Efficient physical field reconstruction method
By combining modulation and demodulation technology with convolutional neural networks and optimizing the control parameters of convolutional neural networks, the computational efficiency and accuracy issues of convolutional neural networks in processing irregular grid data and reconstructing high-frequency features are solved, thus achieving efficient physical field reconstruction.
Patent Information
- Application Number
- CN202510909316.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-09-23
AI Technical Summary
In existing technologies, convolutional neural networks cannot directly process irregular structured or unstructured grid data, resulting in high computational cost and low accuracy when reconstructing complex physical fields. In addition, the frequency domain reconstruction paradigm has low computational efficiency and is difficult to meet the high-frequency feature response capabilities.
Combining modulation and demodulation technology with convolutional neural networks, low-dimensional vector space mapping and frequency domain information transformation are achieved through pre-training of convolutional neural network layers, modulation layers, and demodulation layers, thereby optimizing the control parameters of the convolutional neural network and improving the response capability to frequency domain information flows.
It improves the reconstruction accuracy and high-frequency feature fitting ability of convolutional neural networks for complex physical fields, reduces the computational cost, enhances the deployability of the network, and realizes the effective fusion of multi-scale frequency domain information.
Smart Images
Figure CN120688556A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of physical field reconstruction, and in particular to an efficient physical field reconstruction method. Background Art
[0002] As equipment becomes more complex and sophisticated, the need for comprehensive monitoring of equipment status is becoming increasingly urgent. Deep neural network technology is being proposed to quickly and accurately reconstruct physical field data based on sparse observations or operating conditions. Deep neural networks can leverage the advantages of big data and powerful nonlinear fitting capabilities to characterize physical field characteristics, enabling rapid and accurate reconstruction of physical field data.
[0003] The physical field reconstruction method based on convolutional neural networks reconstructs directly from the spatial domain and requires that the original data collected from the physical field must be a regular matrix. However, in the actual data processing process, the physical field data collection comes from irregular structured grids or unstructured grids, which makes the convolutional neural network unable to process directly. In addition, when reconstructing complex and high-frequency physical features with large gradients, it faces the problems of high computational cost and low reconstruction accuracy.
[0004] To address this problem, one approach proposed using interpolation methods to convert non-matrix data into matrix data. However, this interpolation method can cause a homogenization effect, smoothing physical regions with large gradient characteristics, leading to a serious loss of physical field features. While physical field reconstruction methods based on one-dimensional vectorization offer significant advantages in preserving the original physical field features, the one-dimensional vectorization of the physical field disrupts the frequency domain feature distribution of the physical field, further reducing the convolutional neural network's ability to respond to high frequencies.
[0005] To improve the high-frequency response of convolutional neural networks to physical fields, it has been proposed to transform spatial physical features into the frequency domain and utilize a frequency-domain learning paradigm, such as the Fourier neural operator, to enhance the convolutional neural network's ability to represent high-frequency physical features. The frequency-domain reconstruction paradigm requires the use of Fourier transforms to convert between the spatial and frequency domains. However, physical fields are typically high-dimensional, and high-dimensional Fourier transforms often consume significant computational resources and are very inefficient. This presents the challenge of balancing computational efficiency with the ability to fit high-frequency features.
[0006] To reduce the computational inefficiency of the frequency-domain reconstruction paradigm, a proposed approach to improve the network structure based on the network frequency principle is to make the convolutional neural network converge first on low-frequency features and then on high-frequency features. However, in practical applications, this approach has difficulty meeting the requirements for physical field data reconstruction.
[0007] Therefore, an efficient physical field reconstruction method is proposed, which performs high-frequency enhanced reconstruction of the physical field based on modulation and demodulation technology. 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 an efficient physical field reconstruction method, which performs high-frequency enhanced reconstruction of the physical field based on modulation and demodulation technology, and is used to solve the problems of poor response ability of deep neural networks to high-frequency features and low accuracy of physical field reconstruction in complex physical field environments.
[0009] The technical solutions of the present invention are as follows:
[0010] Provides an efficient physical field reconstruction method, including:
[0011] Constructing a pre-trained convolutional neural network layer, and training the pre-trained convolutional neural network layer using a training set to obtain a convolutional neural network layer;
[0012] Obtain the measurement point information of the detection area;
[0013] Mapping the measurement point information to a low-dimensional vector space through a modulation layer to obtain low-dimensional measurement point information;
[0014] Inputting the low-dimensional measurement point information into a convolutional neural network layer for nonlinear convolution operation, and outputting low-dimensional physical field information;
[0015] The low-dimensional physical field information is subjected to Fourier transform through the demodulation layer to obtain low-frequency frequency domain physical field information, and frequency domain demodulation is performed to transform the low-frequency frequency domain physical field information into high-frequency frequency domain space to obtain high-frequency frequency domain physical field information, and the high-frequency frequency domain physical field information is transformed into the spatial domain through inverse Fourier transform to obtain predicted physical field data.
