP alignment module iteratively optimizes the normalized parameter physical features and node temperature features by introducing a contrastive loss function, forcing the matching features to present a dense and clustered distribution in the embedding space, while constraining the unmatched features to maintain a significant distance interval in the space, and finally achieving the precise semantic alignment of the two types of features. The calculation formula of the contrastive loss function is:

[0041]

[0042] wherein, is the cross-entropy loss; is the target label; is a similarity matrix, and the elements in the similarity matrix are specifically:

[0043]

[0044] wherein, is the similarity between the node temperature feature and the parameter physical feature ; is the node temperature feature vector embedded after L2 normalization of the node temperature feature , is the node temperature feature embedding of the sample, and the sample refers to the data instance input in one training iteration; is the parameter physical feature vector embedded after L2 normalization of the parameter physical feature , is the node temperature feature embedding of the sample, and the sample refers to the data instance input in one training iteration; is the vector dot product, since the node temperature feature vector and the parameter physical feature vector L2 normalization has been completed, and its dot product is mathematically equivalent to cosine similarity. This is a learnable scaling factor used to scale the cosine similarity, preventing the gradient from vanishing due to excessively small dot product values, thereby accelerating model convergence.

[0045] By leveraging the synergistic effect of the diffusion generation module and the CLIP alignment module, the diffusion loss is minimized, the physical rationality of the mapping relationship between the physical feature C and the node temperature feature P is maximized, and the model learning process is ensured to both fit the data distribution and remain true to the physical scenario.

[0046] Therefore, combining the contrastive loss function and the diffusion loss function, the formula for calculating the total loss of the diffusion module is as follows:

[0047]

[0048] in, This is a hyperparameter used to balance two objectives: generation quality and alignment between parametric physical characteristics and node temperature characteristics.

[0049] Step 5: Input the target's material parameters, dynamic system parameters, meteorological parameters, and the target's three-dimensional coordinates into the generation and diffusion model. Then, call the model weights trained in Step 4 to gradually denoise the pure Gaussian noise temperature field, ultimately generating the predicted three-dimensional temperature field. Specifically, this includes:

[0050] like Figure 3 The inverse denoising process shown first uses the material, dynamic system, meteorological parameters, and the target's three-dimensional geometric coordinates as input conditions provided by the test set in step 2. In the inference configuration, the inverse denoising sampling steps are set to 200 steps. The diffusion model scheduler (DDPM) is used to perform uniform resampling within the [0, 1000] time step interval set in the model training in step 4, generating 200 discrete inference time points to accelerate the generation process.

[0051] By calling the converged model weights trained in step 4, iterative denoising is initiated with a pure Gaussian noise temperature field following a standard normal distribution as the initial state. In each inverse calculation step, the CLIP alignment module plays a crucial role, mapping the input material, dynamic system, and meteorological parameters to a latent feature space aligned with the node temperatures, and injecting them as strongly consistent conditional features to generate the diffusion model. Under the joint constraints of three-dimensional geometric coordinates, and guided by the time step of the diffusion model scheduler, the model gradually strips away random components from the Gaussian noise, reconstructing a reasonable temperature distribution that conforms to specific physical laws. Finally, after 200 iterations, the predicted three-dimensional temperature field is output at t=0, achieving the transformation from a Gaussian noise temperature field... Gradually denoise to obtain the predicted temperature field of the original frame number. .

[0052] Step 6, multi-view rendering of the predicted three-dimensional temperature field using a differentiable renderer to generate a two-dimensional temperature image.

[0053] The present application uses a differentiable renderer based on Pytorch3d to perform multi-view rendering on the predicted three-dimensional temperature field to generate a two-dimensional temperature image. The error between the three-dimensional temperature field and the two-dimensional temperature image is analyzed comprehensively to make a comprehensive judgment. In order to intuitively evaluate the prediction accuracy of node temperature, the present application uses the mean absolute error (MAE) and the mean absolute percentage error (MAPE).

[0054]

[0055]

[0056] In the formula: is the number of test sets, is the number of temperature nodes, and is the predicted value and the true value of the node temperature.

[0057] In addition to the node temperature value evaluation standard, the present application also uses the structural similarity index (SSIM) and the peak signal-to-noise ratio (PSNR) to measure the authenticity of the rendered two-dimensional temperature image.

[0058] In this embodiment, the material, power system and meteorological parameters of the test set of 5 examples are used as input conditions to predict the three-dimensional temperature field of the target.

[0059] As described above, the present embodiment is aimed at the target under thermal static working condition, and based on the target three-dimensional temperature field prediction method of generating diffusion model, the three-dimensional temperature field of the target is quickly and accurately predicted, which provides strong technical support for subsequent target infrared characteristic research scheme evaluation and optimization scheme design.

