Multi-fidelity data fusion prediction methods, devices, equipment, media, and products based on flexible feature alignment

Abstract

This invention discloses a multi-fidelity data fusion prediction method, apparatus, device, medium, and product based on flexible feature alignment, belonging to the field of data fusion. The method includes: generating corresponding low-fidelity and high-fidelity data based on the input conditions of the target to be predicted; inputting the low-fidelity and high-fidelity data into a trained multi-fidelity fusion prediction model, and outputting a high-fidelity prediction result that meets preset requirements for the target to be predicted; wherein the multi-fidelity fusion prediction model is trained on an initial fusion prediction model based on low-fidelity and high-fidelity sample data, and the initial fusion prediction model is trained on a basic prediction network architecture based on low-fidelity and high-fidelity sample data processed with flexible feature alignment. This invention can effectively improve the accuracy of prediction results.

Description

Technical Field

[0001] This invention relates to the field of data fusion technology, and in particular to a method, apparatus, equipment, medium and product for multi-fidelity data fusion prediction based on flexible feature alignment. Background Technology

[0002] In fields such as industrial simulation and scientific computing, there is typically a combination of low-fidelity data (LF, easily accessible and inexpensive but with low accuracy) and high-fidelity data (HF, difficult to acquire and expensive but with high accuracy). Multifidelity learning aims to combine the advantages of both to achieve near-high-fidelity predictions within a limited budget of high-cost HF data. For example, in computational fluid dynamics (CFD) simulations, coarse-mesh simulations can complete a calculation for a given operating condition in seconds, but with limited accuracy; while fine-mesh simulations may take hours or even days, but provide highly accurate results. The goal of multifidelity learning is to train a model using a large amount of LF data and a small amount of HF data, enabling it to output predictions with near-fine-mesh accuracy under new operating conditions.

[0003] In related technologies, existing methods usually require a precise pairing relationship between high-fidelity samples and low-fidelity samples. However, in real-world scenarios, high-fidelity samples are often far fewer than low-fidelity samples, making it difficult to establish such a precise pairing relationship. Furthermore, existing sample alignment methods are prone to pattern collapse or overfitting, ultimately resulting in extremely low prediction accuracy.

[0004] Therefore, there is an urgent need for multi-fidelity data fusion prediction methods, devices, equipment, media, and products based on flexible feature alignment to solve the above-mentioned technical problems. Summary of the Invention

[0005] This invention provides a method, apparatus, device, medium, and product for multi-fidelity data fusion prediction based on flexible feature alignment, which can effectively improve prediction accuracy. The technical solution is as follows: On the one hand, a multi-fidelity data fusion prediction method based on flexible feature alignment is provided, the method comprising: Generate corresponding low-fidelity and high-fidelity data based on the input conditions of the target to be predicted; Low-fidelity data and high-fidelity data are respectively input into a trained multi-fidelity fusion prediction model, and the output is a high-fidelity prediction result that meets the preset requirements for the target to be predicted; wherein, the multi-fidelity fusion prediction model is trained on an initial fusion prediction model based on low-fidelity sample data and high-fidelity sample data, and the initial fusion prediction model is trained on a basic prediction network architecture based on low-fidelity sample data and high-fidelity sample data that have undergone flexible feature alignment processing.

[0006] On the other hand, a multi-fidelity data fusion prediction device based on flexible feature alignment is provided, the device comprising: The generation module is used to generate corresponding low-fidelity and high-fidelity data based on the input conditions of the target to be predicted. The prediction module is used to input low-fidelity data and high-fidelity data into a pre-trained multi-fidelity fusion prediction model, and output a high-fidelity prediction result that meets the preset requirements for the target to be predicted. The multi-fidelity fusion prediction model is trained on an initial fusion prediction model based on low-fidelity sample data and high-fidelity sample data, and the initial fusion prediction model is trained on a basic prediction network architecture based on low-fidelity sample data and high-fidelity sample data that have undergone flexible feature alignment processing.

[0007] On the other hand, a computer device is provided, the computer device including a memory and a processor, the memory for storing computer programs, and the processor for executing the computer programs stored in the memory to implement the steps of the multi-fidelity data fusion prediction method based on flexible feature alignment described above.

[0008] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the steps of the above-described multi-fidelity data fusion prediction method based on flexible feature alignment.

[0009] On the other hand, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the multi-fidelity data fusion prediction method based on flexible feature alignment described above.

