An aero-engine gas path fault prediction method based on digital twinning

Abstract

The present application belongs to the field of digital twin technology and aero-engine fault prediction, and provides an aero-engine gas path fault prediction method based on digital twin. The Wiener-BP error correction model proposed by the present application establishes a nonlinear dynamic error compensation model according to the error between the engine digital twin and the sensor measurement value of the real engine, which overcomes the disadvantage of time cumulative error existing in the existing digital twin. At the same time, the Wiener-BP error correction model trained according to the error data has strong generalization ability and can be applied to various types of engine digital twins, and has good universality. The method of the present application obtains the predicted value of the engine gas path parameter through the parameter prediction algorithm, inputs the gas path parameter prediction value into the digital twin after error correction for simulation, obtains the predicted working state of the engine at the future time, and thus realizes the gas path fault prediction.

Description

Technical Field

[0001] This invention belongs to the field of digital twin technology and aero-engine fault prediction, specifically a method for predicting aero-engine air path faults based on digital twins. Background Technology

[0002] Digital twin technology is a simulation process based on virtual-real mapping, capable of dynamically reflecting the real-time state of a corresponding physical object within a digital twin space. Digital twin technology has a wide range of applications, such as product design, industrial manufacturing, and engineering testing. However, in practical applications, due to factors such as modeling accuracy, measurement errors, and unobservable dynamic factors, there is a certain tracking error between the digital twin and the corresponding physical object, affecting the application of digital twins in high-precision scenarios.

[0003] Digital twin models of aero-engines represent the application of digital twin technology in the aviation field. They can reflect the current operating status of the engine in detail, as detailed in the analysis of digital twin applications in aero-engines by Wang Le et al. As the power source of aircraft, the reliability of aero-engines is directly related to aircraft safety. Among all engine failures, gas path failures are the most common type. By simulating and predicting engine operating conditions using digital twin models, potential failures can be detected in advance, improving flight safety, as detailed in the exploration of digital twin applications in aero-engine reliability by Liu Kui et al. Therefore, predicting aero-engine gas path failures places higher demands on the accuracy of digital twins. Existing component-level models struggle to compensate for modeling errors and unknown dynamic disturbances, making them unsuitable for the needs of aero-engine gas path failure prediction. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting aero-engine gas path faults based on digital twins. The method includes an aero-engine digital twin and a parameter prediction algorithm, which can predict the parameter change trend over a future period based on historical and current gas path parameter values. By inputting the predicted parameter values ​​into the digital twin, it determines whether a fault will occur in the future, thereby achieving early prediction of aero-engine gas path faults.

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

[0006] A method for predicting aero-engine gas path faults based on digital twins, comprising the following steps:

[0007] S1: Constructing digital twins of aero-engine components;

[0008] Furthermore, the aero-engine component-level digital twin described in S1 includes an air intake, compressor, combustion chamber, turbine, tail nozzle, and rotor, with input parameters being fuel flow rate Wf, altitude H, and Mach number Ma.

[0009] S2: Calculate the simulation data of the current engine air path parameters using the aero-engine component-level digital twin in S1;

[0010] Furthermore, the engine air path parameter signals mentioned in S2 include rotor relative converted speed, inlet air total temperature, fan adjustable blade angle, compressor adjustable blade angle, turbine inlet temperature, turbine exhaust temperature, turbine exhaust pressure, compressor inlet / outlet temperature, and compressor inlet / outlet pressure.

[0011] S3: During the operation of the aero-engine, the parameters of the air path sensors are collected in real time;

[0012] S4: Denoise the sensor signals acquired in S3;

[0013] Furthermore, the engine sensor parameter signals described in S4 are denoised using wavelet thresholding. For aero-engine sensors, the parameter measurements are discrete data at specific frequencies, requiring the use of discrete wavelet transform.

[0014]

[0015] Among them, (W) ψ f) represents the signal after wavelet decomposition, a0 is the scaling step size, b0 is the basis function factor, and m and n are the scaling and translation coefficients, respectively. The wavelet denoising threshold is determined by the following formula:

[0016]

[0017] Where σ is the standard deviation estimate of the noise, and N is the length of the signal. The threshold function is determined by the following formula:

[0018]

[0019] Where, ω j,k The wavelet coefficients are the wavelet coefficients after wavelet decomposition of the original signal. These are the wavelet coefficients after processing with the appropriate threshold function. The wavelet decomposition signal is then processed using the aforementioned threshold function to obtain the denoised gas path parameter signal.

[0020] S5: Compare the noise-reduced gas path parameter signals with the simulation data of the digital twin to obtain the error between the simulation data and the actual measurement value;

[0021] S6: Establish the Wiener-BP error correction model, use the difference between the simulation data and the real data obtained in S5 to train the model, and use the trained model to correct the digital twin of the aero-engine to obtain the error-corrected digital twin.

