An aero-engine digital twin parameter prediction method based on CEEMDAN-informer

Abstract

The application belongs to the field of digital twinning technology and aero-engine parameter prediction, and discloses an aero-engine digital twinning parameter prediction method based on CEEMDAN-Informer, which uses the CEEMDAN algorithm to decompose the gas path parameter data in the aero-engine twinning model into multiple components with different time scale characteristics, effectively solving the problem of poor nonlinear data trend tracking ability of the existing model. At the same time, the fuzzy entropy is used for grouping and fusing each component, reducing the input dimension of the encoder and effectively improving the model training speed and prediction efficiency. Compared with the existing method, the aero-engine parameter prediction method based on CEEMDAN-Informer and digital twinning technology can effectively improve the parameter prediction accuracy and reduce the prediction error.

Description

Technical Field

[0001] This invention belongs to the fields of digital twin technology and aero-engine parameter prediction, specifically a method for predicting aero-engine digital twin parameters based on CEEMDAN-Informer. Background Technology

[0002] Aero-engine digital twin is a technology that reflects the actual operating state of an aero-engine in a computer through physical modeling, sensor communication, data mapping, and dynamic simulation. By observing the parameter changes of the engine digital twin, the operating state and performance changes of the aero-engine can be monitored and evaluated. In order to achieve early warning of abnormal engine conditions, it is necessary to predict the parameters in the aero-engine digital twin, which places high demands on the prediction accuracy of the aero-engine parameter prediction model.

[0003] As a highly complex mechanical-thermodynamic system, most parameters in aero-engines exhibit strong nonlinearity and poor regularity, making it difficult to directly detect their trends. Existing mainstream parameter prediction models mostly predict directly from the raw parameter data, failing to fully utilize the regularity information at different time scales contained in the parameter sequence. Therefore, they cannot guarantee prediction accuracy when the prediction time is long or the parameters change; see Yang Haimin et al.'s review of time series prediction methods for details. How to discover trends from nonlinear parameter sequences and improve the accuracy of parameter prediction remains a challenge. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide a method for predicting aero-engine digital twin parameters based on CEEMDAN-Informer. The fully adaptive ensemble empirical mode decomposition (CEEMDAN) decomposes engine parameters into multiple components with different time-scale characteristics. These components typically exhibit significant differences in trend and periodicity. The multi-head attention mechanism in Informer can be used to analyze the variation patterns of these components, achieving accurate tracking of parameter trends. The proposed CEEMDAN-Informer prediction model can extract multiple time-scale features from aero-engine digital twin parameters to the greatest extent possible, thereby achieving higher parameter prediction accuracy.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for predicting digital twin parameters of aero-engines based on CEEMDAN-Informer, comprising the following steps:

[0007] S1: Establish digital twin models of aero-engine components, including five components: air intake, turbine, compressor, combustion chamber, and tail nozzle;

[0008] S2: Collect the environmental operating condition parameters of the current real engine as the preset conditions of the engine digital twin model, use the engine digital twin model to simulate the engine operation process, and obtain simulation data of the flight parameters of the aero-engine.

[0009] S3: During the actual engine operation, sensor data is collected in real time to obtain the measured values ​​of engine flight parameters. The measured values ​​of flight parameters are used to correct the errors in the simulation data to obtain the corrected engine flight parameter data.

[0010] S4: Standardize the engine flight parameter correction data described in S3 to reduce the distribution differences between data, enhance the temporal stability of the data, and make the model's input distribution more stable.

[0011] S5: The CEEMDAN algorithm is used to decompose the standardized engine flight parameter correction data to obtain multiple intrinsic components with different time scale characteristics and a residual component. The trends and periodicity of these components are usually quite different, which can be used to comprehensively analyze the variation law of the sequence.

[0012] S6: Use fuzzy entropy to evaluate the complexity of several components mentioned in S5. Group each component according to the size of its fuzzy entropy and the order of magnitude of its peak value. The components in each group have similar fuzzy entropy and the same order of magnitude. Fuse the components in each group to obtain several fused components. It can be guaranteed that the number of fused components is not greater than the number of original components.

[0013] S7: A sequence prediction model is established based on the principles of the Informer model, with each fused component serving as input data. Aero-engine data is essentially a time series; the first 60% of the corrected engine flight parameters are used as the training set, the middle 20% as the validation set, and the last 20% as the test set.

[0014] S8: The training data enters the encoder after input embedding and position encoding. Feature extraction and attention calculation are performed in the encoder. The data then enters the decoder through a fully connected feedforward network and residual normalization. Input data with time offset also enters the decoder after input embedding. Attention calculation is performed in the masked multi-head attention module. The relationship mapping between input and output is completed through decoding. Finally, the predicted value is output through a fully connected layer.