[0010] The technical solution provided by this invention can bring at least the following beneficial effects: This method constructs a network architecture consisting of a shared coding module, a low-fidelity baseline prediction network, and a high-fidelity residual correction network, providing a foundation for training a multi-fidelity fusion prediction model. Then, relying on massive and easily accessible low-fidelity data, the baseline model anchoring training and residual module initialization are completed in stages, consolidating the low-fidelity prediction accuracy and providing a stable feature manifold and reference benchmark for subsequent feature alignment. Next, surrogate features are generated in the latent space through soft attention weights, constructing reverse consistency constraints and fine-tuning the shared coding module in a targeted manner, which not only eliminates the hard pairing dependency in the input space but also avoids feature drift and overfitting under small sample conditions. Finally, the optimal parameters are converged through joint fine-tuning of the entire module using multi-component loss weighting. During inference, only a single forward propagation is needed to output high-fidelity results, ultimately achieving high-precision multi-fidelity fusion in unpaired, small-sample scenarios, significantly reducing the application cost of industrial high-fidelity simulation. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a multi-fidelity data fusion prediction method based on flexible feature alignment provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a fusion prediction network architecture provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of flexible feature alignment and reverse consistency provided in an embodiment of the present invention; Figure 4 This is a structural diagram of a multi-fidelity data fusion prediction device based on flexible feature alignment provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0014] As mentioned earlier, existing methods often struggle to establish accurate pairing relationships because high-fidelity samples are far fewer than low-fidelity samples, ultimately resulting in low accuracy of prediction results.

[0015] Based on this, the concept of the present invention is to achieve high-precision prediction of results when there is a shortage of high-fidelity data by flexibly aligning high-fidelity data and low-fidelity data.

[0016] The following describes the specific implementation of the above concept.

[0017] Please refer to Figure 1 This invention provides a multi-fidelity data fusion prediction method based on flexible feature alignment, the method comprising: Step 100: Generate corresponding low-fidelity data and high-fidelity data based on the input working conditions of the target to be predicted; Step 102: Input the low-fidelity data and high-fidelity data into the trained multi-fidelity fusion prediction model respectively, and output the high-fidelity accuracy prediction result of the target to be predicted, which meets the preset requirements; wherein, the multi-fidelity fusion prediction model is trained on the initial fusion prediction model based on the low-fidelity sample data and high-fidelity sample data, and the initial fusion prediction model is trained on the basic prediction network architecture based on the low-fidelity sample data and high-fidelity sample data after flexible feature alignment processing.

[0018] In this embodiment of the invention, a network architecture consisting of a shared encoding module, a low-fidelity baseline prediction network, and a high-fidelity residual correction network is constructed, providing a foundation for training a multi-fidelity fusion prediction model. Then, relying on massive amounts of readily available low-fidelity data, the baseline model anchoring training and residual module initialization are completed in stages, solidifying the low-fidelity prediction accuracy and providing a stable feature manifold and reference benchmark for subsequent feature alignment. Next, surrogate features are generated in the latent space using soft attention weights, constructing reverse consistency constraints and fine-tuning the shared encoding module in a targeted manner. This eliminates the hard pairing dependency in the input space and avoids feature drift and overfitting under small sample conditions. Finally, the optimal parameters are converged through joint fine-tuning of the entire module using multi-component loss weighting. During inference, only a single forward propagation is needed to output high-fidelity results, ultimately achieving high-precision multi-fidelity fusion in unpaired, small-sample scenarios, significantly reducing the application cost of industrial high-fidelity simulation.

[0019] The following description Figure 1 The execution method of each step is shown.

[0020] First, for step 100, corresponding low-fidelity data and high-fidelity data are generated based on the input conditions of the target to be predicted.

[0021] Low-fidelity data approximates real physical laws less closely, but it is fast to generate, consumes few resources, and can easily obtain large numbers of samples, numbering in the tens of thousands or more. It can be obtained through simulation calculations using sparse mesh generation, low-order numerical discretization schemes, and simplified physical models.

[0022] High-fidelity data closely approximates real physical laws and has high accuracy, but its generation cost is extremely high. Typically, only a few dozen small samples can be obtained. In contrast to low-fidelity data, simulations using encrypted grids, high-order numerical formats, and complete high-precision physical models are used to solve the same physical problem.

[0023] In this embodiment of the invention, the dimensions of high-fidelity data and low-fidelity data are determined according to the specific application scenario, such as geometric parameters and incoming flow conditions in airfoil aerodynamic characteristic prediction. Further details are omitted here.

[0024] Then, for step 102, the low-fidelity data and high-fidelity data are respectively input into the trained multi-fidelity fusion prediction model, and the high-fidelity accuracy prediction result of the target to be predicted meets the preset requirements is output.

[0025] Existing methods typically require precise pairings in the input space, meaning each low-fidelity sample must correspond to a specific high-fidelity sample. However, in real-world industrial scenarios, such one-to-one correspondences are often nonexistent or extremely difficult to establish. For example, in fluid simulations using different meshes, the node positions of coarse and fine meshes are completely different, making it impossible to establish a direct spatial correspondence. This renders traditional supervised learning methods completely unusable.

