Physical field data fusion method

Abstract

The invention provides a physical field data fusion method, which belongs to the technical field of multi-source data fusion, and comprises the following steps: multiplying low-fidelity physical field data by a sensor mask matrix to obtain low-fidelity sensor data; constructing a multi-order difference format based on a finite difference method and a Taylor expansion theory, and processing the low-fidelity physical field data to obtain low-fidelity multi-order partial differential physical field data; training the pre-trained data fusion model through the low-fidelity multi-order partial differential physical field data and the low-fidelity sensor data to obtain a data fusion model; the data fusion model is finely adjusted and optimized by updating sensor data and updating multi-order partial differential physical field data, and an optimized data fusion model is obtained and used for fusion processing of multi-source heterogeneous multi-physical field data; the optimized data fusion model obtained by the method can perform feature extraction and fusion on the multi-source heterogeneous multi-physics field data to obtain high-fidelity data with consistent representation and high confidence.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-source data fusion, and in particular to a physical field data fusion method. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technologies such as machine learning, deep learning methods have demonstrated significant advantages in a variety of tasks, such as physical field reconstruction, diagnosis and health management, uncertainty quantification, and digital twins. The success of deep learning methods depends on large amounts of accurate data to construct mappings that reflect the true relationship between inputs and outputs.

[0003] However, multi-source, heterogeneous, multiphysics data is common in practical applications. It comes from diverse sources, such as simulations and experiments, and involves different physics fields, such as force, heat, and flow. Multi-source, heterogeneous, multiphysics data also has different types, attributes, structures, and forms. These differences lead to data diversity and inconsistencies in accuracy and resolution, posing significant challenges to deep learning-based proxy models. Therefore, it is crucial to effectively fuse multi-source, heterogeneous, multiphysics data to form a highly reliable multiphysics dataset with consistent representation across different dimensions.

[0004] Currently, there are proposals to directly simulate multi-physics data using refined numerical computational models, discretizing the computational domain and using numerical methods to calculate multi-physics. However, it is difficult to fully align the computational simulation model settings with physical reality. The use of refined computational grids imposes a significant computational burden, making it impossible to obtain high-precision physical field data in real time and reducing the efficiency of obtaining high-precision physical field information.

[0005] Traditionally, real physical field data is obtained through real-world physical experiments. High-precision sensors are used to collect real-time physical field information, but this consumes significant manpower and resources, and the data obtained cannot effectively cover the entire operating space. Furthermore, differences between actual operating conditions and physical experiments, such as the significant differences between spacecraft in-orbit operation and ground-based simulated physical experiments, can lead to shifts in data distribution.

[0006] Data fusion methods based on traditional theories such as Bayesian theory, Kalman filtering, and Dempster-Shafer (DS) have great limitations when dealing with multi-source heterogeneous multi-physics field data fusion problems with multiple scales, multiple precisions, multiple resolutions, strong nonlinearity, and strong coupling. For example, Bayesian-based fusion methods require prior and posterior information, as well as independence between variables, and it is very difficult to decouple multiple physical fields. Kalman filtering is only applicable to linear systems and cannot solve nonlinear, non-Gaussian multi-source heterogeneous multi-physics field data fusion problems. The Dempster-Shafer (DS) theoretical method can fuse data from different sources, but when the inconsistency effect of the data is very prominent, it leads to weak robustness and may cause information loss.

[0007] Therefore, a physical field data fusion method is proposed. Summary of the Invention

[0008] In order to solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a physical field data fusion method to solve the problem that traditional technology cannot effectively fuse multi-source heterogeneous multi-physical field data.

[0009] The technical solutions of the present invention are as follows:

[0010] A physical field data fusion method is provided, including:

[0011] Simulate the physical field based on numerical calculation methods to obtain low-fidelity physical field data;

[0012] Obtain the placement of sensors in the area to be detected and construct a sensor mask matrix;

[0013] multiplying the low-fidelity physical field data by the sensor mask matrix to obtain low-fidelity sensor data;

[0014] Based on the finite difference method and Taylor expansion theory, a multi-order difference format is constructed;

[0015] Processing the low-fidelity physical field data using the multi-order difference format to obtain low-fidelity multi-order partial differential physical field data;

[0016] Build a pre-trained data fusion model;

[0017] Training the pre-trained data fusion model based on the low-fidelity multi-order partial differential physical field data and the low-fidelity sensor data to obtain a data fusion model;

[0018] Data is collected through sensors deployed in the area to be detected to obtain high-fidelity physical field data;

