Solid rocket engine time series parameter prediction method and prediction system based on deep learning

Through deep learning methods and ConvNeXt model, the time series parameter characteristics of solid rocket engines are extracted, and the problem of high cost efficiency in thrust prediction of solid rocket engines is solved, fast and accurate thrust prediction is achieved, experimental costs are reduced, and solid rocket engines are reduced. The development of the company provides data support and guidance.

CN117763933BActive Publication Date: 2024-11-12SHENYANG AEROSPACE UNIVERSITY
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202310173394.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-02-28
Publication Date
2024-11-12
Estimated Expiration
2043-02-28

AI Technical Summary

Technical Problem

The prior art has problems of high cost and low efficiency in solid rocket engine thrust prediction, especially the ground thrust test is irreversible and the calculation accuracy is not high enough, making it difficult to establish an accurate simulation model.

Method used

Using deep learning-based methods, data is collected through ground interview tests, deep neural network model ConvNeXt is established, time series parameter data is preprocessed and trained, and convolutional layer, LN layer, ConvNeXt module, downsampling layer and pooling layer are used. Extract features for prediction.

Benefits of technology

It realizes fast and accurate time series parameter prediction, reduces experimental costs, avoids repeated testing, and provides low-cost and efficient solid rocket engine performance evaluation tools, suitable for all stages of development and production.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117763933B_ABST
    Figure CN117763933B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for predicting time series parameters of solid rocket engines based on deep learning. The prediction method includes: collecting multiple groups of original time series parameter data of solid rocket engines under different working conditions and performing data preprocessing to obtain sample data; establishing a deep neural network model; dividing the preprocessed sample data into training samples and test samples, and inputting the training samples into the deep neural network model for training; inputting the test samples into the trained deep neural network model to obtain the prediction results of the time series parameter data of solid rocket engines. The prediction method and prediction system can correct the data obtained from the ground thrust test experiment, can supplement part of the data lost in the experiment due to unexpected circumstances, avoid repeated test experiments, reduce experimental costs, and thus evaluate the performance of solid engines more efficiently and at a low cost.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of computers, artificial intelligence and parameter identification, and in particular to a time series parameter prediction method and prediction system for a solid rocket engine based on deep learning. Background Art

[0002] Solid rocket engines have the characteristics of simple structure, compactness, high reliability, and high specific impulse. They can meet the engine requirements of various tactical missiles to the greatest extent, and thus have become the first choice for missile weapon power systems. The working process of solid rocket engines is generally divided into four stages: ascent stage, working stage, descent stage, and aftereffect stage, and its design process involves multiple theoretical disciplines.

[0003] The performance parameters of the engine are indicators that reflect and evaluate the engine performance from various angles, among which the main ones are thrust, total impulse and specific impulse. Since the solid rocket engine is composed of a grain, a nozzle, an ignition mechanism and a combustion chamber, the thrust of the solid rocket is affected by many parameters such as the propellant composition, the grain structure size, and the nozzle structure size. Therefore, thrust, as one of the main parameters of the solid rocket engine, needs to be focused on. There are two methods to measure the thrust of the solid rocket engine. One is to collect the thrust of the solid rocket engine through a ground thrust test. The ground thrust test experiment of the solid rocket engine is an irreversible, high-cost, and long-period experiment. To accurately measure the thrust, the designer needs to have certain physical expertise and requires a lot of workload. The other is to estimate the thrust of the solid rocket engine through numerical simulation. This method requires modeling and analysis of the solid rocket engine. However, the modeling design of the solid rocket engine involves a large number of nonlinear relationships and physical equations. It is difficult for designers to establish an accurate simulation model, and the calculation accuracy is often not high enough. Therefore, it is necessary to design a low-cost and more efficient method to predict the thrust of the solid rocket engine.