[0026] Therefore, this embodiment abandons the hard pairing dependency in the input space and achieves pairless fusion through flexible feature alignment in the latent space. To achieve this, this embodiment constructs as follows: Figure 2 The basic prediction network architecture shown includes three main parts: a shared coding module, a low-fidelity baseline prediction network, and a high-fidelity residual correction network.

[0027] Shared encoder module Used to process input data of different fidelity (including low-fidelity data) and high-fidelity data Mapped to a unified To obtain the latent variables in the latent feature space. ,Right now The encoder can employ fully connected neural networks, convolutional neural networks, or other neural network structures suitable for the input data type (e.g., using CNNs for image inputs and fully connected networks for numerical inputs).

[0028] Its core function is to abstract data of different fidelity and input formats into a unified feature representation, providing a unified feature foundation for subsequent flexible feature alignment and residual correction. The encoder's output dimension is a hyperparameter that needs to be optimized according to specific application scenarios (typically ranging from 32 to 256). For LF and HF data, the encoder shares the same network parameters, which ensures that data of both fidelity levels are mapped to the same latent space, providing a prerequisite for subsequent feature alignment.

[0029] Low-fidelity baseline prediction module With latent features As input, output low-fidelity prediction results. High-fidelity residual correction module With latent features Input: Output: Residual .

[0030] By training the basic prediction network architecture built from the above three parts in stages, a multi-fidelity fusion prediction model with significantly improved prediction accuracy is finally obtained. The training process includes the following steps: The initial fusion prediction model is obtained by training the basic prediction network architecture with low-fidelity sample data and high-fidelity sample data that have undergone flexible feature alignment. The initial fusion prediction model is trained using low-fidelity and high-fidelity sample data to obtain a multi-fidelity fusion prediction model.

[0031] In this embodiment of the invention, the initial fusion prediction module is specifically trained through the following two stages: Phase 1: Based on the low-fidelity sample data, the first parameter of the preset shared coding module, low-fidelity baseline prediction network, and high-fidelity residual correction network is iteratively updated in sequence.

[0032] In this stage, the low-fidelity sample data is first mapped and predicted sequentially based on the shared coding module to be updated and the low-fidelity baseline prediction network to obtain the first prediction result for each low-fidelity sample data.

[0033] Specifically, the low-fidelity sample data is first input into the shared encoding module, which maps the data to a unified encoding scheme. The dimensional latent feature space is obtained to obtain the th Potential characteristics of low-fidelity samples Then, a low-fidelity baseline prediction network is used to predict latent features. As input, output the first prediction result for the i-th low-fidelity sample data. : Furthermore, based on the first prediction result, a first low-fidelity supervision loss is calculated, and with the goal of minimizing the first low-fidelity supervision loss, the first parameters of the shared coding module and the low-fidelity baseline prediction network are iteratively updated.

[0034] Specifically, the low-fidelity supervised loss function designed in this embodiment is as follows: In the formula, represents the total number of low-fidelity sample data; This represents the true value of the i-th low-fidelity sample data; The first prediction result is substituted into the above function to calculate the loss of the current iteration round, and the loss is simultaneously propagated back to the shared coding module and the low-fidelity baseline prediction network through the backpropagation algorithm, and the network parameters of the two modules are updated synchronously.

[0035] The core function of joint training is to ensure that the latent features learned by the encoder are precisely the most effective feature representations for low-fidelity data prediction tasks, ultimately forming a task-oriented stable low-fidelity data feature manifold, rather than an unsupervised random feature space.

[0036] Training of the shared coding module and the low-fidelity baseline prediction network is complete when the following conditions are met: The prediction loss for low-fidelity data continues to decrease and stabilize on both the training and validation sets, and the accuracy on the validation set no longer improves. Meanwhile, when similar low-fidelity sample data are input, their corresponding latent features are also close in space, and the feature manifold is smooth and ordered.

[0037] Two core outcomes were achieved upon completion of the training: The low-fidelity baseline prediction network can accurately capture the input-output mapping pattern of low-fidelity data, accurately predict the results of low-fidelity data, and cover the overall trend and large-scale features of the data. At the same time, the shared encoding module forms a stable and adaptable low-fidelity data latent feature manifold, which is the core coordinate system for subsequent flexible feature alignment.

[0038] Furthermore, a baseline prediction value for the low-fidelity sample data is generated based on the updated shared coding module and the low-fidelity baseline prediction network, and a residual value for the low-fidelity sample data is generated based on the updated shared coding module and the high-fidelity residual correction network to be updated; the baseline prediction value and the residual value are summed to obtain a second prediction result with residuals.

[0039] Specifically, once the shared encoding module and the low-fidelity baseline prediction network have been updated, they have both achieved stable feature extraction capabilities and prediction accuracy for low-fidelity results. At this point, their parameters are frozen, and the high-fidelity residual correction network is trained and updated based on them.

