A target three-dimensional temperature field prediction method based on a generative diffusion model

Abstract

The application discloses a target three-dimensional temperature field prediction method based on a generated diffusion model. The method steps are as follows: a simplified geometric model of the target is established, and key parameters such as target material, power system and meteorology are determined; an outer boundary heat balance equation of the target is established, and the surface temperature field under different working conditions is calculated as basic data; a multi-modal feature extraction module is established to perform deep feature extraction and fusion on three-dimensional coordinates, node temperatures and key parameters; a generated diffusion model is established to train the fused features, learn the mapping from feature noise to real temperature features; according to input parameters and three-dimensional shape constraints, the predicted node temperature value is output through a reverse denoising process, and the diffusion loss is calculated, and finally a complete three-dimensional temperature field is generated. The application effectively solves the problem of low efficiency of traditional simulation and realizes fast and accurate prediction of the target temperature field.

Description

Technical Field

[0001] This invention belongs to the field of infrared scene simulation technology, specifically a method for predicting the three-dimensional temperature field of a target based on a generative diffusion model. Background Technology

[0002] With the continuous development of infrared detection technology, accurately acquiring the infrared radiation characteristics of targets has become a core prerequisite for technological breakthroughs and application implementation in related fields. Whether in technology research and development, solution optimization, simulation testing, or practical application scenarios, there is an urgent need for accurate characterization and data support of target infrared characteristics, and its technological value directly affects the performance ceiling and application effectiveness of related systems.

[0003] Traditional methods for predicting the infrared radiation characteristics of targets utilize numerical calculation models. The core idea is to model the heat and radiation exchange between the target and its environment to predict the target's infrared radiation characteristics. This method comprehensively considers the global thermal environment and material parameters of both the target and its surroundings, solving the conservation equations numerically to simulate the infrared radiation characteristics.

[0004] Modeling the infrared characteristics of a target typically requires knowing detailed structural parameters, with dynamic characteristics, material parameters, surface optical parameters, and external environmental parameters as inputs. Calculating the infrared radiation characteristics of a target depends on accurately calculating its temperature field, but traditional simulation methods are time-consuming and labor-intensive, making it difficult to meet the real-time requirements of complex environments.

[0005] Traditional deep learning methods face significant challenges in target parameter inversion and 3D temperature field prediction due to problems such as sample scarcity, low utilization, and weak model universality caused by environmental variability. Existing research is mostly limited to 2D vision tasks such as conversion or fusion between infrared and visible light images, lacking research on direct mapping from parameter space to 3D physical fields. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting the target three-dimensional temperature field based on a generative diffusion model.

[0007] The technical solution to achieve the objective of this invention is as follows: A method for predicting a target three-dimensional temperature field based on a generative 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 the thermal balance equation of the outer boundary of the target and calculate the surface temperature field of the target under different working conditions as the basic data;

[0010] Step 3: Use the multimodal feature extraction module to extract and fuse the parameter physical features, node temperature features, geometric features, and time step features to obtain multimodal fused features;

[0011] Step 4: Use the generation-diffusion module to train the multimodal fusion features, learn the mapping from feature noise to real temperature features, and save the generation-diffusion module model weights after training.

[0012] 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, and call the model weights of the generation and diffusion module in Step 4 to gradually denoise the pure Gaussian noise temperature field, and finally generate the predicted three-dimensional temperature field.

[0013] Step 6: Use a differentiable renderer to render the predicted three-dimensional temperature field from multiple angles to generate a two-dimensional temperature image.

[0014] Compared with existing technologies, the significant advantages of this invention are: 1) This invention trains the fused features based on a generative diffusion model, learning the mapping from feature noise to real temperature features. Using only input parameters and 3D shape constraints, it quickly and accurately predicts the 3D temperature field of the target, meeting the needs of complex and variable environments. 2) This invention utilizes a differentiable renderer to accurately and effectively render 2D temperature images from different orientations based on the predicted 3D temperature field. By using temperature and mesh smoothing, it solves the problem of poor prediction accuracy at individual temperature points, greatly increasing flexibility.

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0016] Figure 1 This is a flowchart of a method for predicting the target three-dimensional temperature field based on a generative diffusion model.

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

[0018] Figure 3 This is a schematic diagram of the forward and backward processes of training the generative diffusion model. Detailed Implementation

[0019] Combination Figure 1 A method for predicting a target three-dimensional temperature field based on a generative diffusion model includes the following steps:

[0020] Step 1: Using Solidworks software, simplify the model based on the actual geometry of the target. Use the ICEM software (a mesh generation tool on the ANSYS platform) to create a computational mesh for the target; both the target and background meshes are unstructured. Through field experiments or literature review, clarify the material, dynamic system, and meteorological parameters of different targets.