[0019] multiplying the high-fidelity physical field data by the sensor mask matrix to obtain high-fidelity sensor data;

[0020] fusing the low-fidelity sensor data and the high-fidelity sensor data through a data fusion model to obtain preliminary fused physical field data;

[0021] Processing the preliminary fused physical field data using the multi-order difference format to obtain fused partial differential physical field data;

[0022] Updating and optimizing the preliminary fused physical field data based on an iterative update coefficient to obtain updated sensor data;

[0023] Acquire updated multi-order partial differential physical field data according to the low-fidelity multi-order partial differential physical field data and the fused partial differential physical field data;

[0024] Fine-tuning the data fusion model according to the updated sensor data and the updated multi-order partial differential physical field data to obtain an optimized data fusion model;

[0025] The optimized data fusion model is used for the fusion processing of multi-source heterogeneous and multi-physics field data.

[0026] Preferably, in the above-mentioned physical field data fusion method, the step of simulating the physical field based on a numerical calculation method to obtain low-fidelity physical field data includes:

[0027] The physical field is simulated by any numerical calculation method such as finite difference method, finite element method and finite volume method to obtain low-fidelity physical field data.

[0028] Preferably, in the above-mentioned physical field data fusion method, the step of obtaining the placement positions of sensors in the area to be detected and constructing a sensor mask matrix comprises:

[0029] According to the size of the physical field in the area to be detected, an initial mask matrix is ​​set;

[0030] The positions where sensors are arranged in the area to be detected are obtained, and their corresponding values ​​in the initial mask matrix are set to "1". The corresponding values ​​of the positions where sensors are not arranged in the area to be detected are set to "0" in the corresponding initial mask matrix to construct a sensor mask matrix.

[0031] Preferably, in the above-mentioned physical field data fusion method, the step of updating and optimizing the preliminary fused physical field data based on the iterative update coefficient to obtain updated sensor data comprises:

[0032] S in =coe·S LF+(1-coe)·S HF

[0033] Among them, S in is to update the sensor data, coe is the iterative update coefficient, S LF For low-fidelity sensor data, S HF For high-fidelity sensor data;

[0034] coe=e -β·epoch

[0035] Among them, β is a hyperparameter and epoch is the number of iterations;

[0036] By continuously optimizing and adjusting the iterative update coefficient, the updated sensor data is made to approach the high-fidelity sensor data.

[0037] Preferably, in the above-mentioned physical field data fusion method, the step of obtaining updated multi-order partial differential physical field data based on the low-fidelity multi-order partial differential physical field data and the fused partial differential physical field data comprises:

[0038] G in =coe·G LF +(1-coe)·G HF

[0039] Among them, G in To update the multi-order partial differential physics data, G LF is low-fidelity multi-order partial differential physical field data, G HF To integrate partial differential physics data;

[0040] By continuously optimizing and adjusting the iterative update coefficient, the updated multi-order partial differential physical field data is made close to the fused partial differential physical field data.

[0041] The main advantages of the technical solution of the present invention are as follows:

[0042] The physical field data fusion method of the present invention greatly reduces the training difficulty and improves the model training efficiency through pre-training, fine-tuning and updating in the process of obtaining the optimized data fusion model, so that the obtained optimized data fusion model can extract and fuse features of multi-source heterogeneous multi-physical field data, and obtain high-fidelity data with consistent representation and high confidence, solving the problem that the traditional technology uses sophisticated numerical calculation models or conducts real physical experiments, which is difficult to conform to physical reality and consumes computing and human resources; in addition, physical partial differential equations are implicitly introduced in the training process, making full use of the rich physical information contained in the physical field data to achieve higher fusion accuracy of the model. The above method is applicable to data fusion scenarios under multi-scale, multi-precision, multi-resolution, strong nonlinearity and strong coupling, so that the multi-source physical field data processed by the model have consistent representation and high reliability in different dimensions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings described herein are used to provide a further understanding of the embodiments of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0044] Figure 1 This is a flow chart of a physical field data fusion method provided by the present invention. DETAILED DESCRIPTION

[0045] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] The technical solutions provided by the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] Example 1:

[0048] The embodiment of the present invention provides a physical field data fusion method, referring to Figure 1 ,include:

[0049] Simulate the physical field based on numerical calculation methods to obtain low-fidelity physical field data;

[0050] Obtain the placement of sensors in the area to be detected and construct a sensor mask matrix;

