CN-122021724-A - Data reconstruction method, device, equipment and storage medium

CN122021724ACN 122021724 ACN122021724 ACN 122021724ACN-122021724-A

Abstract

The embodiment of the invention discloses a data reconstruction method, a device, equipment and a storage medium, wherein the method comprises the steps of obtaining various initial sample sequence data, wherein the data acquisition step length of the initial sample sequence data is not fixed, the sequence length and the data acquisition step length between the initial sample sequence data are different, the data reconstruction processing is respectively carried out on each initial sample sequence data to obtain target sample sequence data corresponding to the initial sample sequence data, the data acquisition step length of the target sample sequence data is fixed, the sequence length and the data acquisition step length between the target sample sequence data are the same, and training is carried out on an initial neural network model based on the target sample sequence data to obtain a target neural network model. The technical scheme of the embodiment of the invention can reconstruct the data of the model training sample data, and perform standardized and unified processing on the sequence length and the data acquisition step length of the sample data, thereby greatly improving the convergence speed and the success rate of model training.

Inventors

ZHU LILI
ZHANG YINGQIN
CHEN HUI
ZHU ANDI
XU MINGYANG
LI XIANG

Assignees

苏州同元软控技术股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (10)

1. A method of data reconstruction, comprising: Acquiring a plurality of initial sample sequence data, wherein the data acquisition step length of the initial sample sequence data is not fixed, and the sequence length and the data acquisition step length between the initial sample sequence data are different; Respectively carrying out data reconstruction processing on each initial sample sequence data to obtain target sample sequence data corresponding to the initial sample sequence data, wherein the data acquisition step length of the target sample sequence data is fixed, and the sequence length and the data acquisition step length between the target sample sequence data are the same; Training the initial neural network model based on the target sample sequence data to obtain a target neural network model.
2. The method according to claim 1, wherein the performing data reconstruction processing on each initial sample sequence data to obtain target sample sequence data corresponding to the initial sample sequence data includes: Determining a target time interval and a target data step length based on the sequence time intervals of all initial sample sequence data; Determining a target interpolation mode based on the data change degree in the initial sample sequence data aiming at each initial sample sequence data; And carrying out interpolation processing on the initial sample sequence data in the target time interval based on the target interpolation mode and the target data step length to obtain target sample sequence data.
3. The method of claim 2, wherein the determining the target time interval and the target data step based on the sequence time interval of all initial sample sequence data comprises: Determining a minimum starting time and a maximum ending time from a sequence time interval of all initial sample sequence data, and forming the target time interval based on the minimum starting time and the maximum ending time; And determining a working condition data type corresponding to the initial sample sequence data, matching the corresponding data node number based on the working condition data type, and determining the target data step length based on the target time interval and the data node number.
4. The method of claim 2, wherein the determining a target interpolation based on the degree of data change in the initial sample sequence data comprises: Determining a data gradient value corresponding to the initial sample sequence data, and determining that the target interpolation mode is linear interpolation under the condition that the data gradient value is smaller than a preset gradient threshold value; And under the condition that the data gradient value is larger than a preset gradient threshold value, determining a data curvature value corresponding to the initial sample sequence data, and under the condition that the data curvature value is smaller than the preset curvature threshold value, determining that the target interpolation mode is cubic spline interpolation.
5. The method according to claim 2, wherein the interpolating the initial sample sequence data in the target time interval based on the target interpolation mode and the target data step length to obtain target sample sequence data includes: Performing interpolation processing on the initial sample sequence data based on the target interpolation mode, and determining a corresponding continuous approximation function based on the interpolated data; And determining a plurality of target data time points in a target time interval based on the target data step length, respectively determining an approximate value corresponding to each target data time point based on the continuous approximate function, and taking the approximate values corresponding to all the target data time points as the target sample sequence data.
6. The method of claim 1, wherein training the initial neural network model based on the target sample sequence data to obtain a target neural network model comprises: inputting the target sample sequence data into the initial neural network model to obtain a model output result; and determining a model analysis loss function according to the model output result, and adjusting the initial neural network model according to the model analysis loss function to obtain the target neural network model.
7. The method according to claim 2, wherein the method further comprises: Acquiring corresponding initial key physical characteristics between the initial sample sequence data and corresponding target key physical characteristics between the target sample sequence data; And adjusting an interpolation mode corresponding to the initial sample sequence data based on the characteristic difference information between the initial key physical characteristics and the target key physical characteristics under the condition that the characteristic difference information meets a preset adjustment condition.
8. A data reconstruction apparatus, the apparatus comprising: The data acquisition module is used for acquiring various initial sample sequence data, wherein the data acquisition step length of the initial sample sequence data is not fixed, and the sequence length and the data acquisition step length between the initial sample sequence data are different; The data processing module is used for respectively carrying out data reconstruction processing on each initial sample sequence data to obtain target sample sequence data corresponding to the initial sample sequence data, wherein the data acquisition step length of the target sample sequence data is fixed, and the sequence length and the data acquisition step length between the target sample sequence data are the same; And the model training module is used for training the initial neural network model based on the target sample sequence data to obtain a target neural network model.
9. A computer device, the computer device comprising: one or more processors; A memory for storing one or more programs; When executed by the one or more processors, causes the one or more processors to implement the data reconstruction method as recited in any one of claims 1-7.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements a data reconstruction method as claimed in any one of claims 1-7.