[0021] This invention studies targets under thermal static conditions, simplifying and modeling them based on their actual geometry; and clarifies the target's materials, propulsion system, and meteorological parameters through experiments or data. Material parameters mainly include: density, specific heat capacity, thermal conductivity, surface emissivity, and solar absorptivity; propulsion system parameters mainly include: surface temperature of the heat source within the target's propulsion compartment, intake air temperature and mass flow rate, and exhaust smoke temperature and mass flow rate; meteorological parameters mainly include: air temperature, wind speed, humidity, direct solar radiation, and diffuse solar radiation.

[0022] Step 2: Establish the target boundary thermal equilibrium equation and calculate the surface temperature field under different operating conditions as the basic data; specifically including:

[0023] Step 21: Establish the target boundary heat balance equation. The main factors affecting the target's outer surface temperature include solar radiation, sky background radiation, ground background radiation, and air convection heat transfer.

[0024]

[0025] In the formula, Net radiative heat flux; It is the heat flow from solar radiation; The absorption rate of solar radiation by the target surface; The target is the absorption rate of longwave radiation from the ground and sky background. Background radiation from the sky; This refers to ground background radiation. The Stefan-Boltzmann constant has a value of 5.67 × 10⁻⁶. -8 W / (m 2 ·K 4 ); The emissivity of the target surface; The convective heat transfer coefficient; The temperature of the target surface; This refers to the air temperature.

[0026] Step 22: Based on Fluent simulation software, and according to the established target outer boundary thermal balance equation and the key parameters such as thermal properties, dynamic system, and meteorology determined in Step 1, the target surface temperature field under different working conditions is calculated using the orthogonal experimental method. The result of the target's three-dimensional temperature field mainly includes the target's three-dimensional coordinates and the corresponding nodal temperature values. The first 20 examples are assigned to the training set, and supervised learning is used to enable the model to master the nonlinear mapping law between multiple physical parameters and the temperature field. The remaining 5 examples are used as a test set not involved in training to verify the model's performance in extrapolating new working conditions.

[0027] Step 3: Establish a multimodal feature extraction module to perform deep feature extraction and fusion on the three-dimensional coordinates and nodal temperature values ​​of the target temperature field in Step 2, as well as the target material, dynamic system, and meteorological parameters in Step 1. The specific steps are as follows:

[0028] Step 31: Construct a convolutional feature encoding module to process material parameters, dynamic system parameters, and meteorological parameters respectively. Through multi-layer convolution operations and mapping with nonlinear activation functions, the low-dimensional original physical parameters are projected into a high-dimensional feature space to obtain the parametric physical features C;

[0029] Step 32: Construct a feedforward neural network of a multilayer perceptron (MLP) to extract features from node temperature values. Using fully connected layers and nonlinear activation functions, project the low-dimensional scalar temperature data into a high-dimensional latent feature space to obtain the node temperature features P.

[0030] Step 33: Construct a PointNet encoder for point cloud feature extraction. Using the three-dimensional coordinates of the target temperature field obtained in Step 2 as input, perform a unified nonlinear high-dimensional mapping on each spatial point through a multilayer perceptron with shared weights, and introduce a max pooling function to globally aggregate the point-by-point high-dimensional features, retaining the key topological structure in the target's three-dimensional space, and obtain the target's geometric features G.

[0031] Step 34: Construct a time encoder for extracting time steps. Using sinusoidal position encoding, different sine and cosine functions in the Transformer are used to convert discrete time steps t into high-dimensional vectors to obtain time step features T.

[0032] Step 35: Using the fusion module, the physical features of the parameters, node temperature features, target geometric features, and time step features extracted in the previous four steps are superimposed using a feature concatenation strategy to achieve effective fusion of multimodal features, providing complete feature input for subsequent model inference. The specific model structure is as follows: Figure 2 As shown.

[0033] Step 4: Use the generation and diffusion module to train the multimodal fusion features, learn the mapping from feature noise to real temperature features, and save the model weights of the generation and diffusion module after training. The generation and diffusion module includes an alignment module and a diffusion module.

[0034] Step 41: Introduce the Contrastive Language-Image Pre-training (CLIP) alignment mechanism to construct a dual-tower mapping structure containing temperature and physical branches. This module uses two independent fully connected mapping layers as projection heads to process the features extracted in Step 3: compressing and mapping the high-dimensional node temperature features P, and upscaling and mapping the low-dimensional physical parameter features C, transforming both into an aligned feature space with consistent dimensions. Based on this, L2 normalization is used to eliminate the influence of dimensions, and a learnable scaling factor is used for dot product interaction. This process effectively completes the measurement and calibration of multimodal features, ensuring the semantic consistency between the physical parameter features P and the node temperature features C, ultimately obtaining the aligned features extracted by the CLIP module, providing high-quality aligned conditional features for subsequent generative models.