[0015] S8: Calculate the error between the predicted value and the actual value, calculate the loss function (Loss value), and repeat the model training process in S8 until the training termination condition is met.

[0016] S9: Use the model trained in S8 to predict the fusion components in S6, obtain the prediction results of the fusion components, and sum the prediction results of the fusion components to obtain the prediction results of the original flight parameters.

[0017] S10: Input the predicted results of each flight parameter into the digital twin of the aero-engine and perform simulation to obtain a digital twin that can reflect the predicted working state of the engine. The predicted values ​​of other parameters can be obtained from the digital twin, thereby realizing the parameter prediction of the aero-engine.

[0018] Furthermore, the standardization process described in S4 is as follows:

[0019]

[0020]

[0021]

[0022] Where X = [x1, x2, ..., x L ] T Correction data for the original engine flight parameters, X′=[x1′,x2′,...,x′ L ] T This refers to the standardized engine flight parameter correction data, where L is the length of the engine flight parameter correction data, and μ is the value of μ. X The mean of the original engine flight parameter correction data X is given. The variance of the original engine flight parameter correction data X.

[0023] Furthermore, the CEEMDAN decomposition steps in S5 are as follows:

[0024] S5-1: Add Gaussian white noise ω to the engine flight parameter correction data s(t). i (t), resulting in a new signal sequence:

[0025]

[0026] Where γ0 is the standard deviation of Gaussian white noise, and n is the number of times white noise is added.

[0027] S5-2: The new signal sequence obtained above EMD decomposition yields the following decomposition form:

[0028]

[0029] Where m is the number of EMD components, IMF1 j(t) represents the j-th EMD component, and res1(t) represents the residual. The first-order CEEMDAN component is obtained from the EMD components.

[0030]

[0031] The first-order modal signal margin is:

[0032]

[0033] S5-3: Add Gaussian white noise ω to the first-order modal signal margin r1(t) i (t), to obtain a new signal sequence

[0034]

[0035] Similar to the steps in S5-2, the signal sequence is processed. EMD decomposition yields the following decomposition form:

[0036]

[0037] The second-order CEEMDAN component is obtained from the EMD component.

[0038]

[0039] The margin of the second-order modal signal is:

[0040]

[0041] S5-4: Repeat step S5-3 until the modal signal margin can no longer be decomposed. At this point, the final decomposition result of the engine flight parameter correction data s(t) is:

[0042]

[0043] Furthermore, the steps for outputting the predicted value in S8 are as follows:

[0044] S8-1: Data embedding and positional encoding are performed on the model input sequence. Data embedding uses convolutional and fully connected networks to map the data into a 512-dimensional feature vector. Positional encoding embeds the temporal information of the sequence into the feature vector, preventing the loss of temporal information due to attention calculations. Positional encoding is calculated using the encoding method in Transformer:

[0045]

[0046]

[0047] Where u represents the dimension index in the input vector, d model The input sequence has the total number of dimensions, and pos represents the position of the calculated input vector within the entire input sequence.

[0048] S8-2: The transformed feature vectors are fed into the encoder for attention calculation. The attention calculation method is as follows:

[0049]

[0050] Where Q,K,V∈R N*d These are the query matrix, key matrix, and value matrix, respectively. N is the length of the input engine flight parameter correction data, d is the hidden dimension in the attention head, and softmax is the normalized exponential function.

[0051] S8-3: The attention calculation results contain some redundancy. Pooling is used to reduce its dimensionality, giving higher weights to the important parts. Each encoding module contains one attention calculation, one pooling operation, and one residual and normalization operation. Multiple layers of encoding modules can be stacked in the encoder. After three layers of encoding modules, the encoder's output vector is obtained.

[0052] S8-4: Constructing a sequence with a time offset:

[0053] X feedDecoder =concat(X) Token ,X Placeholder )

[0054] Among them, X Token X is a known sequence preceding the time point to be predicted. Placeholder This is a zero-padded sequence to be predicted, also known as a placeholder sequence. The sequence with time offset is embedded into the input and then enters the decoder. Attention is calculated in the masked multi-head attention module. The relationship between input and output is mapped through decoding. Finally, the predicted value is output through a fully connected layer.