Claims

1. A method for predicting a target three-dimensional temperature field based on a generative diffusion model, characterized in that, Comprise: Step 1, a simplified geometric model of the target is established, and target parameters are determined; Step 2, a target outer boundary heat balance equation is established, and the target surface temperature field under different working conditions is calculated as basic data; Step 3, a multi-modal feature extraction module is used to extract and fuse parameter physical features, node temperature features, geometric features and time step features, and obtain multi-modal fusion features; Step 4, the multi-modal fusion features are trained using a generative diffusion module, the mapping from feature noise to real temperature features is learned, and the model weight of the trained generative diffusion module is saved; Step 5, the material parameters, power system parameters and meteorological parameters of the target and the three-dimensional coordinates of the target are input into the generative diffusion model, and the model weight of the generative diffusion module in step 4 is called, and the pure Gaussian noise temperature field is gradually denoised, and finally a predicted three-dimensional temperature field is generated; Step 6, a differentiable renderer is used to render the predicted three-dimensional temperature field in multiple directions to generate a two-dimensional temperature image.

2. The method of claim 1, wherein, The target parameters determined in step 1 include target material parameters, power system parameters and meteorological parameters, wherein the material parameters include density, specific heat capacity, thermal conductivity, surface emissivity and solar absorption rate; the power system parameters include the surface temperature of the heat source in the target power cabin, the temperature and mass flow rate of the inlet air, and the temperature and mass flow rate of the exhaust smoke; and the meteorological parameters include air temperature, wind speed, humidity, direct solar radiation and scattered solar radiation. 3.The target three-dimensional temperature field prediction method based on the generated diffusion model according to claim 1, wherein, The specific method for establishing the target outer boundary heat balance equation and calculating the surface temperature field under different working conditions as basic data is as follows: Step 21, a target boundary heat balance equation is established, specifically: ； wherein is the solar heat flux; is the surface absorptivity of the target to solar radiation; is the absorptivity of the target to ground and sky background long wave radiation; is the reflectivity of the ground to solar radiation; and represent the emissivity of the ground and the target surface, respectively; , and represent the ground temperature, the equivalent temperature of the sky background radiation and the surface temperature of the target, respectively. ； wherein is the net radiative heat flux; is the solar radiative heat flux; is the absorptivity of the target surface to solar radiation; is the absorptivity of the target to ground and sky background long wave radiation; is the sky background radiation; is the ground background radiation; is the Stefan-Boltzmann constant; is the emissivity of the target surface; is the convective heat transfer coefficient; is the temperature of the target surface; is the air temperature; Step 22, based on the Fluent simulation calculation software, according to the established target outer boundary heat balance equation and the target parameters determined in step 1, the orthogonal experiment method is used to calculate the target surface temperature field under different working conditions.

4. The method of claim 1, wherein, The multi-modal feature extraction module comprises a convolution feature encoding module, a multi-layer perception feedforward neural network, a PointNet architecture encoder, a time encoder and a fusion module; the convolution feature encoding module is used to process the target parameters, and through multi-layer convolution operation and mapping of a nonlinear activation function, the low-dimensional target parameters are projected into a high-dimensional feature space to obtain parameter physical features; the multi-layer perception feedforward neural network extracts features of node temperature values to obtain node temperature features; the PointNet architecture encoder is used to obtain geometric features of the target three-dimensional model according to the three-dimensional coordinates of the target temperature field; and the fusion module is used to stack the dimensions of the parameter physical features, the node temperature features, the target geometric features and the time step features by using a feature splicing strategy, so as to effectively fuse the multi-modal features.

5. The method of claim 4, wherein, The PointNet architecture encoder takes the three-dimensional coordinates of the target temperature field as input, performs uniform nonlinear high-dimensional mapping on each spatial point through a multi-layer perception machine with shared weights, and introduces a max pooling function to globally aggregate the point-by-point high-dimensional features, thereby retaining the key topological structure in the target three-dimensional space and obtaining the geometric features of the target.

6. The method of claim 1, wherein, The generation diffusion module comprises an alignment module and a diffusion module, the diffusion module takes the multi-modal fusion feature as an input feature, realizes gradual noise addition to the node temperature feature to obtain a pure Gaussian noise temperature field, and iteratively optimizes the diffusion module model parameters to learn a mapping relationship between the feature and a target output; The alignment module is used to realize semantic alignment of the parameter physical feature and the node temperature feature, and through L2 normalization, the loss is minimized by optimizing the directional alignment between the parameter physical feature and the node temperature feature, and the normalized parameter physical feature and node temperature feature are taken as vector embeddings of the parameter physical feature and the node temperature feature of the alignment module.

7. The method of claim 1, wherein, A loss function used in the training process of the generation diffusion module is: ； wherein is a hyperparameter, is a diffusion loss function, is a contrastive loss function.

8. The method of claim 7, wherein, The diffusion loss function is specifically: ； wherein is the time step; is the real noise; is the model prediction noise; is the node temperature feature; is the parameter physical feature, is the first time step noisy temperature field; is the mathematical expectation, i.e. the mean.

9. The method of claim 7, wherein the generating diffusion model based target three-dimensional temperature field prediction method is characterized by, The contrast loss function is specifically: ； wherein is the cross-entropy loss; is the target label; is the similarity matrix, denotes the transpose, and B is the batch size.​ 10. The method of claim 9, wherein, The elements in the similarity matrix are specifically: , in, For the first Temperature characteristics of each node With the Physical characteristics of each parameter The degree of similarity between them; To study the temperature characteristics of nodes The embedded node temperature feature vector after L2 normalization; For the physical characteristics of parameters The embedded parameter physical feature vector after L2 normalization; This is a learnable scaling factor.

Images