[0035] Step 42: The diffusion module uses the multimodal features fused in Step 3 (specifically including parametric physical features, node temperature features, target geometric features, and time step features) as input features. Based on this input, the training process of the diffusion module is initiated, and the model parameters are iteratively optimized to learn the mapping relationship between the features and the target output. The diffusion module adopts a U-Net architecture based on conditional point clouds. This network consists of a downsampled encoder and an upsampled decoder, utilizing skip connections to fuse deep semantics and shallow details. In the middle layers of the network, the alignment features extracted by the CLIP module in Step 41 are conditionally injected through a cross-attention mechanism. Figure 3 As shown in the forward noise addition process, the diffusion module training is performed on the original real temperature field. Gradually adding noise to obtain a pure Gaussian noise temperature field The total number of time steps T in the diffusion process is 1000, and the time step interval of [0, 1000] is set for the diffusion process. After training, the model weights are saved.

[0036] During the training of the diffusion module, a diffusion loss function is added to calculate the diffusion loss between the actual noise and the noise predicted by the network. In the construction of the diffusion loss function, the node temperature feature P and the physical feature C of the parameters are incorporated into the conditional constraint system. Through this constraint mechanism, the model optimization objective is shifted from unbiased learning of the general data distribution to conditional distribution learning that is strictly limited by physical laws and the correlation between node temperature and the data.

[0037] The formula for calculating the diffusion loss function is:

[0038]

[0039] In the formula: For time steps; This is real noise; Predict noise for the model; For node temperature characteristics; For physical characteristics of parameters; For the first Noisy temperature field at time step; The mathematical expectation, i.e. the average value, is used to calculate the weighted average level of the squared noise prediction error.

[0040] The CLIP alignment module iteratively optimizes the normalized physical features and node temperature features by introducing a contrastive loss function. This forces matching features to exhibit a dense clustered distribution in the embedding space, while simultaneously constraining mismatched features to maintain a significant distance interval within the space, ultimately achieving semantically accurate alignment between the two types of features. The formula for calculating the contrastive loss function is:

[0041]

[0042] in, Cross-entropy loss; For target labels; It is The similarity matrix, specifically the elements of the similarity matrix are:

[0043]

[0044] 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 first The node temperature feature embedding of the nth sample, where a sample refers to the nth input in a training iteration. One data instance; For the physical characteristics of parameters The embedded parameter physical feature vector after L2 normalization. For the first The node temperature feature embedding of the nth sample, where a sample refers to the nth input in a training iteration. One data instance; It is a vector dot product, due to the nodal temperature feature vector. With parameter physical eigenvectors 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: Use a differentiable renderer to render the predicted three-dimensional temperature field from multiple angles to generate a two-dimensional temperature image.

[0053] This invention utilizes the PyTorch3d differentiable renderer to render the predicted 3D temperature field from multiple angles to generate a 2D temperature image. A comprehensive evaluation is then performed by analyzing the errors between the 3D temperature field and the 2D temperature image. To intuitively evaluate the prediction accuracy of nodal temperatures, this invention employs mean absolute error (MAE) and mean absolute percentage error (MAPE).

[0054]

[0055]

[0056] In the formula: The number of test sets, This represents the number of temperature nodes. and These are the predicted and actual values ​​of the node temperature.

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

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

[0059] As described above, this embodiment, based on a target three-dimensional temperature field prediction method using a generated diffusion model, rapidly and accurately predicts the three-dimensional temperature field of a target under thermal static conditions, providing strong technical support for the evaluation and optimization of subsequent infrared characteristic research schemes for the target.