[0004] With the rapid development of artificial intelligence, deep learning technology (deep neural network) has made great breakthroughs in image recognition, user preference matching, natural language processing and other fields. Deep learning technology has strong versatility, and many research topics or projects have achieved great success through deep learning technology. When faced with a large number of data samples, compared with other methods, deep learning can automatically extract important features of data samples for learning prediction, which greatly improves work efficiency while ensuring accuracy. Deep learning can extract the thrust characteristics of solid rocket engines through its powerful feature extraction ability and find the connection between its internal data, thereby efficiently solving the problem of difficult prediction of solid rocket engine thrust, and deep learning has low requirements for relevant physical expertise in solid rockets. Therefore, deep learning has a good application prospect in solving practical engineering problems.

[0005] Therefore, how to apply deep learning to the prediction of time series parameters of solid rocket engines (such as thrust prediction) has become an urgent problem to be solved. Summary of the invention

[0006] The purpose of the present invention is to provide a solid rocket engine time series parameter prediction method and prediction system based on deep learning to solve the problems of high cost and low efficiency in existing time series parameter prediction methods.

[0007] On the one hand, the present invention provides a method for predicting time series parameters of solid rocket engines based on deep learning, comprising:

[0008] S1: using a solid rocket ground test run to collect multiple groups of original time series parameter data of a solid rocket engine under different working conditions, and performing data preprocessing on the original time series parameter data to obtain sample data;

[0009] S2: Establish a deep neural network model ConvNeXt;

[0010] S3: Divide the preprocessed sample data into training samples and test samples, and input the training samples into the deep neural network model ConvNeXt for model training;

[0011] S4, inputting the test sample into the trained deep neural network model to obtain the prediction result of the solid rocket engine time series parameter data.

[0012] Preferably, in S1, the data preprocessing includes:

[0013] S11: In the original time series parameter data with a data length of L, the first N points of each data are taken as the main focus segment, and the last LN points are taken as the non-main focus segment;

[0014] S12, downsampling the non-primary focus segment, and then splicing the downsampled non-primary focus segment data with the primary focus segment into new solid rocket engine time series parameter data;

[0015] S13: Slice the new solid rocket engine time series parameter data to generate samples required by the neural network.

[0016] Further preferably, in S13, during the slicing process, the time window size in each sample is Li, the prediction time length is Lt, the sample length is Li+Lt, the slicing interval is S, and the samples after slicing are (1, Li+Lt), (1+S, Li+Lt+S),…, (1+nS, Li+Lt+nS).

[0017] Further preferably, in S3, the model training includes:

[0018] S31: Initialize the parameters of the deep neural network model;

[0019] S32: Train and optimize the deep neural network model using the training samples;

[0020] S321: Perform network training of the deep neural network model ConvNeXt using a back propagation algorithm, calculate errors, and optimize and adjust parameters of the entire neural network;

[0021] S332, after completing one training, execute steps S321-S322 multiple times in a loop. After each loop is completed, determine whether the maximum number of model training times has been reached. If so, complete all model training.

[0022] Further preferably, in S321, the loss function used to calculate the error is the mean square error MSE, and the formula of the mean square error is shown in formula (1):

[0023]

[0024] Among them, MSE is the mean square error, is the predicted time series parameter data value of the model, y (i) represents the real time series parameter data value obtained from the ground test, and m represents the number of predictions.

[0025] Further preferably, during the model training process, the first layer learning rate of the deep neural network model is 0.001 and the training rounds are 200; the second layer learning rate is 0.0001 and the training rounds are 1000, and the model converges after the maximum number of model training times is reached.

[0026] Further preferably, in the model training, the ConvNeXt model architecture used includes a convolution layer, an LN layer, a ConvNeXt module, a downsampling layer and a pooling layer, wherein the convolution layer contains a filter of size 4, number 16 and stride 4; there are 4 ConvNeXt modules, namely A, B, C, and D; there are 3 downsampling layers, which are respectively located after ConvNeXt modules A, B, and C; the pooling layer is used for global average pooling and to replace the fully connected layer. The data extracted by the convolution layer is processed by the LN layer, then by the ConvNeXt module A, and then by the downsampling layer, and modules B and C are repeated in sequence. Finally, the extracted features are further extracted by module D, processed by the LN layer, and finally output the predicted values ​​of the time series parameters of the solid rocket engine through global average pooling.

