Mass physical field data inversion method based on feature point driving
By constructing a multidimensional observation matrix and using a feature point-driven approach, and reconstructing the physical field using neural networks and Gaussian kernel functions, the problem of large computational load and poor adaptability of traditional methods under limited measurement point conditions is solved, and efficient and accurate inversion and reconstruction of the physical field in complex systems is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-01-13
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies struggle to efficiently infer the continuous physical field distribution in complex systems under limited measurement point conditions. This is especially true in aerospace and weaponry, where traditional methods suffer from high computational complexity, high cost, or poor adaptability, making it difficult to meet real-time and accuracy requirements.
By collecting multi-source physical field data, a multi-dimensional observation matrix is constructed for preprocessing. The correlation between the measurement points and the global physical field is calculated using neural networks and the Hilbert-Schmidt independence criterion. Key feature points are screened, feature fusion is performed using a multilayer perceptron, and the physical field is reconstructed using weighted interpolation with a Gaussian kernel function.
It enables efficient and accurate reconstruction of the continuous spatial distribution of complex physical fields under conditions with very few measurement points, improving the scientific nature and stability of the inversion. It is applicable to the rapid reconstruction of various physical fields such as temperature field, stress field, and flow field, and enhances the safety and intelligence level of system operation.
Smart Images
Figure CN121901643A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data inversion and physical modeling technology, specifically to a method for inverting massive physical field data based on feature point-driven approaches. Background Technology
[0002] With the rapid development of aerospace and weaponry, complex systems contain a variety of coupled physical field processes, such as structural stress field, temperature field, flow field, and electromagnetic field. The distribution of these physical fields directly affects the operational safety and health of the system.
[0003] Existing physical field inversion and monitoring methods mainly rely on full-field measurements or high-dimensional numerical simulations. For example, the Finite Element Method (FEM) and Finite Volume Method (FVM) are commonly used for temperature or stress field reconstruction, but their computational demands are high, making it difficult to meet real-time requirements. On the other hand, full-field testing based on experimental measurements (such as infrared thermography and strain holographic measurement) is costly, complex to deploy, and often difficult to implement in extreme environments such as flight, launch, and deep sea. Furthermore, while existing deep learning-based inversion methods can achieve rapid prediction of high-dimensional data, they typically rely on large-scale labeled samples, have poor adaptability to different operating conditions, and exhibit significantly reduced inversion accuracy in sparse or noisy environments.
[0004] Therefore, there is an urgent need for a method that can efficiently deduce the distribution of continuous physical fields under limited measurement points, in order to achieve inversion modeling "from point to surface". Especially in application scenarios such as thermal stress estimation of rocket control cabins, heat dissipation monitoring of aircraft power compartments, and health assessment of satellite payload panels, if high-dimensional fields can be rapidly inverted based on a small number of feature points, it will greatly improve the safety and intelligence of system operation. Summary of the Invention
[0005] The purpose of this invention is to provide a feature-point driven method for inverting massive physical field data to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a feature-point driven method for inverting massive physical field data, comprising the following steps:
[0007] S1. Collect multi-source physical field data of the target system under different operating conditions, construct a multi-dimensional observation matrix, and perform wavelet threshold denoising, normalization, outlier removal, interpolation completion and time alignment preprocessing on the data in sequence.
[0008] S2. For each measuring point, construct time or working condition characteristics, and use the mutual information estimation method based on neural networks and the Hilbert-Schmidt independence criterion to calculate the correlation index between the measuring point and the global physical field. After normalizing the two indices, a weighted combination is used to obtain the comprehensive correlation score.
[0009] S3. Determine an adaptive threshold based on the mean, standard deviation and adjustment factor of the comprehensive correlation scores of all measurement points, and select measurement points with comprehensive correlation scores not lower than the threshold as the set of key feature points.
[0010] S4. Stack the preprocessed observation data of key feature points into an input matrix, input it into a multilayer perceptron for feature fusion, and output a low-dimensional fused feature vector.