[0040] First, a single low-fidelity sample data is input into the already trained shared encoding module, and a latent feature is output. Next, this latent feature is fed into both the already trained low-fidelity baseline prediction network and the untrained, updated high-fidelity residual correction network. The low-fidelity baseline prediction network then outputs a baseline prediction value. A residual value is output by the high-fidelity residual correction network. ; Next, the baseline predicted value and the residual value are summed to obtain the second prediction result with residuals for the i-th low-fidelity sample data. : The design philosophy of this residual learning paradigm is that low-fidelity networks can already capture the main trends and large-scale features of the data, while the differences between high-fidelity and low-fidelity predictions mainly lie in details and high-frequency components. Therefore, a small residual module can achieve the accuracy improvement from low-fidelity to high-fidelity predictions. This design significantly reduces the learning difficulty of the residual module, reduces the number of model parameters, and improves training stability and generalization ability.

[0041] Finally, the second low-fidelity supervision loss is calculated based on the second prediction result, and the first parameter of the high-fidelity residual correction network is iteratively updated with the objective function of minimizing the second low-fidelity supervision loss.

[0042] Specifically, the second prediction result with residuals is substituted back into the low-fidelity supervised loss function to calculate the loss for the current iteration, and the loss is then backpropagated to the high-fidelity residual correction network via the backpropagation algorithm to update the network parameters.

[0043] It is worth noting that, since the predicted values ​​output by the low-fidelity baseline prediction network are already very close to the true values ​​after the previous training, the process of optimizing this loss is essentially to force the output of the high-fidelity residual correction network on the latent features of the low-fidelity data to approach 0, so as to avoid it interfering with the existing accurate baseline predictions.

[0044] The training of a high-fidelity residual correction network is complete when the following two conditions are met: The prediction accuracy of the low-fidelity baseline prediction network did not decrease significantly compared with the baseline accuracy in the first stage, indicating that the high-fidelity residual correction network did not destroy the baseline prediction capability; the output amplitude of the high-fidelity residual correction network remained stable in a very small range close to 0.

[0045] After training, the high-fidelity residual correction network has an initial state of "outputting zero-mean small residuals in the low-fidelity domain". This not only ensures the stability of the low-fidelity prediction benchmark, but also provides a reasonable parameter starting point for subsequent learning of the true residuals between the high-fidelity prediction results and the low-fidelity prediction results, avoiding the training oscillation and overfitting risks caused by random initialization.

[0046] The training and update process in Phase Two is as follows: Figure 3 As shown, it includes the following steps: In the latent feature space, the latent features of low-fidelity sample data and high-fidelity sample data are flexibly aligned to obtain the latent surrogate features of each high-fidelity sample data on the manifold of low-fidelity sample data. Reverse consistency calculation is performed on the potential original features and potential surrogate features of the high-fidelity sample data, and the second parameter of the optimized shared encoder module is iteratively updated based on the calculation results.

[0047] After completing the training process in Phase 1, the shared coding module, low-fidelity baseline prediction network, and high-fidelity residual correction network in the basic prediction network architecture have all completed their initial parameter updates. At this point, although the accuracy of the prediction results has been improved to some extent, the key technical problem has not yet been solved—that is, the low-fidelity samples and high-fidelity samples cannot achieve a one-to-one correspondence in the input space.

[0048] To address this fundamental issue affecting prediction accuracy, this embodiment proposes a method to transfer the alignment operation from the input space to the latent feature space of the encoder output, based on the initial update of the network architecture. By calculating the soft attention weights of high-fidelity samples to all low-fidelity samples, a flexible, manifold distance-based correspondence is established, thereby completely eliminating the dependence on hard-paired data.

[0049] In this embodiment of the invention, the calculation process of potential agent features is as follows: The low-fidelity sample data and high-fidelity sample data are respectively input into the shared encoding module that has completed the first parameter iterative update for mapping processing, and the latent features of the low-fidelity sample and the latent features of the high-fidelity sample are output in sequence. Specifically, low-fidelity sample data and high-fidelity sample data are input into the shared encoding module, and the latent features of the j-th high-fidelity sample are output sequentially. Latent features of the i-th low-fidelity sample .

[0050] Furthermore, the squared Euclidean distance between the latent features of a single high-fidelity sample and the latent features of all low-fidelity samples is calculated: In the formula, Latent features of the j-th high-fidelity sample Latent features of the i-th low-fidelity sample The square of the Euclidean distance.

[0051] The approximation of the two in the latent space is quantified by calculating the squared Euclidean distance between them, providing a basis for subsequent soft weight allocation.

[0052] It is worth noting that the Euclidean distance squared is chosen in this embodiment because of its ease of calculation and differentiability, which facilitates subsequent gradient backpropagation and model optimization.

[0053] Furthermore, the calculation results are weighted according to preset temperature parameters to obtain the weight coefficient of the latent features of each low-fidelity sample relative to the latent features of the same high-fidelity sample. : In the formula, Let be the correlation between the latent features of the i-th low-fidelity sample and the latent features of the j-th high-fidelity sample; For temperature parameters; This represents the total number of low-fidelity samples.