[0027] Further preferably, the ConvNeXt module consists of a Depthwise Conv layer, an LN layer, a convolution layer and a regularization layer, wherein the Depthwise Conv layer contains Depthwise Convs of size 7, stride 1, and numbers 24, 48, 96, and 192 respectively, the number of convolution kernels in the first layer of the convolution layer is 24, 48, 96, and 192, the number of convolution kernels in the second layer is 96, 192, 384, and 768, and the size and stride of the convolution kernels of the two layers are both 1.

[0028] Further preferably, the downsampling layer consists of an LN layer and a convolutional layer, the convolutional kernel size in the convolutional layer is 2, the numbers are 24, 48, and 96 respectively, and the step size is 2.

[0029] The present invention also provides a solid rocket engine time series parameter prediction system based on deep learning, which is used to execute the above prediction method.

[0030] The prediction method and prediction system provided by the present invention have the following beneficial effects:

[0031] (1) Overcoming the limitations of traditional solid rocket engine ground tests, deep learning methods are used to predict time series parameters. Data is put into the model, and the model performs convolution, pooling, and regularization operations on the time series parameters to extract features. Finally, all features are put together for analysis and prediction, so that the characteristics of time series parameters can be fully utilized to predict time series parameters for a period of time in the future. The method of using deep learning methods to predict time series parameters is fast and accurate, and has become an effective tool for solid rocket engine performance evaluation;

[0032] (2) Prediction of solid rocket engine time series parameters based on deep learning methods is an emerging method that can supplement some of the time series parameter data lost in the experiment due to unexpected situations, avoid repeated test experiments, and reduce experimental costs, thereby evaluating the performance of solid engines more efficiently and at a lower cost, providing data support and directional guidance for the pre-research, prototype, test and batch production stages of solid rocket engine development, finalization and production.

[0033] (3) A deep neural network (ConvNeXt) was applied and modified to make it more suitable for the prediction of time series parameters of solid rocket engines, and it has superiority in the prediction of time series parameters of solid rocket engines. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of a method for predicting time series parameters of solid rocket engines based on deep learning provided by the present invention.

[0035] Figure 2This is a model structure diagram of the method for predicting time series parameters of solid rocket engines based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0036] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0037] like Figure 1 , Figure 2 As shown in FIG. 1 , a method for predicting time series parameters of solid rocket engines based on deep learning includes the following steps (taking thrust time series data as a time series parameter for illustration):

[0038] S1: using a solid rocket ground test run to collect multiple groups of original thrust time series data of a solid rocket engine under different working conditions, and performing data preprocessing on the original thrust time series data to obtain sample data;

[0039] Wherein, the data preprocessing includes:

[0040] S11: In the original thrust time series data with a data length of L, the first N points of each data are taken as the main focus segment (the part with large data fluctuation and large value), and the last LN points are taken as the non-main focus segment (the part with small data fluctuation and small value);

[0041] S12, downsampling the non-primary segment, and then splicing the downsampled non-primary segment data with the primary segment into new solid rocket engine thrust time series data;

[0042] S13: Slice the new solid rocket engine thrust time series data to generate samples required by the neural network.

[0043] Among them, during the slicing process, the time window size in each sample is Li, the prediction time length is Lt, the sample length is Li+Lt, the slicing interval is S, and the samples after slicing are (1, Li+Lt), (1+S, Li+Lt+S),…, (1+nS, Li+Lt+nS).

[0044] S2: Establish a deep neural network model ConvNeXt;

[0045] S3: Divide the preprocessed sample data into training samples and test samples, and input the training samples into the deep neural network model ConvNeXt for model training;

[0046] Wherein, the model training includes:

[0047] S31: Initialize the parameters of the deep neural network model;

[0048] S32: Using the training sample to train and optimize the deep neural network model, including:

[0049] S321: Perform network training of the deep neural network model ConvNeXt using a back propagation algorithm, calculate errors, and optimize and adjust parameters of the entire neural network;

[0050] S332, after completing one training, execute steps S321-S322 multiple times in a loop. After each loop is completed, determine whether the maximum number of model training times has been reached. If so, complete all model training.