[0051] Multiplying the low-fidelity physical field data with the sensor mask matrix to obtain low-fidelity sensor data;

[0052] Based on the finite difference method and Taylor expansion theory, a multi-order difference format is constructed;

[0053] Process low-fidelity physical field data through multi-order difference format to obtain low-fidelity multi-order partial differential physical field data;

[0054] Build a pre-trained data fusion model;

[0055] The pre-trained data fusion model is trained based on low-fidelity multi-order partial differential physical field data and low-fidelity sensor data to obtain a data fusion model;

[0056] Data is collected through sensors deployed in the area to be detected to obtain high-fidelity physical field data;

[0057] Multiplying the high-fidelity physical field data with the sensor mask matrix to obtain high-fidelity sensor data;

[0058] The low-fidelity sensor data and high-fidelity sensor data are fused through the data fusion model to obtain preliminary fused physical field data;

[0059] Processing the preliminary fusion physical field data through a multi-order difference format to obtain fusion partial differential physical field data;

[0060] Update and optimize the preliminary fused physical field data based on the iterative update coefficient to obtain updated sensor data;

[0061] Obtain updated multi-order partial differential physics data according to the low-fidelity multi-order partial differential physics data and the fused partial differential physics data;

[0062] Fine-tune the data fusion model according to the updated sensor data and the updated multi-order partial differential physical field data to obtain an optimized data fusion model;

[0063] The optimized data fusion model is used for the fusion processing of multi-source heterogeneous and multi-physics field data.

[0064] In the above embodiments, the physical field is simulated based on a numerical calculation method to obtain low-fidelity physical field data; the placement positions of sensors in the area to be detected are obtained, a sensor mask matrix is ​​constructed, and the low-fidelity physical field data is multiplied by the sensor mask matrix to obtain low-fidelity sensor data; based on the finite difference method and Taylor expansion theory, a multi-order difference format is constructed, and the low-fidelity physical field data is processed through the multi-order difference format to obtain low-fidelity multi-order partial differential physical field data; the constructed pre-trained data fusion model is trained based on the low-fidelity multi-order partial differential physical field data and the low-fidelity sensor data to obtain a data fusion model;

[0065] Data is collected through sensors deployed in the area to be detected to obtain high-fidelity physical field data, and the high-fidelity physical field data is multiplied with the sensor mask matrix to obtain high-fidelity sensor data; low-fidelity sensor data and high-fidelity sensor data are fused through a data fusion model to obtain preliminary fused physical field data; preliminary fused physical field data are processed through a multi-order difference format to obtain fused partial differential physical field data;

[0066] Based on the iterative update coefficient, the preliminary fused physical field data is updated and optimized to obtain updated sensor data; based on the low-fidelity multi-order partial differential physical field data and the fused partial differential physical field data, updated multi-order partial differential physical field data are obtained; based on the updated sensor data and the updated multi-order partial differential physical field data, the data fusion model is fine-tuned to obtain an optimized data fusion model;

[0067] The optimized data fusion model is used for the fusion processing of multi-source heterogeneous and multi-physics field data.

[0068] In the above embodiments, the pre-trained data fusion model is trained by using low-fidelity sensor data as low-fidelity low-order Taylor modes and low-fidelity multi-order partial differential physical field data as low-fidelity high-order Taylor modes.

[0069] In the above embodiments, the pre-trained data fusion model is constructed based on any deep neural network such as a fully connected network, a convolutional neural network, a Fourier neural operator, or a combination of multiple neural networks.

[0070] In the above embodiments, the low-fidelity physical field data obtained by numerical calculation simulation is continuous data, and the high-fidelity physical field data obtained by data acquisition by sensors is discrete data. Continuous data contains rich distribution information, but such data is often limited by factors such as test conditions, test equipment reliability, and calculation simulation accuracy, and cannot achieve high accuracy. Discrete data can usually achieve very high single-point accuracy, but this column of data often does not contain continuous distribution information, so it is impossible to obtain a high spatial resolution. By constructing an optimized data fusion model, continuous data and discrete data are effectively fused, and their advantages are complemented to improve the accuracy and spatiotemporal resolution of physical field data.