[0022] Furthermore, the digital twin error correction model in S6 is based on the Wiener model and a BP neural network. The Wiener model can be viewed as a series connection of a linear dynamic element and a nonlinear static gain. For the nonlinear static gain part, since neural networks have a good fitting ability for nonlinear functions, a BP neural network is used to replace the nonlinear static gain part in the Wiener model. For the Wiener-BP model, the input of its linear dynamic element is u(t), and the output of its linear dynamic element is:

[0023] x(t)=b0u(t)+b1u(t-1)+b2u(t-2)+…+b n u(tn)-a1x(t-1)-…-a m x(tm)

[0024] Among them, a j b j These are all parameters to be trained on the model, and asynchronous inputs of the form u(tn) and x(tm) are implemented using delay operators.

[0025] For the nonlinear static gain stage, a backpropagation (BP) neural network is used. The input to the nonlinear static gain stage is the output x(t) of the linear dynamic stage, and the outputs of its hidden layer nodes are:

[0026]

[0027] in, The weights are input to the i-th hidden layer node for the nonlinear static gain circuit. Let f be the activation threshold of the i-th hidden layer neuron. f is the Sigmoid activation function, expressed as follows:

[0028]

[0029] The output of the nonlinear static gain circuit is:

[0030]

[0031] in, Let be the weights from the i-th hidden layer node to the output of the nonlinear static gain stage. Let be the activation threshold for the i-th hidden layer neuron, and I be the total number of hidden layer nodes. At this point, the process of establishing a digital twin error correction model based on the Wiener model and a BP neural network is complete. The mapping relationship between the output y(t) and the input u(t) of the Wiener-BP model is as follows:

[0032]

[0033] Among them, the parameters to be trained are a j b j , At the start of the simulation, random initialization is performed, and the error between the output value of the digital twin and the true value is used as training data. The gradient descent method based on the minimum mean square error is used to correct the parameters, and finally the error correction model of the engine digital twin is obtained.

[0034] S7: Use the air path parameter data during engine operation as training sample data to establish an aero-engine parameter prediction algorithm;

[0035] Furthermore, the engine parameter prediction algorithm in S7 is based on a Long Short-Term Memory (LSTM) neural network. The input to the LSTM model is the engine operating air path parameter data for the current and historical periods, and the output is the predicted value of the engine air path parameters for a future period.

[0036] S8: Input the predicted values ​​of the air path parameters output by the parameter prediction algorithm in S7 into the error-corrected digital twin to obtain the operating status of the aero-engine at the predicted time.

[0037] S9: Compare the operating status of the twin model in S8 at the prediction time with the engine airflow fault characteristics to determine whether an airflow fault has occurred. If the twin model fails at the prediction time, it indicates that the engine has a high probability of experiencing the same fault after the prediction time has elapsed, thus achieving the purpose of fault prediction.

[0038] Compared with the prior art, the present invention has at least the following beneficial effects:

[0039] (1) This invention proposes an error correction model for aero-engine digital twins based on the Wiener model and a BP neural network. The proposed Wiener-BP error correction model can establish a nonlinear dynamic error compensation model based on the error between the digital twin of the engine and the sensor measurements of the real engine, overcoming the drawback of existing digital twins having time-accumulated errors. At the same time, the Wiener-BP error correction model trained based on error data has strong generalization ability and can be applied to various types of engine digital twins, showing good universality.

[0040] (2) This invention proposes a method for predicting aero-engine gas path faults based on digital twins and parameter prediction algorithms. The predicted values ​​of engine gas path parameters are obtained through parameter prediction algorithms. These predicted values ​​are then input into an error-corrected digital twin for simulation to obtain the expected operating state of the engine at future times, thereby achieving gas path fault prediction. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the present invention.

[0042] Figure 2 This is a test chart showing the error correction effect of the Wiener-BP invention. Detailed Implementation

[0043] 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.

[0044] like Figure 1 As shown, this invention proposes a method for predicting aero-engine gas path faults based on digital twins, comprising the following steps:

[0045] S1: Construct a component-level digital twin of an aero-engine, including components such as the air intake, compressor, combustion chamber, turbine, tail nozzle, and rotor. The input parameters of the twin are fuel flow rate Wf, altitude H, and Mach number Ma.