[0051] In S321, the loss function used to calculate the error is the mean square error MSE, and the formula for the mean square error is shown in formula (1):

[0052]

[0053] Among them, MSE is the mean square error, is the predicted thrust of the model, y (i) represents the actual thrust obtained from the ground test, and m represents the number of predictions;

[0054] During the model training process, the first layer learning rate of the deep neural network model was 0.001 and the number of training rounds was 200; the second layer learning rate was 0.0001 and the number of training rounds was 1000. The model converged after the maximum number of model training times was reached.

[0055] Among them, Figure 2 As shown in the figure, the ConvNeXt model architecture used in model training includes a convolution layer, a regularization layer (LN layer), a ConvNeXt module, a downsampling layer and a pooling layer, wherein the convolution layer contains a filter of size 4, number 16 and stride 4; there are 4 ConvNeXt modules, namely A, B, C and D; there are 3 downsampling layers, which are respectively located after ConvNeXt modules A, B and C; the pooling layer is used for global average pooling and to replace the fully connected layer. The data extracted by the convolution layer is processed by the LN layer, then by the ConvNeXt module A, and then by the downsampling layer, and modules B and C are repeated in sequence. Finally, the extracted features are further extracted by module D, processed by the LN layer, and finally the thrust prediction value of the solid rocket engine is output by global average pooling.

[0056] Among them, Figure 2As shown, the ConvNeXt module consists of a Depthwise Conv layer, a LN layer, a convolution layer and a regularization layer, wherein the Depthwise Conv layer contains Depthwise Convs of size 7, stride 1, and the numbers are 24, 48, 96, and 192 respectively. The number of convolution kernels in the first layer of the convolution layer is 24, 48, 96, and 192, and the number of convolution kernels in the second layer is 96, 192, 384, and 768, and the size and stride of the convolution kernels of the two layers are both 1.

[0057] The downsampling layer consists of an LN layer and a convolution layer. The convolution kernel size in the convolution layer is 2, the number is 24, 48, and 96 respectively, and the step size is 2.

[0058] S4, inputting the test sample into the trained deep neural network model to obtain the solid rocket engine thrust timing data prediction result.

[0059] Integrating complex data from different sensors into feature learning can achieve data fusion, enrich the information of the system, and improve the performance of the general model. Generally speaking, the original thrust test data of solid rocket engines can be made more suitable for input into deep neural networks after data preprocessing. During the operation process, only the test samples are input into the proposed parameter recognition network to obtain the prediction results, and the thrust timing prediction scheme is automatically implemented, which greatly improves the applicability of the model in actual industrial scenarios.

[0060] The solid rocket engine thrust-time series data prediction method and prediction system based on convolutional neural network proposed in this application use multiple groups of real thrust-time series data obtained from ground thrust test experiments as samples, train the samples through convolutional neural network, obtain convolutional neural network model, and thus obtain thrust-time series data at future moments. Comparing the thrust value predicted by the present invention with the real value, the thrust value obtained by deep learning CNN network method, deep learning DNN network method, deep learning CNN network method, and deep learning LSTM network method, it can be known that the present invention can more efficiently and accurately correct the data obtained from the ground thrust test experiment, and can also complete part of the thrust data lost in the experiment due to unexpected circumstances, avoid repeating thrust test experiments, and reduce experimental costs, so as to more efficiently and low-costly evaluate the performance of solid engines, and provide data support and directional guidance for the pre-research, prototype, test and batch production stages of solid rocket engine research and development and finalization production.

[0061] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting time series parameters of solid rocket engines based on deep learning, characterized in that: include: S1: using a solid rocket ground test run to collect multiple groups of original time series parameter data of a solid rocket engine under different working conditions, and performing data preprocessing on the original time series parameter data to obtain sample data; S2: Establish a deep neural network model ConvNeXt; S3: Divide the preprocessed sample data into training samples and test samples, and input the training samples into the deep neural network model ConvNeXt for model training; S4: inputting the test sample into the trained deep neural network model to obtain the prediction result of the solid rocket engine time series parameter data; In S1, the data preprocessing includes: S11: In the original time series parameter data with a data length of L, the first N points of each data are taken as the main focus segment, and the last LN points are taken as the non-main focus segment; S12, downsampling the non-primary focus segment, and then splicing the downsampled non-primary focus segment data with the primary focus segment into new solid rocket engine time series parameter data; S13: Slice the new solid rocket engine time series parameter data to generate samples required by the neural network.