[0011] S5. Based on the spatial coordinates of key feature points and the fused feature vector, a Gaussian kernel function is used to construct spatial weights, and a weighted interpolation method is used to realize the inversion and reconstruction of the physical field at any spatial location.
[0012] Furthermore, in step S1, the expression for the multidimensional observation matrix is: ,in, Indicates the length of the time series. This indicates the number of measurement points; the preprocessed data is denoted as... Multi-source physical field data includes at least one of temperature field data, stress field data, flow field data, and electromagnetic field data.
[0013] Furthermore, in step S2, the mutual information estimation formula based on the neural network is as follows:
[0014]
[0015] in, The joint probability distribution of measurement point characteristics and global physical field quantities. and Their respective edge distributions, This serves as a reference quantity for the overall physical field; the Hilbert–Schmidt independence criterion is used to calculate the dependence between the measurement point and the global field, as shown in the formula:
[0016]
[0017] The two indicators mentioned above are normalized and weighted to form a comprehensive correlation score for each measurement point. .
[0018] Furthermore, in step S2, the overall correlation score is calculated. The calculation formula is
[0019]
[0020] in, , These are the weighting coefficients. .
[0021] Furthermore, in step S3, the formula for calculating the adaptive threshold is:
[0022]
[0023] in , They are respectively The mean and standard deviation; As a regulating factor;
[0024] when At that time, the measuring point Included in the key feature point set, the expression for the key feature point set is: , where M is the total number of measurement points.
[0025] Furthermore, in step S4, the set of key feature points obtained in S3 is... The observation data from each measuring point are stacked into an input matrix, and the expression for the input matrix is as follows:
[0026]
[0027] in ;
[0028] matrix The forward propagation formula for a multilayer perceptron is:
[0029]
[0030] in, To fuse feature vectors; , These are the weights and bias parameters, respectively; is the activation function for the Gaussian error linear unit.
[0031] Furthermore, in step S5, the spatial coordinates of each key feature point obtained from S3... The fused feature vector obtained from S4 For any spatial location The physical field distribution is inverted and reconstructed;
[0032] Spatial weights are constructed using a Gaussian kernel function. The formula for calculating the Gaussian kernel function is as follows:
[0033]
[0034] in, The spatial location of the feature point; The Gaussian kernel width parameter;
[0035] The spatial location of the physical field is calculated using a weighted interpolation-based reconstruction method. The estimated value The formula for calculating the reconstructed physical field value is:
[0036]
[0037] in, To fuse feature vectors with feature points The corresponding components.
[0038] Compared with the prior art, the beneficial effects of the present invention are:
[0039] 1. The feature point evaluation and selection method described in this invention accurately selects representative key feature points through nonlinear correlation analysis, avoiding the redundancy and bias problems that may occur in traditional point selection methods that rely on manual experience or linear correlation, and improving the scientificity and stability of feature selection.
[0040] 2. The fusion feature construction and kernel function reconstruction method described in this invention can effectively recover the continuous spatial distribution of complex physical fields under conditions of very few measurement points, overcome the shortcomings of high cost of traditional full-field measurement and large amount of numerical simulation calculation, and realize efficient and accurate inversion of high-dimensional physical fields.
[0041] 3. The feature-point-driven massive physical field data inversion method described in this invention is applicable to the rapid reconstruction of various physical fields such as temperature, stress, and flow fields. It can play a crucial role in high-reliability scenarios such as rocket module thermal stress assessment, aircraft power compartment heat dissipation monitoring, and satellite payload structural health management, demonstrating significant engineering application value. Attached Figure Description
[0042] Figure 1 This is an overall flowchart of the present invention;
[0043] Figure 2 A schematic diagram of the feature point evaluation and selection method of the present invention;
[0044] Figure 3 A network structure diagram is constructed for the feature point-driven fusion features of this invention;
[0045] Figure 4 This is a schematic diagram of the physical field kernel interpolation reconstruction method of the present invention. Detailed Implementation
[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0047] Please see Figure 1-4 This invention provides a feature-point driven method for inverting massive physical field data, comprising the following steps:
[0048] S1. Collect multi-source physical field data of the target system under different operating conditions, construct a multi-dimensional observation matrix, and perform wavelet threshold denoising, normalization, outlier removal, interpolation completion and time alignment preprocessing on the data in sequence.