[0071] The beneficial effects of the above technology are: the pre-trained data fusion model constructed is trained by low-fidelity multi-order partial differential physical field data and low-fidelity sensor data to obtain a data fusion model; the data fusion model is fine-tuned and optimized by the obtained updated sensor data and updated multi-order partial differential physical field data to obtain an optimized data fusion model for fusion processing of multi-source heterogeneous multi-physical field data; compared with traditional technologies, the physical field data fusion method greatly reduces the training difficulty and improves the model training efficiency through pre-training, fine-tuning and updating in the process of obtaining the optimized data fusion model, so that the obtained optimized data fusion model can be used for multi-source heterogeneous multi-physical field data. Feature extraction and fusion are performed on heterogeneous multi-physics field data from different sources to obtain high-fidelity data with consistent representation and high confidence, which solves the problem that traditional technologies use sophisticated numerical calculation models or conduct real physical experiments, which are difficult to conform to physical reality and consume computing and human resources. In addition, physical partial differential equations are implicitly introduced in the training process to make full use of the rich physical information contained in the physical field data and achieve higher fusion accuracy of the model. The above method is suitable for data fusion scenarios under multi-scale, multi-precision, multi-resolution, strong nonlinearity and strong coupling, so that the multi-source physical field data processed by the model have consistent representation and high reliability in different dimensions.

[0072] Example 2:

[0073] An embodiment of the present invention provides a physical field data fusion method, comprising the steps of: performing simulation processing on a physical field based on a numerical calculation method to obtain low-fidelity physical field data; and

[0074] The physical field is simulated by any numerical calculation method such as finite difference method, finite element method and finite volume method to obtain low-fidelity physical field data.

[0075] In the above embodiments, the physical field is simulated by using any numerical calculation method such as the finite difference method, the finite element method, and the finite volume method, thereby achieving the acquisition of low-fidelity physical field data.

[0076] Example 3:

[0077] An embodiment of the present invention provides a physical field data fusion method, comprising the steps of obtaining the placement positions of sensors in a to-be-detected area and constructing a sensor mask matrix;

[0078] According to the size of the physical field in the area to be detected, an initial mask matrix is ​​set;

[0079] The positions where sensors are arranged in the area to be detected are obtained, and their corresponding values ​​in the initial mask matrix are set to "1". The corresponding values ​​of the positions where sensors are not arranged in the area to be detected are set to "0" in the corresponding initial mask matrix to construct a sensor mask matrix.

[0080] In the above embodiments, an initial mask matrix is ​​set according to the size of the physical field of the area to be detected, and the corresponding values ​​of the positions where sensors are arranged in the area to be detected in the initial mask matrix are set to "1", and the corresponding values ​​of the positions where sensors are not arranged in the corresponding initial mask matrix are set to "0" to construct a sensor mask matrix.

[0081] The beneficial effect of the above technology is that by constructing an initial mask matrix with the same size as the physical field of the area to be detected and setting the values ​​according to the sensor layout position, the construction of the sensor mask matrix is ​​realized, and then the matrix processing of low-fidelity physical field data and high-fidelity physical field data based on the sensor mask matrix is ​​realized, which facilitates the training and fine-tuning of the model in subsequent steps.

[0082] Example 4:

[0083] An embodiment of the present invention provides a physical field data fusion method, comprising the steps of updating and optimizing preliminary fused physical field data based on an iterative update coefficient to obtain updated sensor data;

[0084] S in =coe·S LF +(1-coe)·S HF

[0085] Among them, S in is to update the sensor data, coe is the iterative update coefficient, S LF For low-fidelity sensor data, S HF For high-fidelity sensor data;

[0086] coe=e -β·epoch

[0087] Among them, β is a hyperparameter and epoch is the number of iterations;

[0088] By continuously optimizing and adjusting the iterative update coefficient, the updated sensor data is made close to the high-fidelity sensor data.

[0089] In the above embodiments, the preliminary fused physical field data obtained based on the low-fidelity sensor data and the high-fidelity sensor data are fine-tuned based on the iterative update coefficient, and the obtained updated sensor data is made close to the high-fidelity sensor data by optimizing and adjusting the iterative update coefficient.

[0090] In the above embodiments, an iterative update coefficient is designed to constrain the update process.

[0091] The beneficial effects of the above technology are: by obtaining updated sensor data to fine-tune the data fusion model, optimize the performance of the data fusion model, and realize the fusion of multi-source heterogeneous and multi-physical field data.

[0092] Example 5:

[0093] An embodiment of the present invention provides a physical field data fusion method, comprising the steps of obtaining updated multi-order partial differential physical field data based on low-fidelity multi-order partial differential physical field data and fused partial differential physical field data; and comprising:

[0094] G in =coe·G LF +(1-coe)·G HF

[0095] Among them, G in To update the multi-order partial differential physics data, G LF is low-fidelity multi-order partial differential physical field data, G HF To integrate partial differential physics data;

[0096] By continuously optimizing and adjusting the iterative update coefficients, the updated multi-order partial differential physical field data is made close to the fused partial differential physical field data.