[0046] S2: Using the current environmental operating condition parameters as preset parameters, the simulation data of the current engine air path parameters is obtained by using the digital twin simulation of the aero-engine component level;

[0047] S2-1: The engine air circuit parameters mentioned in S2 include the rotor relative equivalent speed, total inlet air temperature, fan adjustable blade angle, compressor adjustable blade angle, turbine inlet temperature, turbine exhaust temperature, turbine exhaust pressure, compressor inlet / outlet temperature, and compressor inlet / outlet pressure.

[0048] S3: Real-time acquisition of air path sensor parameter signals during actual aircraft engine operation;

[0049] S4: Use wavelet thresholding to denoise the acquired sensor signals. The specific steps are as follows:

[0050] S4-1: Perform wavelet decomposition on the original sensor signal using discrete wavelet transform:

[0051]

[0052] Among them, (W) ψf) represents the signal after wavelet decomposition, a0 is the scaling step size, b0 is the basis function factor, and m and n are the scaling coefficient and translation coefficient, respectively.

[0053] S4-2: Calculate the wavelet denoising threshold:

[0054]

[0055] Where σ is the standard deviation estimate of the noise, and N is the length of the signal.

[0056] S4-3: Use the following threshold function to process the wavelet decomposition components:

[0057]

[0058] Where, ω j,k The wavelet coefficients are the wavelet coefficients after wavelet decomposition of the original signal. These are the wavelet coefficients after being processed by the corresponding threshold function.

[0059] S4-4: The processed signal components are fused to obtain the noise-reduced gas path parameter signal.

[0060] S5: Compare the noise-reduced gas path parameter signals with the simulation data of the digital twin to obtain the error between the simulation data and the actual measurement value;

[0061] S6: Establish the Wiener-BP error correction model, use the difference between the simulation data and the real data obtained in S5 to train the model, and use the trained model to correct the digital twin of the aero-engine to obtain the error-corrected digital twin.

[0062] S6-1: The Wiener-BP error correction model described in S6 is based on the Wiener model and a BP neural network. The Wiener model can be viewed as a series connection of a linear dynamic element and a nonlinear static gain. For the nonlinear static gain part, since neural networks have a better ability to fit nonlinear functions, a BP neural network is used to replace the nonlinear static gain part in the Wiener model.

[0063] S6-2: For the linear dynamic element in Wiener-BP, if its input is u(t), then its output is:

[0064] x(t)=b0u(t)+b1u(t-1)+b2u(t-2)+…+b n u(tn)-a1x(t-1)-…-a m x(tm)

[0065] Among them, aj b j These are all parameters to be trained on the model, and asynchronous inputs of the form u(tn) and x(tm) are implemented using delay operators.

[0066] S6-3: For the nonlinear static gain stage in Wiener-BP, a BP neural network is used for implementation. Taking a single-hidden-layer BP neural network as an example, the input of the nonlinear static gain stage is the output x(t) of the linear dynamic stage, and the output of its hidden layer nodes is:

[0067]

[0068] in, The weights are input to the i-th hidden layer node for the nonlinear static gain circuit. Let f be the activation threshold of the i-th hidden layer neuron. f is the Sigmoid activation function, expressed as follows:

[0069]

[0070] The output of the nonlinear static gain circuit is:

[0071]

[0072] in, Let be the weights from the i-th hidden layer node to the output of the nonlinear static gain stage. Let be the activation threshold of the i-th hidden layer neuron, and I be the total number of hidden layer nodes.

[0073] S6-4: Based on the linear dynamic element described in S6-2 and the nonlinear static gain element described in S6-3, the mapping relationship between the output y(t) and the input u(t) of the Wiener-BP model is obtained:

[0074]

[0075] Among them, a j b j , These are the parameters to be trained in the Wiener-BP model.

[0076] S6-5: Initialize the training parameters in S6-4 with random values, use the error between the simulation data of the digital twin and the actual measurement values ​​as training data, and use the gradient descent method based on the minimum mean square error to update the weights.

[0077] S6-6: The simulation data output by the engine digital twin is summed with the error compensation value output by the Wiener-BP error correction model to obtain the error-corrected simulation data.

[0078] S7: Use the air path parameter data during engine operation as training sample data to establish an aero-engine parameter prediction algorithm;

[0079] S7-1: The aero-engine parameter prediction algorithm described in S7 is based on LSTM. The input of LSTM is the engine operating air path parameter data for the current and historical periods, and the output is the predicted value of the engine air path parameters for a future period.

[0080] S8: Input the predicted values ​​of the engine air path parameters into the error-corrected digital twin to obtain the operating status of the aero-engine at the predicted time.

[0081] S9: Compare the operating status of the twin model in S8 at the prediction time with the engine airflow fault characteristics to determine whether an airflow fault has occurred. If the twin model fails at the prediction time, it indicates that the engine has a high probability of experiencing the same fault after the prediction time has elapsed, thus achieving the purpose of fault prediction.