[0054] Specifically, this embodiment introduces a temperature parameter. The soft attention weights (i.e., approximations of the flexible feature alignment matrix) of HF samples to LF samples are calculated using the Softmax function, thereby achieving flexible association between HF samples and multiple LF samples.

[0055] Temperature parameters It controls the sharpness of the weight distribution: when When smaller (e.g.) The weights are concentrated on the most recent few low-fidelity samples, which is close to a hard allocation; when When it is large (e.g.) The weights are relatively even, and all low-fidelity samples participate in the aggregation. This soft allocation mechanism is the core of this invention for achieving unpaired fusion. It avoids a rigid one-to-one correspondence and allows each high-fidelity sample to establish a flexible correspondence with multiple low-fidelity samples, adapting to unpaired industrial scenarios.

[0056] Furthermore, based on the latent features of each low-fidelity sample and its corresponding weight coefficient, the latent proxy features of the high-fidelity sample are calculated: In the formula, Let be the potential proxy features of the j-th high-fidelity sample.

[0057] Specifically, utilizing soft weights We perform a weighted summation of the latent variables for all low-fidelity samples to generate "latent surrogate features" of high-fidelity samples on the low-fidelity manifold. The proxy feature can be understood as the "projection" of the high-fidelity sample onto the low-fidelity manifold. It is a key intermediate quantity for subsequent calculation of reverse consistency loss. It integrates the feature information of multiple low-fidelity samples, and each weight is determined by the potential distance between the high-fidelity sample and each low-fidelity sample. It can accurately reflect the corresponding features of the high-fidelity sample on the low-fidelity manifold, providing support for establishing the structural relationship between high-fidelity and low-fidelity samples.

[0058] For example, suppose we have 3 high-fidelity sample data and 10 low-fidelity sample data. After the above processing, we can obtain the potential proxy features of each high-fidelity sample data with respect to the 10 low-fidelity sample data. That is, the 3 high-fidelity sample data have 3 potential features and 3 potential proxy features. This flexible alignment can be adapted to unpaired industrial scenarios.

[0059] It is worth noting that, in order to ensure that the features of high-fidelity data can be interpreted by the low-fidelity prediction network and improve the robustness of the model under small sample conditions, this embodiment proposes a reverse consistency mechanism, that is, to force the output of the high-fidelity sample data after passing through the low-fidelity benchmark prediction network to be consistent with the output of the surrogate features after passing through the low-fidelity benchmark prediction network.

[0060] This mechanism enables iterative updates of the second parameter of the shared encoder module, ensuring that high-fidelity data maintains structural consistency with the low-fidelity manifold in the latent space, avoiding feature drift in small sample cases, and eliminating the MMD or GAN distribution alignment methods that are prone to failure in small sample cases, thereby improving model stability.

[0061] In this embodiment of the invention, the iterative update process of the second parameter is as follows: The latent features and latent surrogate features of the high-fidelity sample data are respectively input into the low-fidelity baseline prediction network after the first parameter iteration update, and the corresponding prediction results are obtained in sequence. The reverse consistency loss of the latent features and latent surrogate features is calculated based on the prediction results, and the second parameter iteration update of the shared coding module is performed with the goal of minimizing the reverse consistency loss value, so that the latent feature structure of the high-fidelity sample data tends to be consistent with the manifold structure of the low-fidelity sample data.

[0062] Specifically, the latent features of the j-th high-fidelity sample are calculated directly through the output of the low-fidelity baseline prediction network. The corresponding potential agent features are then processed by the low-fidelity baseline prediction network. The mean square error between them is denoted as This refers to the reverse consistency loss. Its physical meaning is: it forces the projection of high-fidelity samples onto the low-fidelity manifold to be interpretable by the low-fidelity model, thereby achieving implicit structural anchoring in the case of no pairing and avoiding feature drift under small sample conditions.

[0063] The reverse consistency loss function is calculated using the following formula: In the formula, This represents the total number of high-fidelity sample data.

[0064] To facilitate averaging, the scale of the loss value is ensured to be independent of the sample size. The design principle of this loss function is: if HF samples maintain structural consistency with the LF manifold in the latent space, then after averaging... The output should be derived from its surrogate features on the LF manifold. The outputs are close, and by minimizing this loss, the structures of HF and LF are aligned.

[0065] After calculating the reverse consistency loss for the current iteration, the loss is propagated back to the shared coding module via the backpropagation algorithm for parameter fine-tuning, i.e., the second parameter is iteratively updated.

