CN-121118346-B - Temperature field prediction method and medium

CN121118346BCN 121118346 BCN121118346 BCN 121118346BCN-121118346-B

Abstract

The application provides a temperature field prediction method and medium, wherein the temperature field prediction method comprises the steps of obtaining a temperature field data set of a rocket engine, inputting the temperature field data set into a pre-trained temperature field prediction model to obtain temperature field prediction time sequence data of a prediction time period, the temperature field prediction model comprises a time sequence compression module, a time sequence decompression module and a time sequence output module which are sequentially connected, the time sequence compression module is configured to compress the temperature field time sequence data to obtain temperature field compression time sequence data, the time sequence prediction module is configured to predict according to working condition data and the temperature field compression time sequence data to obtain compression prediction data of each prediction time in at least one prediction time, and the time sequence decompression module is configured to decompress the compression prediction data of each prediction time to obtain the temperature field prediction data of each prediction time. Therefore, the rapid and accurate prediction of the rocket combustion chamber temperature field can be realized.

Inventors

ZHANG YANG
ZHU WEIYI
ZHU XIAO
ZHANG BO

Assignees

江苏大学

Dates

Publication Date: 20260508
Application Date: 20250804

Claims (6)

1. A method of temperature field prediction, comprising: acquiring a temperature field data set of the rocket engine, wherein the temperature field data set comprises working condition data and temperature field time sequence data of an acquisition time period; Inputting the temperature field data set into a pre-trained temperature field prediction model to obtain temperature field prediction time sequence data of a prediction time period, wherein the prediction time period is after the acquisition time period, and the prediction time period comprises at least one prediction moment; The temperature field prediction model comprises a time sequence compression module, a time sequence decompression module and a time sequence output module which are sequentially connected, wherein the time sequence compression module is used for compressing the temperature field time sequence data to obtain temperature field compression time sequence data, the time sequence prediction module is used for predicting according to the working condition data and the temperature field compression time sequence data to obtain compression prediction data of each prediction moment in the at least one prediction moment, the time sequence decompression module is used for decompressing the compression prediction data of each prediction moment to obtain temperature field prediction data of each prediction moment, and the time sequence output module is used for integrating the temperature field prediction data of each prediction moment to form temperature field prediction time sequence data of the prediction time period; The at least one prediction time comprises an earliest prediction time, and the time sequence prediction module is configured to predict according to the working condition data and the temperature field compression time sequence data to obtain compression prediction data of the earliest prediction time; The at least one prediction time further comprises one or more subsequent prediction times after the earliest prediction time, and the time sequence prediction module is further configured to update temperature field compression time sequence data for predicting the previous prediction time according to compression prediction data of the previous prediction time of the subsequent prediction time for each of the one or more subsequent prediction times, and predict based on the working condition data and the updated temperature field compression time sequence data to obtain compression prediction data of the subsequent prediction time; the time sequence prediction module comprises a noise adding layer, a diffusion prediction layer and a time sequence shifting layer, wherein the earliest prediction moment is as follows: The noise adding layer is configured to add Gaussian white noise to the temperature field compression time sequence data, and fuse the noise-added temperature field compression time sequence data with the working condition data to obtain temperature field noise working condition time sequence data; The diffusion prediction layer is configured to perform denoising prediction on the temperature field noise working condition time sequence data to obtain compression prediction data at the earliest prediction moment; the time sequence shifting layer is configured to delete the temperature field compression data at the earliest time in the temperature field compression time sequence data and add the compression prediction data at the earliest prediction time to the temperature field compression time sequence data; For each of said subsequent predicted instants: The noise adding layer is further configured to add white gaussian noise to temperature field compression time sequence data updated based on compression prediction data at a previous prediction time of the subsequent prediction time, and fuse the temperature field compression time sequence data after noise adding with the working condition data to obtain updated temperature field noise working condition time sequence data; the diffusion prediction layer is further configured to perform denoising prediction on the updated temperature field noise working condition time sequence data to obtain compression prediction data of the subsequent prediction moment; the timing shift layer is further configured to delete the earliest time temperature field compression data in the updated temperature field compression timing data, and to add the compression prediction data of the subsequent prediction time to the updated temperature field compression timing data.
2. The temperature field prediction method according to claim 1, wherein the time sequence decompression module comprises an intermediate bridge unit, at least two up-sampling units connected in series, and a convolution unit, each up-sampling unit comprising at least one third feature extraction layer, a second residual connection layer, and a first feature reconstruction layer connected in sequence, wherein the first feature reconstruction layer comprises a normalization layer, an activation layer, and a deconvolution layer.
3. The temperature field prediction method according to claim 2, wherein the other up-sampling units except the last up-sampling unit of the at least two up-sampling units connected in series further comprises a feature calibration layer and a third residual connection layer, wherein the feature calibration layer comprises a channel space attention layer, a convolution layer, a normalization layer and an activation layer which are sequentially connected, and the third residual connection layer connects the output of the first feature reconstruction layer with the output of the feature calibration layer.
4. The temperature field prediction method according to claim 1, further comprising a training step of a temperature field prediction model, the training step comprising a joint training step and a diffusion prediction training step; The combined training step comprises the steps of freezing a time sequence prediction module and a time sequence output module of a temperature field prediction model, and training a time sequence compression module and a time sequence decompression module of the temperature field prediction model by utilizing historical temperature field data of at least one historical moment; The diffusion prediction training step comprises the steps of thawing a time sequence prediction module of a temperature field prediction model, freezing parameters of a time sequence compression module and a time sequence decompression module, and denoising and prediction training are carried out on the time sequence prediction module of the temperature field prediction model by utilizing at least one historical temperature field time sequence data and corresponding working condition data.
5. The method of temperature field prediction according to claim 4, wherein the training step further comprises: Dividing a data set of different working condition tasks into a support set and a query set, wherein each element of the data set comprises working condition data and corresponding actual temperature field time sequence data; An internal circulation step of completing the joint training step and the diffusion prediction training step for each element in the support set to obtain a temperature field prediction model for completing internal circulation training; And an outer circulation step of traversing the query set, calculating query loss based on the temperature field prediction model completing the inner circulation training, and if the query loss is not converged, adjusting parameters of the temperature field prediction model completing the inner circulation training, returning to the inner circulation step, and performing inner circulation training on the adjusted temperature field prediction model until the query loss is converged.
6. A computer readable storage medium having stored thereon instructions which, when executed by a processor, implement the steps of the temperature field prediction method of any of claims 1-5.