[0097] In the above embodiments, low-fidelity multi-order partial differential physical field data and fused partial differential physical field data are processed based on iterative update coefficients to obtain updated multi-order partial differential physical field data, and the iterative update coefficients are continuously optimized and adjusted so that the updated multi-order partial differential physical field data are close to the fused partial differential physical field data.

[0098] The beneficial effect of the above technology is that by obtaining and updating multi-order partial differential physical field data, it is convenient to further fine-tune the data fusion model, optimize the feature extraction performance and feature fusion performance of the data fusion model, and thus improve the model's fusion accuracy of the physical field data.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In addition, "front", "back", "left", "right", "upper" and "lower" in this document are all referenced to the placement states shown in the accompanying drawings.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A physical field data fusion method, characterized in that: include: Simulate the physical field based on numerical calculation methods to obtain low-fidelity physical field data; Obtain the placement of sensors in the area to be detected and construct a sensor mask matrix; multiplying the low-fidelity physical field data by the sensor mask matrix to obtain low-fidelity sensor data; Based on the finite difference method and Taylor expansion theory, a multi-order difference format is constructed; Processing the low-fidelity physical field data using the multi-order difference format to obtain low-fidelity multi-order partial differential physical field data; Build a pre-trained data fusion model; Training the pre-trained data fusion model based on the low-fidelity multi-order partial differential physical field data and the low-fidelity sensor data to obtain a data fusion model; Data is collected through sensors deployed in the area to be detected to obtain high-fidelity physical field data; multiplying the high-fidelity physical field data by the sensor mask matrix to obtain high-fidelity sensor data; fusing the low-fidelity sensor data and the high-fidelity sensor data through a data fusion model to obtain preliminary fused physical field data; Processing the preliminary fused physical field data using the multi-order difference format to obtain fused partial differential physical field data; Updating and optimizing the preliminary fused physical field data based on an iterative update coefficient to obtain updated sensor data; Acquire updated multi-order partial differential physical field data according to the low-fidelity multi-order partial differential physical field data and the fused partial differential physical field data; Fine-tuning the data fusion model according to the updated sensor data and the updated multi-order partial differential physical field data to obtain an optimized data fusion model; The optimized data fusion model is used for the fusion processing of multi-source heterogeneous and multi-physics field data.

2. A physical field data fusion method according to claim 1, characterized in that: The steps of simulating the physical field based on a numerical calculation method to obtain low-fidelity physical field data include: The physical field is simulated by any numerical calculation method such as finite difference method, finite element method and finite volume method to obtain low-fidelity physical field data.

3. The physical field data fusion method according to claim 1, characterized in that: The step of obtaining the placement positions of sensors in the area to be detected and constructing a sensor mask matrix includes: According to the size of the physical field in the area to be detected, an initial mask matrix is ​​set; The positions where sensors are arranged in the area to be detected are obtained, and their corresponding values ​​in the initial mask matrix are set to "1". The corresponding values ​​of the positions where sensors are not arranged in the area to be detected are set to "0" in the corresponding initial mask matrix to construct a sensor mask matrix.

4. The physical field data fusion method according to claim 1, characterized in that: The step of updating and optimizing the preliminary fused physical field data based on the iterative update coefficient to obtain updated sensor data includes: S in =coe·S LF +(1-coe)·S HG Among them, S in is to update the sensor data, coe is the iterative update coefficient, S LF is low-fidelity sensor data, S HF For high-fidelity sensor data; coe=e -β·epoch Among them, β is a hyperparameter and epoch is the number of iterations; By continuously optimizing and adjusting the iterative update coefficient, the updated sensor data is made to approach the high-fidelity sensor data.

5. A physical field data fusion method according to claim 4, characterized in that: The step: obtaining updated multi-order partial differential physical field data according to the low-fidelity multi-order partial differential physical field data and the fused partial differential physical field data; include: G in =coe·G LF +(1-coe)·G HF Among them, G in To update the multi-order partial differential physics data, G LF is low-fidelity multi-order partial differential physical field data, G HF To integrate partial differential physics data; By continuously optimizing and adjusting the iterative update coefficient, the updated multi-order partial differential physical field data is made close to the fused partial differential physical field data.

Images