[0016] Preferably, in the provided efficient physical field reconstruction method, the step of: constructing a pre-trained convolutional neural network layer, training the pre-trained convolutional neural network layer using a training set to obtain a convolutional neural network layer; comprises:
[0017] Obtain sample measurement point information and standard sample physical field data in the training set;
[0018] Processing the sample measurement point information through a modulation layer to obtain low-dimensional sample measurement point information;
[0019] The pre-trained convolutional neural network layer processes the low-dimensional sample measurement point information and outputs low-dimensional sample physical field information;
[0020] Demodulating the low-dimensional sample physical field information through a demodulation layer to obtain predicted sample physical field data;
[0021] The deviation between the predicted sample physical field data and the standard sample physical field data is analyzed based on the loss function, and the control parameters of the pre-trained convolutional neural network layer are optimized.
[0022] The main advantages of the technical solution of the present invention are as follows:
[0023] The efficient physical field reconstruction method of the present invention combines the modulation and demodulation technology in the communication field with the convolutional neural network. By optimizing the control parameters in the convolutional neural network, the response capability to the spectrum distribution characteristics of the frequency domain information flow is improved, the difficulty of the convolutional neural network in fitting high-frequency features is reduced, and its reconstruction effect in processing complex physical features with large gradient characteristics is improved. It realizes the effective fusion of multi-scale frequency domain information of the physical field, enhances the fitting accuracy of the high-frequency features of the physical field, unifies the characterization capability of the convolutional neural network for information in each frequency band, and thus effectively improves the reconstruction accuracy of the physical field. The above method can not only retain the complete original physical field characteristics, but also enhance the characterization capability of the convolutional neural network for the high-frequency information of complex physical fields, so that the convolutional neural network can accurately reconstruct complex high-frequency physical field features with large gradient characteristics, greatly reducing the computational cost, and thus improving the deployability of the convolutional neural network. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] 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:
[0025] Figure 1 A schematic flow chart of an efficient physical field reconstruction method provided by the present invention;
[0026] Figure 2 A schematic diagram of the modulation layer structure of an efficient physical field reconstruction method provided by the present invention;
[0027] Figure 3 Schematic diagram of the demodulation layer structure of an efficient physical field reconstruction method provided by the present invention. DETAILED DESCRIPTION
[0028] 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.
[0029] The technical solutions provided by the embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0030] Example 1:
[0031] The embodiment of the present invention provides an efficient physical field reconstruction method, referring to Figure 1 、 Figure 2 and Figure 3 ,include:
[0032] Construct a pre-trained convolutional neural network layer, train the pre-trained convolutional neural network layer using the training set, and obtain the convolutional neural network layer;
[0033] Obtain the measurement point information of the detection area;
[0034] Mapping the measurement point information to a low-dimensional vector space through the modulation layer to obtain low-dimensional measurement point information;
[0035] Input the low-dimensional measurement point information into the convolutional neural network layer for nonlinear convolution operation, and output low-dimensional physical field information;
[0036] The low-dimensional physical field information is Fourier transformed through the demodulation layer to obtain low-frequency frequency domain physical field information, and frequency domain demodulation is performed to transform the low-frequency frequency domain physical field information into high-frequency frequency domain space to obtain high-frequency frequency domain physical field information. The high-frequency frequency domain physical field information is transformed into the spatial domain through the inverse Fourier transform to obtain predicted physical field data.
[0037] In the above embodiments, the pre-trained convolutional neural network layer is trained using a training set to obtain a convolutional neural network layer.
[0038] In the above embodiments, the measurement point information of the detection area is mapped to the low-dimensional vector space through the modulation layer to obtain the low-dimensional measurement point information, and the low-dimensional measurement point information is input into the convolutional neural network layer for nonlinear convolution operation to output low-dimensional physical field information, and the low-dimensional physical field information is Fourier transformed through the demodulation layer to obtain low-frequency frequency domain physical field information, and frequency domain demodulation is performed to transform the low-frequency frequency domain physical field information into the high-frequency frequency domain space to obtain high-frequency frequency domain physical field information, and the high-frequency frequency domain physical field information is transformed into the spatial domain through the inverse Fourier transform to obtain predicted physical field data.