Description

Temperature field prediction method and medium Technical Field The application relates to the technical field of intersection of thermodynamic simulation and machine learning, in particular to a temperature field prediction method and a medium. Background Accurate prediction of the rocket engine combustion chamber temperature field is a key link of design optimization of the aerospace propulsion system. In addition to calculations using traditional numerical simulations, modeling and predicting the temperature field of rocket engine combustion chambers by projection-based reduced order models has also received significant attention. However, the conventional method for reducing the order is an intrinsic orthogonal decomposition method of linear dimension reduction, and for an engine temperature field with strong nonlinear characteristics, the method has larger errors between the reduced order and the predicted result and actual data, because the method has the problems that 1, the modal cut-off errors are that the cut-off low-energy modes in the conventional projection-based reduced order model possibly discard nonlinear characteristics such as flame front mutation and the like due to strong nonlinear correlation of the combustion process, 2, the working condition coverage is insufficient, the working condition related to training data has good effect, and model failure can be caused for data out of range, and 3, the training data is high in requirement, and a large amount of engine temperature field data is difficult to obtain for the temperature field of a rocket engine combustion chamber, whether test data or numerical simulation. Disclosure of Invention In order to solve the problems, the application provides a temperature field prediction method and a medium, which can realize rapid and accurate prediction of a rocket combustion chamber temperature field. The temperature field prediction method comprises the steps of obtaining a temperature field data set of a rocket engine, inputting the temperature field data set into a pre-trained temperature field prediction model to obtain temperature field prediction time sequence data of a prediction time period, wherein the prediction time period comprises at least one prediction time after the acquisition time period, the temperature field prediction model comprises a time sequence compression module, a time sequence decompression module and a time sequence output module which are connected in sequence, the time sequence compression module is configured to compress the temperature field time sequence data to obtain temperature field compression time sequence data, the time sequence prediction module is configured to predict according to the working condition data and the temperature field compression time sequence data to obtain compression prediction data of each prediction time in at least one prediction time, the time sequence decompression module is configured to decompress the compression prediction data of each prediction time to obtain temperature field prediction data of each prediction time, and the time sequence output module is configured to integrate the temperature field prediction data of each prediction time period to form the temperature field prediction time sequence data of each prediction time period. In a second aspect, a computer readable storage medium is provided, on which instructions are stored which, when executed by a processor, implement the steps of a temperature field prediction method as provided by any of the embodiments of the present application. In summary, the temperature field prediction method and the computer readable storage medium provided by the application have the advantages that the temperature field data set is input into the temperature field prediction model to obtain the temperature field prediction time sequence data of a prediction time period, the temperature field prediction module comprises a time sequence compression module, a time sequence decompression module and a time sequence output module which are sequentially connected, so that a compression and decompression structure is introduced into the model, and the high-dimensional temperature field data is subjected to dimension reduction compression and decompression reduction, thereby remarkably reducing the storage space requirement and the calculation load of the subsequent time sequence prediction module. And the temperature field time sequence data after compression is combined with the working condition data to predict the temperature field data, so that the aim of rapidly and accurately predicting the temperature field data at a plurality of prediction moments under the working condition can be fulfilled, and the problem that the traditional method is easy to fit when the training sample is insufficient is solved. Drawings In order to more clearly illustrate the embodiments of the application or the technical solutions in the prior art, the drawing