Claims

1. A method for predicting a target three-dimensional temperature field based on a generative diffusion model, characterized in that, include: Step 1: Establish a simplified geometric model of the target and determine the target parameters; Step 2: Establish the thermal equilibrium equation of the target's outer boundary and calculate the target surface temperature field under different operating conditions as the basic data. The specific method is as follows: Step 21: Establish the target boundary heat balance equation, specifically as follows: ； In the formula, It is solar heat flow; The target is the surface absorptivity of solar radiation; The target is the absorption rate of longwave radiation from the ground and sky background. The reflectivity of the Earth to solar radiation; and These represent the emissivity of the ground and the target surface, respectively. , as well as These represent the ground temperature, the equivalent temperature of the sky background radiation, and the surface temperature of the target, respectively. ； In the formula, Net radiative heat flux; It is the heat flow from solar radiation; The absorption rate of solar radiation by the target surface; The target is the absorption rate of longwave radiation from the ground and sky background. Background radiation from the sky; This refers to ground background radiation. It is the Stefan-Boltzmann constant; The emissivity of the target surface; The convective heat transfer coefficient; The temperature of the target surface; Air temperature; Step 22: Based on Fluent simulation software, according to the established target outer boundary thermal balance equation and the target parameters determined in Step 1, calculate the target surface temperature field under different working conditions using the orthogonal experimental method. Step 3: Use the multimodal feature extraction module to extract and fuse the parameter physical features, node temperature features, geometric features, and time step features to obtain multimodal fused features; Step 4: Use the generation and diffusion module to train the multimodal fusion features, learn the mapping from feature noise to real temperature features, and save the model weights of the generation and diffusion module after training. The generation and diffusion module includes an alignment module and a diffusion module. The diffusion module uses the multimodal fusion features as input features to gradually add noise to the node temperature features to obtain a pure Gaussian noise temperature field. The model parameters of the diffusion module are iteratively optimized to learn the mapping relationship between features and target output. The alignment module is used to achieve semantic alignment between the physical features of the parameters and the temperature features of the nodes. By using L2 normalization, the directional alignment between the physical features of the parameters and the temperature features of the nodes is optimized to minimize the loss. The normalized physical features of the parameters and the temperature features of the nodes are then used as vector embeddings of the physical features of the parameters and the temperature features of the nodes in the alignment module. 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, and call the model weights of the generation and diffusion module in Step 4 to gradually denoise the pure Gaussian noise temperature field, and finally generate the predicted three-dimensional temperature field. Step 6: Use a differentiable renderer to render the predicted three-dimensional temperature field from multiple angles to generate a two-dimensional temperature image.

2. The method for predicting the target three-dimensional temperature field based on the generative diffusion model according to claim 1, characterized in that, The target parameters determined in step 1 include target material parameters, power system parameters, and meteorological parameters. Among them, material parameters include: density, specific heat capacity, thermal conductivity, surface emissivity, and solar absorptivity; power system parameters include: surface temperature of heat sources inside the target power compartment, temperature and mass flow rate of intake air, and temperature and mass flow rate of exhaust smoke; meteorological parameters include: air temperature, wind speed, humidity, direct solar radiation, and diffuse solar radiation.

3. The method for predicting the target three-dimensional temperature field based on the generative diffusion model according to claim 1, characterized in that, The multimodal feature extraction module includes a convolutional feature encoding module, a feedforward neural network of a multilayer perceptron, a PointNet-based encoder, a temporal encoder, and a fusion module. The convolutional feature encoding module processes the target parameters, projecting the low-dimensional target parameters into a high-dimensional feature space through multilayer convolution operations and mapping with nonlinear activation functions to obtain the physical features of the parameters. The feedforward neural network of the multilayer perceptron extracts features from the node temperature values ​​to obtain the temperature features of the nodes. The PointNet-based encoder obtains the geometric features of the target 3D model based on the 3D coordinates of the target temperature field. The fusion module uses a feature concatenation strategy to overlay the physical features of the parameters, the node temperature features, the target geometric features, and the temporal step features to achieve effective fusion of multimodal features.

4. The method for predicting the target three-dimensional temperature field based on the generative diffusion model according to claim 3, characterized in that, The PointNet architecture encoder takes the three-dimensional coordinates of the target temperature field as input, performs a unified nonlinear high-dimensional mapping on each spatial point through a multilayer perceptron with shared weights, and introduces a max pooling function to globally aggregate the point-by-point high-dimensional features, preserving the key topological structure in the target's three-dimensional space to obtain the target's geometric features.

5. The method for predicting the target three-dimensional temperature field based on the generative diffusion model according to claim 1, characterized in that, The loss function used during the training of the generation diffusion module is: ； In the formula, For hyperparameters, Let be the diffusion loss function. For comparison loss functions.

6. The method for predicting the target three-dimensional temperature field based on the generative diffusion model according to claim 5, characterized in that, The diffusion loss function is as follows: ； In the formula, For time steps; This is real noise; Predict noise for the model; For node temperature characteristics; For parametric physical characteristics, For the first Noisy temperature field at time step; This represents the mathematical expectation, or average value.

7. The method for predicting the target three-dimensional temperature field based on the generative diffusion model according to claim 5, characterized in that, The contrastive loss function is as follows: ； In the formula, Cross-entropy loss; For target labels; It is The similarity matrix, This indicates transpose, and B is the batch size.

8. The method for predicting the target three-dimensional temperature field based on the generative diffusion model according to claim 7, characterized in that, The elements in the similarity matrix are as follows: , 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