[0066] The encoder alignment fine-tuning is complete when the following conditions are met: The reverse consistency loss continued to decrease and stabilized, no longer exhibiting significant fluctuations; The latent features of high-fidelity sample data are uniformly distributed in space near the manifold of low-fidelity sample data, with no obvious outliers or drift phenomena. The output of the low-fidelity baseline prediction network for the latent features of high-fidelity sample data deviates from the output of the corresponding surrogate features by a very small margin.

[0067] In summary, after completing the training and update of the above two stages, an initial fusion prediction model can be established based on the updated parameters of the shared encoding module, the low-fidelity baseline prediction network, and the high-fidelity residual correction network.

[0068] In this embodiment of the invention, the multi-fidelity fusion prediction model is trained through the following steps: Based on the training loss of the initial fusion prediction model, a total loss function is established, which includes low-fidelity supervision loss, high-fidelity supervision loss, reverse consistency loss, and latent space representation smoothing loss.

[0069] Specifically, as mentioned above, low-fidelity supervision loss function and reverse consistency loss function have been discussed. To balance the low-fidelity baseline prediction accuracy, high-fidelity prediction accuracy, latent spatial structure alignment, and feature smoothness of the final prediction model, this embodiment also designs a high-fidelity supervision loss function, calculated using mean squared error (MSE), and optimizes the model using HF samples to ensure HF prediction accuracy. The formula is as follows: in, For the first The true value of a high-fidelity sample data. The corresponding predicted value; The latent space representation smoothing loss function, calculated using mean squared error, is used to improve the local smoothness of latent features and avoid excessive feature fluctuations. The formula is as follows: in, For the first The local neighborhood mean of the latent features of each sample.

[0070] By assigning weights to the four loss functions using preset proportions, the total loss function of the model is obtained: to , , These are hyperparameter weights, which need to be tuned according to the specific application scenario. Typical initial values ​​are set as follows: , , , It can be fine-tuned based on the model training results.

[0071] Finally, based on the low-fidelity sample data and the high-fidelity sample data, the model parameters of the initial fusion prediction model are iteratively updated with the goal of minimizing the total loss function value. The update ends when the preset convergence condition is met, and the multi-fidelity fusion prediction model is output.

[0072] Specifically, the parameters of all network modules in the initial fusion prediction model are unfrozen, and all low-fidelity and high-fidelity sample data are used simultaneously for end-to-end joint training with the total loss function as the optimization objective. The encoding module, the baseline prediction network, and the residual correction network are then adapted to each other to further improve the prediction accuracy of the final result while maintaining structural alignment. Once training is complete, the final multi-fidelity fusion prediction model is obtained.

[0073] Please refer to Figure 4 This invention provides a multi-fidelity data fusion prediction device based on flexible feature alignment, the device comprising: The generation module 400 is used to generate corresponding low-fidelity data and high-fidelity data based on the input conditions of the target to be predicted. The prediction module 402 is used to input low-fidelity data and high-fidelity data into the trained multi-fidelity fusion prediction model, and output a high-fidelity prediction result that meets the preset requirements for the target to be predicted; wherein, the multi-fidelity fusion prediction model is trained on the initial fusion prediction model based on the low-fidelity sample data and the high-fidelity sample data, and the initial fusion prediction model is trained on the basic prediction network architecture based on the low-fidelity sample data and the high-fidelity sample data after flexible feature alignment processing.

[0074] In this embodiment of the invention, an initial fusion prediction model is obtained by training a basic prediction network architecture based on low-fidelity sample data and high-fidelity sample data that have undergone flexible feature alignment processing, including: The first parameter of the preset shared coding module, low-fidelity baseline prediction network and high-fidelity residual correction network is updated sequentially based on the low-fidelity sample data. In the latent feature space, the latent features of low-fidelity sample data and high-fidelity sample data are flexibly aligned to obtain the latent surrogate features of each high-fidelity sample data on the manifold of low-fidelity sample data. Reverse consistency calculation is performed on the potential original features and potential surrogate features of high-fidelity sample data, and the second parameter of the optimized shared encoder module is iteratively updated based on the calculation results; An initial fusion prediction model is established based on the updated shared coding module, the low-fidelity baseline prediction network, and the high-fidelity residual correction network.

[0075] In this embodiment of the invention, the step of sequentially updating the first parameter of the preset shared encoding module, the low-fidelity baseline prediction network, and the high-fidelity residual correction network based on low-fidelity sample data includes: Based on the shared coding module to be updated and the low-fidelity baseline prediction network, the low-fidelity sample data is mapped and predicted sequentially to obtain the first prediction result for each low-fidelity sample data. The first low-fidelity supervision loss is calculated based on the first prediction result, and the first parameter is iteratively updated for the shared coding module and the low-fidelity benchmark prediction network with the goal of minimizing the first low-fidelity supervision loss. The baseline prediction values ​​of the low-fidelity sample data are generated based on the updated shared coding module and the low-fidelity baseline prediction network, and the residual values ​​of the low-fidelity sample data are generated based on the updated shared coding module and the high-fidelity residual correction network to be updated. The baseline forecast and the residual values ​​are summed to obtain the second forecast result with the residuals. The second low-fidelity supervision loss is calculated based on the second prediction result, and the first parameter of the high-fidelity residual correction network is iteratively updated with the objective function of minimizing the second low-fidelity supervision loss.