[0039] The beneficial effects of the above technology are: mapping the measurement point information to the low-dimensional vector space through the modulation layer, avoiding the requirement that the traditional physical field reconstruction paradigm must be defined in a high-dimensional space of the same dimension as the physical field, thereby realizing spatial dimensionality reduction and frequency domain frequency reduction. The convolutional neural network processes the low-dimensional measurement point information and outputs low-dimensional physical field information. The demodulation layer performs demodulation processing to map the modulated frequency domain information concentrated in the low-dimensional vector space back to the high-frequency space, and then assembles it into a complete high-dimensional space physical field to output predicted physical field data. Compared with traditional technologies, the efficient physical field reconstruction method of the embodiment of the present invention combines the modulation and demodulation technology in the communication field with the convolutional neural network, and improves the control of the frequency domain information flow by optimizing the control parameters in the convolutional neural network. The above method not only retains the complete original physical field characteristics, but also enhances the convolutional neural network’s ability to represent high-frequency information of complex physical fields, enabling the convolutional neural network to accurately reconstruct complex, high-frequency physical field characteristics with large gradient characteristics, greatly reducing the computational cost, and thus improving the deployability of the convolutional neural network.
[0040] Example 2:
[0041] An embodiment of the present invention provides an efficient physical field reconstruction method, comprising the steps of: constructing a pre-trained convolutional neural network layer, training the pre-trained convolutional neural network layer using a training set, and obtaining a convolutional neural network layer; comprising:
[0042] Obtain sample measurement point information and standard sample physical field data in the training set;
[0043] The sample measurement point information is processed through the modulation layer to obtain low-dimensional sample measurement point information;
[0044] The pre-trained convolutional neural network layer processes the low-dimensional sample measurement point information and outputs low-dimensional sample physical field information;
[0045] The low-dimensional sample physical field information is demodulated by the demodulation layer to obtain the predicted sample physical field data;
[0046] The deviation between the sample physical field data and the standard sample physical field data is predicted based on the loss function analysis, and the control parameters of the pre-trained convolutional neural network layer are optimized.
[0047] In the above embodiments, the sample measurement point information in the training set is processed through the modulation layer, the pre-trained convolutional neural network layer and the demodulation layer to obtain the predicted sample physical field data. The deviation between the predicted sample physical field data and the standard sample physical field data is analyzed based on the loss function, and the control parameters of the pre-trained convolutional neural network layer are optimized.
[0048] The beneficial effects of the above technology are: through the sample measurement point information and standard sample physical field data in the training set, the optimized training of the pre-trained convolutional neural network layer is realized to obtain the convolutional neural network layer; specifically, by optimizing the control parameters in the low-frequency band of the pre-trained convolutional neural network layer, the prediction performance of the convolutional neural network layer for low-frequency physical field data is improved, and through the modulation of the spectrum by the modulation layer and the demodulation layer, the frequency domain information flow of the control parameters of the convolutional neural network is restricted to the low-dimensional space, which effectively reduces the training cost of the pre-trained convolutional neural network layer and improves the optimization training efficiency.
[0049] 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.
[0050] 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. An efficient physical field reconstruction method, characterized in that: include: Constructing a pre-trained convolutional neural network layer, and training the pre-trained convolutional neural network layer using a training set to obtain a convolutional neural network layer; Obtain the measurement point information of the detection area; Mapping the measurement point information to a low-dimensional vector space through a modulation layer to obtain low-dimensional measurement point information; Inputting the low-dimensional measurement point information into a convolutional neural network layer for nonlinear convolution operation, and outputting low-dimensional physical field information; The low-dimensional physical field information is subjected to Fourier transform through the demodulation layer to obtain low-frequency frequency domain physical field information, and frequency domain demodulation is performed to transform the low-frequency frequency domain physical field information into high-frequency frequency domain space to obtain high-frequency frequency domain physical field information, and the high-frequency frequency domain physical field information is transformed into the spatial domain through inverse Fourier transform to obtain predicted physical field data.
2. An efficient physical field reconstruction method according to claim 1, characterized in that: The step of constructing a pre-trained convolutional neural network layer and training the pre-trained convolutional neural network layer using a training set to obtain a convolutional neural network layer includes: Obtain sample measurement point information and standard sample physical field data in the training set; Processing the sample measurement point information through a modulation layer to obtain low-dimensional sample measurement point information; The pre-trained convolutional neural network layer processes the low-dimensional sample measurement point information and outputs low-dimensional sample physical field information; Demodulating the low-dimensional sample physical field information through a demodulation layer to obtain predicted sample physical field data; The deviation between the predicted sample physical field data and the standard sample physical field data is analyzed based on the loss function, and the control parameters of the pre-trained convolutional neural network layer are optimized.