[0082] Figure 2 The figure shown is the test result of the Wiener-BP error correction model. It can be seen that compared with the simulation data of the original digital twin model, the data after Wiener-BP error correction is closer to the sensor measurement data of the real engine, that is, it can more accurately reflect the working state of the real engine.

[0083] 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 air path faults in aero-engines based on digital twins, characterized in that, The steps are as follows: S1: Constructing digital twins of aero-engine components; The component-level digital twin of an aero-engine includes the air intake, compressor, combustion chamber, turbine, tail nozzle, and rotor. The input parameters are fuel flow rate Wf, altitude H, and Mach number Ma. S2: Calculate the simulation data of the current engine air path parameters using the aero-engine component-level digital twin from step S1; Engine airflow parameters include rotor relative equivalent speed, total inlet air temperature, fan adjustable blade angle, compressor adjustable blade angle, turbine inlet temperature, turbine exhaust temperature, turbine exhaust pressure, compressor inlet / outlet temperature, and compressor inlet / outlet pressure. S3: During the operation of the aero-engine, real-time acquisition of air path sensor signals; S4: Denoise the gas path sensor signal acquired in step S3; Denoising was achieved using wavelet thresholding. Discrete wavelet transform is first applied to the gas path sensor signal: in, The signal from the gas path sensor after wavelet decomposition. To adjust the scaling step size, As a basis function factor, p and q These are the scaling coefficient and translation coefficient, respectively; the wavelet denoising threshold is determined by the following formula: in, For the standard deviation estimation of noise, The length of the gas path sensor signal; the threshold function is determined by the following formula: in, The wavelet coefficients are the wavelet coefficients after wavelet decomposition of the original gas path sensor signal. These are the wavelet coefficients after being processed by the corresponding threshold function; the above threshold function is used to process the gas path sensor signal after wavelet decomposition to obtain the noise-reduced gas path sensor signal. S5: Compare the noise-reduced air path sensor signal with the simulation data of the aero-engine component-level digital twin to obtain the error between the simulation data and the actual measurement value; S6: Establish a Wiener-BP error correction model, use the difference between the simulation data and the real data obtained in step S5 to train the Wiener-BP error correction model, and use the trained Wiener-BP error correction model to correct the aero-engine component-level digital twin, so as to obtain the aero-engine component-level digital twin after error correction. The Wiener-BP error correction model is based on the Wiener model and a BP neural network. The Wiener model is viewed as a series connection of a linear dynamic element and a nonlinear static gain; the BP neural network replaces the nonlinear static gain component in the Wiener model. For the Wiener-BP error correction model, the input to its linear dynamic element is... Then the output of its linear dynamic element is: in, , These are all parameters to be trained in the Wiener-BP error correction model. , Asynchronous input is implemented using a delay operator; For the nonlinear static gain circuit, a backpropagation (BP) neural network is used for implementation. The input of the nonlinear static gain circuit is the output of the linear dynamic circuit. The output of its hidden layer nodes is: in, Input to the nonlinear static gain circuit i The weights of each hidden layer node. For the first i The activation threshold of each hidden layer neuron; The Sigmoid activation function is expressed as follows: The output of the nonlinear static gain circuit is: in, For the first i The weights from each hidden layer node to the output of the nonlinear static gain stage. For the first i The activation threshold of each hidden layer neuron. I This represents the total number of hidden layer nodes; at this point, the Wiener-BP error correction model based on the Wiener model and BP neural network has been established. Output of the Wiener-BP error correction model With input The mapping relationship between them is: Among them, the parameters to be trained are , , , , , At the start of the simulation, random initialization is performed, and the error between the output value of the digital twin and the true value is used as training data. The gradient descent method based on the minimum mean square error is used to correct the parameters, and finally the error correction model of the engine digital twin is obtained. S7: Use the air path sensor signals during engine operation as training sample data to establish an aero-engine parameter prediction algorithm; The aero-engine parameter prediction algorithm is based on a long short-term memory neural network; the input of the LSTM model is the engine operation air path sensor signal in the current and historical time period, and the output is the predicted value of the engine air path sensor signal in the future period. S8: Input the predicted value of the air path sensor signal output by the parameter prediction algorithm in step S7 into the error-corrected aero-engine component-level digital twin to obtain the operating status of the aero-engine at the predicted time. S9: Compare the operating status of the aero-engine component-level digital twin after error correction in step S8 at the prediction time with the engine air path fault characteristics to determine whether an air path fault has occurred; if the aero-engine component-level digital twin after error correction at the prediction time has a fault, it means that the engine is highly likely to have the same fault after the prediction time, thereby achieving the purpose of fault prediction.

Images