[0049] S2. For each measuring point, construct time or working condition characteristics, and use the mutual information estimation method based on neural networks and the Hilbert-Schmidt independence criterion to calculate the correlation index between the measuring point and the global physical field. After normalizing the two indices, a weighted combination is used to obtain the comprehensive correlation score.
[0050] S3. Determine an adaptive threshold based on the mean, standard deviation and adjustment factor of the comprehensive correlation scores of all measurement points, and select measurement points with comprehensive correlation scores not lower than the threshold as the set of key feature points.
[0051] S4. Stack the preprocessed observation data of key feature points into an input matrix, input it into a multilayer perceptron for feature fusion, and output a low-dimensional fused feature vector.
[0052] S5. Based on the spatial coordinates of key feature points and the fused feature vector, a Gaussian kernel function is used to construct spatial weights, and a weighted interpolation method is used to realize the inversion and reconstruction of the physical field at any spatial location.
[0053] Specifically, the following steps are included:
[0054] S1. Data Acquisition and Preprocessing: The system acquires multi-source physical field data under different operating conditions to form the original observation matrix, as shown in the following formula:
[0055]
[0056] in, Indicates the length of the time series; Indicates the number of measurement points.
[0057] Wavelet thresholding was used to denoise the data to eliminate environmental noise interference. Then, normalization, outlier removal, interpolation completion, and time alignment were performed. The processed data is denoted as follows: .
[0058] S2. Evaluation of nonlinear correlation of feature points: such as Figure 2 As shown, in order to determine the contribution of each observation point to the overall physical field distribution, a nonlinear correlation assessment is performed on the relationship between all measurement points and the target physical field.
[0059] For each measuring point The preprocessed observation sequence is denoted as And construct the corresponding time or operating condition characteristics. A neural network-based mutual information estimation method is adopted, utilizing a neural estimator. The lower bound of the mutual information between the measurement point features and the target output is calculated using the following formula:
[0060]
[0061] in, The joint probability distribution of measurement point characteristics and global physical field quantities. and Their respective marginal distributions; It serves as a reference value for the overall physical field.
[0062] The dependence between the measurement point and the global field is calculated using the Hilbert–Schmidt independence criterion (HSIC), as shown in the following formula:
[0063]
[0064] The two indicators mentioned above are normalized and weighted to form a comprehensive correlation score for each measurement point. The formula is as follows:
[0065]
[0066] in, , These are the weighting coefficients.
[0067] S3. Key Feature Point Selection: An adaptive threshold is used, as shown in the following formula:
[0068]
[0069] in, , They are respectively The mean and standard deviation; It is a regulating factor.
[0070] when At that time, the measuring point The formula for incorporating key feature points into the set is as follows:
[0071]
[0072] Where M is the total number of measuring points.
[0073] This method can automatically select a small number of high-value measurement points, improving inversion efficiency without affecting inversion accuracy.
[0074] S4. Feature Fusion Construction: such as Figure 3 As shown, the set of key feature points obtained in S3 The observation data from each measuring point are stacked into an input matrix, as shown in the following formula:
[0075]
[0076] in, .
[0077] matrix The forward propagation formula for an input multilayer perceptron (MLP) is as follows:
[0078]
[0079] in, To fuse feature vectors; , These are the weights and bias parameters, respectively; is the activation function for the Gaussian error linear unit.
[0080] Fusion feature vectors It simultaneously contains the topological relationships and physical coupling information between key measurement points in a low-dimensional space.
[0081] S5. Physical Field Inversion and Reconstruction: such as Figure 4 As shown, the spatial coordinates of each key feature point obtained from S3 The fused feature vector obtained from S4 For any spatial location The physical field distribution is inverted and reconstructed.