[0076] In this embodiment of the invention, a flexible alignment process is performed on the latent features of low-fidelity sample data and high-fidelity sample data in the latent feature space to obtain the latent surrogate features of each high-fidelity sample data on the manifold of the low-fidelity sample data, including: The low-fidelity sample data and high-fidelity sample data are respectively input into the shared encoding module that has completed the first parameter iterative update for mapping processing, and the latent features of the low-fidelity sample and the latent features of the high-fidelity sample are output in sequence. Calculate the squared Euclidean distance between the latent features of a single high-fidelity sample and the latent features of all low-fidelity samples: In the formula, Latent features of the j-th high-fidelity sample Latent features of the i-th low-fidelity sample The square of the Euclidean distance; The calculation results are weighted according to preset temperature parameters to obtain the weight coefficient of the latent features of each low-fidelity sample relative to the latent features of the same high-fidelity sample. : In the formula, Let be the correlation between the latent features of the i-th low-fidelity sample and the latent features of the j-th high-fidelity sample; For temperature parameters; The total number of low-fidelity samples; Based on the latent features of each low-fidelity sample and its corresponding weight coefficient, the latent proxy features of the high-fidelity sample are calculated: In the formula, Let be the potential proxy features of the j-th high-fidelity sample.

[0077] In this embodiment of the invention, reverse consistency calculation is performed on the latent features and latent surrogate features of high-fidelity sample data, and the second parameter of the optimized shared encoder module is iteratively updated based on the calculation results, including: The latent features and latent proxy features of the high-fidelity sample data are respectively input into the low-fidelity baseline prediction network that has completed the first parameter iteration update, and the corresponding prediction results are obtained in sequence. The reverse consistency loss of latent features and latent proxy features is calculated based on the prediction results. The second parameter of the shared coding module is iteratively updated with the goal of minimizing the reverse consistency loss value, so that the latent feature structure of high-fidelity sample data tends to be consistent with the manifold structure of low-fidelity sample data.

[0078] In this embodiment of the invention, an initial fusion prediction model is trained based on low-fidelity sample data and high-fidelity sample data to obtain a multi-fidelity fusion prediction model, including: Based on the training loss of the initial fusion prediction model, a total loss function is established that includes low-fidelity supervision loss, high-fidelity supervision loss, reverse consistency loss, and latent space representation smoothing loss. Based on the low-fidelity sample data and the high-fidelity sample data, the model parameters of the initial fusion prediction model are iteratively updated with the goal of minimizing the total loss function value. The update ends when the preset convergence condition is met, and the multi-fidelity fusion prediction model is output.

[0079] It should be noted that the multi-fidelity data fusion prediction device based on flexible feature alignment provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the multi-fidelity data fusion prediction device based on flexible feature alignment provided in the above embodiments and the multi-fidelity data fusion prediction method based on flexible feature alignment belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0080] Embodiments of this application also provide a computer device including a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the multi-fidelity data fusion prediction method based on flexible feature alignment provided in the above-described method embodiments.

[0081] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the multi-fidelity data fusion prediction method based on flexible feature alignment provided in the above-described method embodiments.

[0082] Embodiments of this application also provide a computer program product, which includes a computer program. A processor of a computer device reads the computer program from a computer-readable storage medium and executes the computer program, causing the computer device to perform any of the multi-fidelity data fusion prediction methods based on flexible feature alignment described in the above embodiments.

[0083] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0084] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0085] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A multi-fidelity data fusion prediction method based on flexible feature alignment, characterized in that, The method includes: Generate corresponding low-fidelity and high-fidelity data based on the input conditions of the target to be predicted; Low-fidelity data and high-fidelity data are respectively input into a trained multi-fidelity fusion prediction model, and the output is a high-fidelity prediction result that meets the preset requirements for the target to be predicted; wherein, the multi-fidelity fusion prediction model is trained on an initial fusion prediction model based on low-fidelity sample data and high-fidelity sample data, and the initial fusion prediction model is trained on a basic prediction network architecture based on low-fidelity sample data and high-fidelity sample data that have undergone flexible feature alignment processing.

2. The method as described in claim 1, characterized in that, The basic prediction network architecture is trained using low-fidelity and high-fidelity sample data that have undergone flexible feature alignment to obtain an initial fusion prediction model, including: The first parameter of the preset shared coding module, low-fidelity baseline prediction network and high-fidelity residual correction network is updated sequentially based on the low-fidelity sample data. In the latent feature space, the latent features of low-fidelity sample data and high-fidelity sample data are flexibly aligned to obtain the latent surrogate features of each high-fidelity sample data on the manifold of low-fidelity sample data. Reverse consistency calculation is performed on the potential original features and potential surrogate features of high-fidelity sample data, and the second parameter of the optimized shared encoder module is iteratively updated based on the calculation results; An initial fusion prediction model is established based on the updated shared coding module, the low-fidelity baseline prediction network, and the high-fidelity residual correction network.