[0055] Compared with the prior art, the present invention has at least the following beneficial effects:

[0056] This invention proposes a method for predicting aero-engine digital twin parameters based on CEEMDAN-Informer. It uses the CEEMDAN algorithm to decompose the gas path parameter data in the aero-engine twin model into multiple components with different time-scale characteristics, effectively solving the problem of poor trend tracking ability of existing models for nonlinear data. Simultaneously, fuzzy entropy is used to group and fuse the components, reducing the input dimensionality of the encoder and effectively improving model training speed and prediction efficiency. Compared with existing methods, the aero-engine parameter prediction method based on CEEMDAN-Informer and digital twin technology proposed in this invention can effectively improve parameter prediction accuracy and reduce prediction error.

[0057] Additional aspects and advantages of the invention will be set forth in the detailed description which follows, and in part will be obvious from the description or may be learned by specific embodiments thereof. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the present invention.

[0059] Figure 2 The figure shows the simulation results of a specific embodiment of the present invention. Detailed Implementation

[0060] Specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the invention.

[0061] like Figure 1 As shown, the present invention proposes a method comprising the following steps:

[0062] S1: Establish digital twin models of aero-engine components, including five components: air intake, turbine, compressor, combustion chamber, and tail nozzle;

[0063] S2: Collect the environmental operating condition parameters of the current real engine as the preset conditions of the engine digital twin model, use the engine digital twin model to simulate the engine operation process, and obtain simulation data of the flight parameters of the aero-engine.

[0064] S3: Collect sensor data in real time during actual engine operation to obtain measured values ​​of engine flight parameters. Use the measured values ​​of flight parameters to correct errors in the simulation data to obtain corrected engine flight parameter data that reflects the engine's flight status.

[0065] S4: Standardize the engine flight parameter correction data described in S3 to reduce the distribution differences between data, enhance the temporal stability of the data, and make the model's input distribution more stable.

[0066] S5: The CEEMDAN algorithm is used to decompose the standardized engine flight parameter correction data to obtain multiple intrinsic components with different time scale characteristics and a residual component. The trends and periodicity of these components are usually quite different, which can be used to comprehensively analyze the variation law of the sequence.

[0067] S6: Use fuzzy entropy to evaluate the complexity of several components mentioned in S5. Group each component according to the size of its fuzzy entropy and the order of magnitude of its peak value. The components in each group have similar fuzzy entropy and the same order of magnitude. Fuse the components in each group to obtain several fused components. It can be guaranteed that the number of fused components is not greater than the number of original components.

[0068] S7: A sequence prediction model is established based on the principles of the Informer model, with each fused component serving as input data. Aero-engine data is essentially a time series; the first 60% of the corrected engine flight parameters are used as the training set, the middle 20% as the validation set, and the last 20% as the test set.

[0069] S8: The training data enters the encoder after input embedding and position encoding. Feature extraction and attention calculation are performed in the encoder. The data then enters the decoder through a fully connected feedforward network and residual normalization. Input data with time offset also enters the decoder after input embedding. Attention calculation is performed in the masked multi-head attention module. The relationship mapping between input and output is completed through decoding. Finally, the predicted value is output through a fully connected layer.

[0070] S9: Calculate the error between the predicted value and the actual value, calculate the loss function (Loss value), and repeat the model training process in S8 until the training termination condition is met.

[0071] S9: Use the model trained in S9 to predict the fusion components in S6, obtain the prediction results of the fusion components, and sum the prediction results of the fusion components to obtain the prediction results of the original flight parameters.

[0072] S10: Input the predicted results of each flight parameter into the digital twin of the aero-engine and perform simulation to obtain a digital twin that can reflect the predicted working state of the engine. The predicted values ​​of other parameters can be obtained from the digital twin, thereby realizing the parameter prediction of the aero-engine.

[0073] Furthermore, the standardization process described in S4 is as follows:

[0074]

[0075]

[0076]

[0077] Where X = [x1, x2, ..., x L ] T Correction data for the original engine flight parameters, X′=[x1′,x′2,...,x′ L ] T This refers to the standardized engine flight parameter correction data, where L is the length of the engine flight parameter correction data, and μ is the value of μ. X The mean of the original engine flight parameter correction data X is given. The variance of the original engine flight parameter correction data X.

[0078] Furthermore, the CEEMDAN decomposition steps in S5 are as follows:

[0079] S5-1: Add Gaussian white noise ω to the engine flight parameter correction data s(t). i (t), resulting in a new signal sequence:

[0080]

[0081] Where γ0 is the standard deviation of Gaussian white noise, and n is the number of times white noise is added.

[0082] S5-2: The new signal sequence obtained above EMD decomposition yields the following decomposition form:

[0083]

[0084] Where m is the number of EMD components, Let res1(t) be the j-th EMD component and res1(t) be the residual. The first-order CEEMDAN component is obtained from the EMD components.