2. The method for predicting time series parameters of solid rocket engines based on deep learning according to claim 1, characterized in that: In S13, during the slicing process, the time window size in each sample is Li, the prediction time length is Lt, the sample length is Li+Lt, the slicing interval is S, and the samples after slicing are (1, Li+Lt), (1+S, Li+Lt+S),…, (1+nS, Li+Lt+nS).

3. The method for predicting time series parameters of solid rocket engines based on deep learning according to claim 1, characterized in that: In S3, the model training includes: S31: Initialize the parameters of the deep neural network model; S32: Train and optimize the deep neural network model using the training samples; S321: Perform network training of the deep neural network model ConvNeXt using a back propagation algorithm, calculate errors, and optimize and adjust parameters of the entire neural network; S332, after completing one training, execute steps S321-S322 multiple times in a loop. After each loop is completed, determine whether the maximum number of model training times has been reached. If so, complete all model training.

4. The method for predicting time series parameters of solid rocket engines based on deep learning according to claim 3 is characterized in that: In S321, the loss function used to calculate the error is the mean square error MSE, and the formula for the mean square error is shown in formula (1): Among them, MSE is the mean square error, is the predicted time series parameter data value of the model, y (i) represents the real time series parameter data value obtained from the ground test, and m represents the number of predictions.

5. The method for predicting time series parameters of solid rocket engines based on deep learning according to claim 1, characterized in that: During the model training process, the first layer learning rate of the deep neural network model was 0.001 and the number of training rounds was 200; the second layer learning rate was 0.0001 and the number of training rounds was 1000. The model converged after the maximum number of model training times was reached.

6. The method for predicting time series parameters of solid rocket engines based on deep learning according to claim 1, characterized in that: In model training, the ConvNeXt model architecture used includes a convolution layer, an LN layer, a ConvNeXt module, a downsampling layer and a pooling layer, wherein the convolution layer contains a filter of size 4, number 16 and stride 4; there are 4 ConvNeXt modules, namely A, B, C and D; there are 3 downsampling layers, which are respectively located after ConvNeXt modules A, B and C; the pooling layer is used for global average pooling and to replace the fully connected layer. The data extracted by the convolution layer is processed by the LN layer, then by the ConvNeXt module A, and then by the downsampling layer, and modules B and C are repeated in sequence. Finally, the extracted features are further extracted by module D, processed by the LN layer, and finally output the predicted values ​​of the time series parameters of the solid rocket engine through global average pooling.

7. The method for predicting time series parameters of solid rocket engines based on deep learning according to claim 6, characterized in that: The ConvNeXt module consists of a Depthwise Conv layer, an LN layer, a convolution layer and a regularization layer, wherein the Depthwise Conv layer contains Depthwise Convs of size 7, stride 1, and numbers 24, 48, 96, and 192 respectively. The number of convolution kernels in the first layer of the convolution layer is 24, 48, 96, and 192, and the number of convolution kernels in the second layer is 96, 192, 384, and 768. The size and stride of the convolution kernels in the two layers are both 1.

8. The method for predicting time series parameters of solid rocket engines based on deep learning according to claim 6, characterized in that: The downsampling layer consists of an LN layer and a convolution layer. The convolution kernel size in the convolution layer is 2, the number is 24, 48, and 96 respectively, and the step size is 2.

9. A solid rocket engine time series parameter prediction system based on deep learning, used to execute the prediction method described in any one of claims 1-8.

Citation Information

Patent Citations

  • A method for improving the number of parameters of deep convolution neural network

    CN109272107A

  • Aero-engine fault diagnosis method based on data driving

    CN111175054A

  • Fault diagnosis method, device and equipment for industrial system and storage medium

    CN111340238A

  • Tunneling parameter prediction method of hard rock tunnel boring machine based on deep learning

    CN112163316A

  • Aircraft trajectory prediction method based on long short-term memory network

    CN114048889A