3. The method as described in claim 2, characterized in that, The step of iteratively updating the first parameters of the preset shared encoding module, low-fidelity baseline prediction network, and high-fidelity residual correction network sequentially based on low-fidelity sample data includes: Based on the shared coding module to be updated and the low-fidelity baseline prediction network, the low-fidelity sample data is mapped and predicted sequentially to obtain the first prediction result for each low-fidelity sample data. The first low-fidelity supervision loss is calculated based on the first prediction result, and the first parameter is iteratively updated for the shared coding module and the low-fidelity benchmark prediction network with the goal of minimizing the first low-fidelity supervision loss. The baseline prediction values ​​of the low-fidelity sample data are generated based on the updated shared coding module and the low-fidelity baseline prediction network, and the residual values ​​of the low-fidelity sample data are generated based on the updated shared coding module and the high-fidelity residual correction network to be updated. The baseline forecast and the residual values ​​are summed to obtain the second forecast result with the residuals. The second low-fidelity supervision loss is calculated based on the second prediction result, and the first parameter of the high-fidelity residual correction network is iteratively updated with the objective function of minimizing the second low-fidelity supervision loss.

4. The method as described in claim 2, characterized in that, In the latent feature space, a flexible alignment process is performed on the latent features of the low-fidelity sample data and the high-fidelity sample data to obtain the latent surrogate features of each high-fidelity sample data on the low-fidelity sample data manifold, including: The low-fidelity sample data and high-fidelity sample data are respectively input into the shared encoding module that has completed the first parameter iterative update for mapping processing, and the latent features of the low-fidelity sample and the latent features of the high-fidelity sample are output in sequence. Calculate the squared Euclidean distance between the latent features of a single high-fidelity sample and the latent features of all low-fidelity samples: In the formula, Latent features of the j-th high-fidelity sample Latent features of the i-th low-fidelity sample The square of the Euclidean distance; The calculation results are weighted according to preset temperature parameters to obtain the weight coefficient of the latent features of each low-fidelity sample relative to the latent features of the same high-fidelity sample. : In the formula, Let be the correlation between the latent features of the i-th low-fidelity sample and the latent features of the j-th high-fidelity sample; For temperature parameters; The total number of low-fidelity samples; Based on the latent features of each low-fidelity sample and its corresponding weight coefficient, the latent proxy features of the high-fidelity sample are calculated: In the formula, Let be the potential proxy features of the j-th high-fidelity sample.

5. The method as described in claim 4, characterized in that, Reverse consistency calculation is performed on the latent features and latent surrogate features of the high-fidelity sample data, and the second parameter of the optimized shared encoder module is iteratively updated based on the calculation results, including: The latent features and latent proxy features of the high-fidelity sample data are respectively input into the low-fidelity baseline prediction network that has completed the first parameter iteration update, and the corresponding prediction results are obtained in sequence. The reverse consistency loss of latent features and latent surrogate features is calculated based on the prediction results. The second parameter of the shared coding module is iteratively updated with the goal of minimizing the reverse consistency loss value, so that the latent feature structure of high-fidelity sample data tends to be consistent with the manifold structure of low-fidelity sample data.

6. The method as described in claim 2, characterized in that, The initial fusion prediction model is trained using low-fidelity and high-fidelity sample data to obtain a multi-fidelity fusion prediction model, including: Based on the training loss of the initial fusion prediction model, a total loss function is established that includes low-fidelity supervision loss, high-fidelity supervision loss, reverse consistency loss, and latent space representation smoothing loss. The model parameters of the initial fusion prediction model are iteratively updated based on the low-fidelity sample data and the high-fidelity sample data with the goal of minimizing the total loss function value. The update ends when the preset convergence condition is met, and the multi-fidelity fusion prediction model is output.

7. A multi-fidelity data fusion prediction device based on flexible feature alignment, characterized in that, The device includes: The generation module is used to generate corresponding low-fidelity and high-fidelity data based on the input conditions of the target to be predicted. The prediction module is used to input low-fidelity data and high-fidelity data into a pre-trained multi-fidelity fusion prediction model, and output a high-fidelity prediction result that meets the preset requirements for the target to be predicted. The multi-fidelity fusion prediction model is trained on an initial fusion prediction model based on low-fidelity sample data and high-fidelity sample data, and the initial fusion prediction model is trained on a basic prediction network architecture based on low-fidelity sample data and high-fidelity sample data that have undergone flexible feature alignment processing.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images