[0085]

[0086] The first-order modal signal margin is:

[0087]

[0088] S5-3: Add Gaussian white noise ω to the first-order modal signal margin r1(t) i (t), to obtain a new signal sequence

[0089]

[0090] Similar to the steps in S5-2, the signal sequence is processed. EMD decomposition yields the following decomposition form:

[0091]

[0092] The second-order CEEMDAN component is obtained from the EMD component.

[0093]

[0094] The margin of the second-order modal signal is:

[0095]

[0096] S5-4: Repeat step S5-3 until the modal signal margin can no longer be decomposed. At this point, the final decomposition result of the engine flight parameter correction data s(t) is:

[0097]

[0098] Furthermore, the steps for outputting the predicted value in S8 are as follows:

[0099] S8-1: Data embedding and positional encoding are performed on the model input sequence. Data embedding uses convolutional and fully connected networks to map the data into a 512-dimensional feature vector. Positional encoding embeds the temporal information of the sequence into the feature vector, preventing the loss of temporal information due to attention calculations. Positional encoding is calculated using the encoding method in Transformer:

[0100]

[0101] Where u represents the dimension index in the input vector, d model The input sequence has the total number of dimensions, and pos represents the position of the calculated input vector within the entire input sequence.

[0102] S8-2: The transformed feature vectors are fed into the encoder for attention calculation. The attention calculation method is as follows:

[0103]

[0104] Where Q,K,V∈R N*d These are the query matrix, key matrix, and value matrix, respectively. N is the length of the input engine flight parameter correction data, d is the hidden dimension in the attention head, and softmax is the normalized exponential function.

[0105] S8-3: The attention calculation results contain some redundancy. Pooling is used to reduce its dimensionality, giving higher weights to the important parts. Each encoding module contains one attention calculation, one pooling operation, and one residual and normalization operation. Multiple layers of encoding modules can be stacked in the encoder. After three layers of encoding modules, the encoder's output vector is obtained.

[0106] S8-4: Constructing a sequence with a time offset:

[0107] X feedDecoder =concat(X) Token ,X Placeholder )

[0108] Among them, X Token X is a known sequence preceding the time point to be predicted. Placeholder This is a zero-padded sequence to be predicted, also known as a placeholder sequence. The sequence with time offset is embedded into the input and then enters the decoder. Attention is calculated in the masked multi-head attention module. The relationship between input and output is mapped through decoding. Finally, the predicted value is output through a fully connected layer.

[0109] Example Experiment 1:

[0110] The experimental data in this example comes from maintenance test data of an engine research institute. The dataset length is 2850 seconds, and the selected prediction parameter is N1r. The dataset is divided into a 60% test set, a 20% validation set, and a 20% test set. The error evaluation metrics used are MAE, RMSE, and MAPE. N1r was predicted using the original Informer model and the improved CEEMDAN-Informer model in this invention at different prediction time steps. The prediction errors are shown in Table 1.

[0111] Table 1

[0112]

[0113] As can be seen from the error comparison in Table 1, the improved CEEMDAN-Informer model has a smaller prediction error than the original model under different prediction time steps, and can achieve more accurate engine parameter prediction.

[0114] Prediction curve as follows Figure 2 As shown, according to Figure 2 Similarly, it can be concluded that the prediction error between the CEEMDAN-Informer model and the true value is smaller than the prediction error of the original model.

[0115] The embodiments of the present invention have been described above with reference to the accompanying drawings, but the present invention is not limited to the specific embodiments described above. The above embodiments are merely illustrative and instructive, and not restrictive. Those skilled in the art can make various modifications based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and these modifications are all within the scope of protection of the present invention.