Description

Data reconstruction method, device, equipment and storage medium Technical Field The embodiment of the invention relates to the technical field of data processing and model training, in particular to a data reconstruction method, a device, equipment and a storage medium. Background In the field of engineering simulation of computational fluid mechanics, structural dynamics and the like, a high-fidelity full-order model is a foundation stone for design and analysis. Currently, snapshot-based eigen-orthogonal decomposition and other order reduction methods are mature, but their successful application strongly depends on a basic assumption that training snapshot data is uniformly sampled in the time dimension. This assumption is often broken in practical high-fidelity simulations. To accurately capture complex physical phenomena (such as shock wave propagation, vortex shedding collision) such as transients and nonlinearities, modern simulators commonly employ an adaptive variable time step strategy, which results in non-uniform distribution of directly obtained snapshot sets on a time axis. The high-fidelity simulator uses adaptive time steps that result in snapshots that are unevenly distributed in time (e.g., small steps when the physical changes are severe and large steps when the changes are gradual). Simulations under different parameter conditions may end at different time points due to small differences in dynamics, or the adaptive strategy may produce a completely different number of time steps, resulting in different snapshot sequence lengths for each condition. This is fatal for recurrent neural networks such as LSTM, as they generally require: The fixed time step length is that the time interval of the input sequence is constant so as to learn the evolution rule in time by the network, and the equal-length input sequence is that the sequence length of all samples in one training batch is consistent so as to perform efficient batch matrix operation and back propagation. Aiming at the challenge, academia and industry have explored a plurality of paths, and the field generally faces a long-standing technical bottleneck that the non-uniformity of original snapshot data generated by high-fidelity simulation in the time dimension is irreconcilable with the requirement of the LSTM network on the uniformity and the equal length of an input sequence in the field of data-driven model reduction, especially in the process of constructing a dynamic proxy model by adopting a cyclic neural network such as a long-term memory network. Disclosure of Invention The embodiment of the invention provides a data reconstruction method, a device, equipment and a storage medium, the technical scheme of the embodiment of the invention can ensure the stability of gradient back propagation in the deep learning training process and greatly improve the convergence speed and success rate of model training. In a first aspect, an embodiment of the present invention provides a data reconstruction method, where the method includes: The method comprises the steps of obtaining multiple initial sample sequence data, wherein the data acquisition step length of the initial sample sequence data is not fixed, the sequence length and the data acquisition step length between the initial sample sequence data are different, data reconstruction processing is conducted on each initial sample sequence data to obtain target sample sequence data corresponding to the initial sample sequence data, the data acquisition step length of the target sample sequence data is fixed, the sequence length and the data acquisition step length between the target sample sequence data are the same, and training is conducted on an initial neural network model based on the target sample sequence data to obtain a target neural network model. In a second aspect, an embodiment of the present invention provides a data reconstruction apparatus, including: The system comprises a data acquisition module, a data processing module and a model training module, wherein the data acquisition module is used for acquiring various initial sample sequence data, the data acquisition step length of the initial sample sequence data is not fixed, the sequence length and the data acquisition step length between the initial sample sequence data are different, the data processing module is used for respectively carrying out data reconstruction processing on each initial sample sequence data to obtain target sample sequence data corresponding to the initial sample sequence data, the data acquisition step length of the target sample sequence data is fixed, the sequence length and the data acquisition step length between the target sample sequence data are the same, and the model training module is used for training an initial neural network model based on the target sample sequence data to obtain a target neural network model. In a third aspect, an embodiment of the present invention provides a c