[0082] The spatial weights are constructed using a Gaussian kernel function, as shown in the following formula:
[0083]
[0084] in, The spatial location of the feature point; The Gaussian kernel width parameter;
[0085] The spatial location of the physical field is calculated using a weighted interpolation-based reconstruction method. The estimated value The formula is as follows:
[0086]
[0087] in, To fuse feature vectors with feature points The corresponding components.
[0088] The above reconstruction process ensures that spatially adjacent feature points contribute more to the reconstructed value, while regions far from feature points have a relatively weaker influence, thus obtaining physical field results that conform to spatial continuity.
[0089] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for inverting massive physical field data based on feature points, characterized in that: Includes the following steps: S1. Collect multi-source physical field data of the target system under different operating conditions, construct a multi-dimensional observation matrix, and perform wavelet threshold denoising, normalization, outlier removal, interpolation completion and time alignment preprocessing on the data in sequence. S2. For each measuring point, construct time or working condition characteristics, and use the mutual information estimation method based on neural networks and the Hilbert-Schmidt independence criterion to calculate the correlation index between the measuring point and the global physical field. After normalizing the two indices, weighted combination is used to obtain the comprehensive correlation score. S3. Determine an adaptive threshold based on the mean, standard deviation and adjustment factor of the comprehensive correlation scores of all measurement points, and select measurement points with comprehensive correlation scores not lower than the threshold as the set of key feature points. S4. Stack the preprocessed observation data of key feature points into an input matrix, input it into a multilayer perceptron for feature fusion, and output a low-dimensional fused feature vector. S5. Based on the spatial coordinates of key feature points and the fused feature vector, a Gaussian kernel function is used to construct spatial weights, and a weighted interpolation method is used to realize the inversion and reconstruction of the physical field at any spatial location.
2. The method for inverting massive physical field data based on feature points as described in claim 1, characterized in that: In step S1, the expression for the multidimensional observation matrix is: ,in, Indicates the length of the time series. This indicates the number of measurement points; the preprocessed data is denoted as... Multi-source physical field data includes at least one of temperature field data, stress field data, flow field data, and electromagnetic field data.
3. The method for inverting massive physical field data based on feature point-driven approach according to claim 1, characterized in that: In step S2, the mutual information estimation formula based on the neural network is as follows: ， in, The joint probability distribution of measurement point characteristics and global physical field quantities. and Their respective edge distributions, This serves as a reference quantity for the overall physical field; the Hilbert–Schmidt independence criterion is used to calculate the dependence between the measurement point and the global field, as shown in the formula: ， The two indicators mentioned above are normalized and weighted to form a comprehensive relevance score for each measurement point. .
4. The method for inverting massive physical field data based on feature point-driven approach according to claim 3, characterized in that: In step S2, the overall correlation score is calculated. The calculation formula is: ， in, , These are the weighting coefficients. .
5. The method for inverting massive physical field data based on feature point-driven approach according to claim 1, characterized in that: In step S3, the formula for calculating the adaptive threshold is: ， in , They are respectively The mean and standard deviation; As a regulating factor; when At that time, the measuring point Included in the key feature point set, the expression for the key feature point set is: , where M is the total number of measurement points.
6. The method for inverting massive physical field data based on feature points according to claim 1, characterized in that: In step S4, the set of key feature points obtained in S3 is... The observation data from each measuring point are stacked into an input matrix, and the expression for the input matrix is as follows: ， in ; matrix The forward propagation formula for a multilayer perceptron is: ， in, To fuse feature vectors; , These are the weights and bias parameters, respectively; is the activation function for the Gaussian error linear unit.
7. The method for inverting massive physical field data based on feature point-driven approach according to claim 1, characterized in that: In step S5, the spatial coordinates of each key feature point obtained from S3 are described. The fused feature vector obtained from S4 For any spatial location The physical field distribution is inverted and reconstructed; Spatial weights are constructed using a Gaussian kernel function. The formula for calculating the Gaussian kernel function is as follows: ， in, The spatial location of the feature point; The Gaussian kernel width parameter; The spatial location of the physical field is calculated using a weighted interpolation-based reconstruction method. The estimated value The formula for calculating the reconstructed physical field value is: ， in, To fuse feature vectors with feature points The corresponding components.