Claims

1. A method for predicting digital twin parameters of aero-engines based on CEEMDAN-Informer, characterized in that, The steps are as follows: S1: Establish digital twin models of aero-engine components, including five components: air intake, turbine, compressor, combustion chamber, and tail nozzle; S2: Collect the environmental operating condition parameters of the current real engine as the preset conditions of the aero-engine component-level digital twin model, use the aero-engine component-level digital twin model to simulate the engine operation process, and obtain the simulation data of the aero-engine flight parameters. S3: During the actual engine operation, sensor data is collected in real time to obtain the measured values ​​of engine flight parameters. The measured values ​​of flight parameters are used to correct the errors in the simulation data to obtain the corrected engine flight parameter data. S4: Standardize the engine flight parameter correction data from step S3; The standardized processing method is as follows: Where X = [x1, x2, ..., x L ] T Correction data for the original engine flight parameters, X′=[x′1,x′2,...,x′ L ] T This refers to the standardized engine flight parameter correction data, where L is the length of the engine flight parameter correction data, and μ is the value of μ. X The mean of the engine flight parameter correction data X, The variance of the data X used to correct engine flight parameters; S5: The CEEMDAN algorithm is used to decompose the standardized engine flight parameter correction data to obtain multiple intrinsic components with different time scale characteristics and a residual component. The trend and periodicity among these components are used to comprehensively analyze the variation law of the flight parameter record value sequence. The CEEMDAN decomposition steps are as follows: S5-1: Add Gaussian white noise ω to the engine flight parameter correction data s(t). i (t), resulting in a new signal sequence: Where γ0 is the standard deviation of Gaussian white noise, and n is the number of times white noise is added; S5-2: The new signal sequence obtained above EMD decomposition yields the following decomposition form: Where m is the number of EMD components, IMF1 j (t) represents the j-th EMD component, and res1(t) represents the residual; the first-order CEEMDAN component is obtained from the EMD component. The first-order modal signal margin is: S5-3: Add Gaussian white noise ω to the first-order modal signal margin r1(t) i (t), to obtain a new signal sequence Similar to the steps in S5-2, the signal sequence is processed. EMD decomposition yields the following decomposition form: The second-order CEEMDAN component is obtained from the EMD component. The margin of the second-order modal signal is: S5-4: Repeat step S5-3 until the modal signal margin can no longer be decomposed; at this point, the final decomposition result of the engine flight parameter correction data s(t) is: S6: Use fuzzy entropy to evaluate the complexity of the components mentioned in step S5. Group the components according to the size of the fuzzy entropy and the order of magnitude of the peak value. The components in each group have similar fuzzy entropy and the same order of magnitude. Merge the components in each group to obtain several fused components. It can be ensured that the number of fused components is not greater than the number of original components. S7: Establish a sequence prediction model based on the principle of the Informer model, with each fusion component as input data; take the first 60% of the engine flight parameter correction data as the training set, the middle 20% as the validation set, and the last 20% as the test set. S8: The training data enters the encoder after input embedding and position encoding. Feature extraction and attention calculation are performed in the encoder. The data then enters the decoder through a fully connected feedforward network and residual normalization. Input data with time offset also enters the decoder after input embedding. Attention calculation is performed in the masked multi-head attention module. The relationship mapping between input and output is completed through decoding. Finally, the predicted value is output through a fully connected layer. The steps for outputting the predicted value are as follows: S8-1: Data embedding and positional encoding are performed on the input sequence of the sequence prediction model. Data embedding uses convolutional and fully connected networks to map the data into a 512-dimensional feature vector. Positional encoding embeds the temporal information of the sequence into the feature vector to prevent the loss of temporal information due to attention calculation. Positional encoding is calculated using the encoding method in Transformer. Where u represents the dimension index in the input vector, d model This represents the total number of dimensions of the input sequence, and pos represents the position of the calculated input sequence within all input sequences. S8-2: The transformed feature vectors are fed into the encoder for attention calculation. The attention calculation method is as follows: Where Q,K,V∈R N*d These are the query matrix, key matrix, and value matrix, respectively. N is the length of the input engine flight parameter correction data, d is the hidden dimension in the attention head, and softmax is the normalized exponential function. S8-3: There will be some redundancy in the attention calculation results. Pooling is used to reduce the dimensionality of it so that the important parts have higher weights. Each encoding module contains an attention calculation, a pooling, and a residual and normalization. Multiple encoding modules are stacked in the encoder. After three layers of encoding modules, the output vector of the encoder is obtained. S8-4: Constructing a sequence with a time offset: X feedDecoder =concat(X Token ,X Placeholder ) Among them, X Token X is a known sequence preceding the time point to be predicted. Placeholder The sequence to be predicted is filled with zeros, also known as a placeholder sequence; the sequence with time offset is embedded into the input and then enters the decoder. Attention is calculated in the masked multi-head attention module. The relationship mapping between input and output is completed through decoding. Finally, the predicted value is output through a fully connected layer. S9: Calculate the error between the predicted value and the actual value, calculate the loss function Loss value, and repeat the model training process in step S8 until the training termination condition is met. S10: Use the model trained in step S9 to predict the fusion components in S6, obtain the prediction results of the fusion components, and sum the prediction results of the fusion components to obtain the prediction results of the original flight parameters. S11: Input the predicted results of each flight parameter into the digital twin of the aero-engine and perform simulation to obtain a digital twin that can reflect the predicted working state of the engine. Based on the digital twin, the predicted values ​​of other parameters can be obtained, thereby realizing the parameter prediction of the aero-engine.

Images