Aero-engine health monitoring method and device
By using interpretable deep neural network model in aero engine health monitoring, the problem of inaccurate similar conversion coefficients in the prior art is solved, the accuracy and reliability of health monitoring are improved, and resource usage is reduced , more accurate steady-state baseline prediction is achieved.
Patent Information
- Application Number
- CN202410993196.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-24
- Publication Date
- 2024-11-12
AI Technical Summary
In the existing model-based aero engine health monitoring technology, the problem of inaccurate similar conversion coefficients leads to inaccurate prediction results of steady-state baseline models, affecting the reliability and accuracy of health monitoring.
The interpretable deep neural network model is used to correct the similarity conversion coefficients, and the preliminary deep neural network model is fitted through ground and high-altitude steady-state data, and interpretive analysis is carried out to streamline the model to generate more accurate corrected similar conversion coefficients for use in Prediction and health monitoring of steady-state baseline models.
It improves the reliability and accuracy of AVIC health monitoring results, reduces the use of software and hardware resources, achieves more accurate steady-state baseline prediction data, and enhances the ability to locate faults.
Smart Images

Figure CN118940407A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to an aero-engine health monitoring method, and belongs to the technical field of aero-engine design and control. Background Art
[0002] The health of aircraft engines is directly related to the safety of aircraft. Engine failures caused by degradation of gas path components account for more than 90% of total engine failures, and their maintenance costs account for more than 60% of their total costs. By monitoring and analyzing the engine status, faults can be discovered in a timely manner and located and quantified, allowing the diagnosis of faults to achieve a leap from timed diagnosis to situation-based diagnosis. It is of great significance to ensure flight safety, reduce maintenance costs and improve operating efficiency, so health monitoring is an indispensable part of the aviation industry.
[0003] Aircraft engine health monitoring is a continuous task that involves measuring, filtering, analyzing, and detecting anomalies from performance parameters to monitor performance status and operational health. This process aims to promptly detect abnormal conditions, failures, or performance degradation in engine operation and provide corresponding feedback and decision support to ensure safe, reliable, and efficient operation of aircraft engines. There are two common approaches to aircraft engine health monitoring: model-free and model-based. In the model-free method, Zhou Han et al. [Zhou Han, Mo Liping, Liu Yuan, et al. Aircraft engine fault diagnosis based on feature optimization and improved KNN [J]. Aviation Computing Technology, 2023, 53(02): 45-49] established an aircraft engine fault diagnosis model based on feature optimization and improved KNN, and studied the impact of fault characteristics on aircraft engines; Liu Lijun et al. [Liu Lijun, Lei Yu, Yu Zhen. Application of bidirectional LSTM model in aircraft engine gas path fault diagnosis [J]. Aerospace Control, 2020, 38(05): 67-72.] proposed an aircraft engine gas path component fault diagnosis model based on bidirectional long short-term memory neural network to accurately locate the location and category of engine gas path component faults. However, the model-free method also has some limitations and is easily affected by data quality and noise. At the same time, when facing unknown situations, it will inevitably affect the generalization ability of health monitoring. In the model-based approach, Feng Lu et al. [Feng Lu; Yu Chen, Jinquan Huang, Dongdong Zhang, Nan Liu. An integrated nonlinear model-based approach to gas turbine engine sensor fault diagnostics [J]. Part G: Journal of Aerospace Engineering. Volume 228, Issue 11. 2014. PP 2007-2021] proposed a nonlinear adaptive Kalman filter model that can provide detailed analysis of the operating status of aircraft engines. However, the model-based approach still faces the problem of inaccurate similarity conversion coefficients when converting the baseline model to high altitude.
[0004] The health status of aircraft engines is of great significance to flight safety. On this basis, research on the health monitoring of aircraft engines has important research value. Summary of the invention
[0005] The technical problem to be solved by the present invention is to overcome the problem of inaccurate similarity conversion coefficients in the existing model-based aircraft engine health monitoring technology, and to provide an aircraft engine health monitoring method that can provide a more accurate corrected similarity conversion coefficient for the steady-state baseline model in real time, and occupy less software and hardware resources.
[0006] The present invention specifically adopts the following technical solutions to solve the above technical problems:
[0007] A method for monitoring the health of an aircraft engine uses a steady-state baseline model to predict measurable parameters of the aircraft engine, and monitors the health of the aircraft engine according to the deviation between the predicted value and the measured value of the measurable parameter; the similarity conversion coefficient of the steady-state baseline model is a modified similarity conversion coefficient generated by an interpretable deep neural network model; the interpretable deep neural network model is established by the following method:
[0008] S1. Fitting a preliminary engine steady-state baseline model based on ground steady-state data;
[0009] S2. Calculate the modified similarity conversion coefficient of the measurable parameter according to the high-altitude steady-state data;
[0010] S3, establishing a preliminary deep neural network model, wherein the input features of the preliminary deep neural network model are the control variables and environmental variables of the aircraft engine, and the output of the preliminary deep neural network model is the modified similarity conversion coefficient of the measurable parameter obtained in step S2;
[0011] S4. Perform an interpretative analysis on the preliminary deep neural network model, simplify the input features and neurons therein, and obtain the interpretable deep neural network model.
[0012] Preferably, the measurable parameters include: compressor outlet pressure P3, low-pressure turbine outlet temperature T5, low-pressure rotor speed N1, and high-pressure rotor speed N2.
[0013] Preferably, an integrated gradient algorithm is used to perform an interpretative analysis on the preliminary deep neural network model.
[0014] Based on the same inventive concept, the following technical solutions can also be obtained:
[0015] An aircraft engine health monitoring device comprises a steady-state baseline model for predicting a measurable parameter of the aircraft engine, and a health monitoring module for monitoring the health status of the aircraft engine according to a deviation between a predicted value and a measured value of the measurable parameter; the similarity conversion coefficient of the steady-state baseline model is a modified similarity conversion coefficient generated by an interpretable deep neural network model; the interpretable deep neural network model is established by the following method:
[0016] S1. Fitting a preliminary engine steady-state baseline model based on ground steady-state data;
[0017] S2. Calculate the modified similarity conversion coefficient of the measurable parameter according to the high-altitude steady-state data;
[0018] S3, establishing a preliminary deep neural network model, wherein the input features of the preliminary deep neural network model are the control variables and environmental variables of the aircraft engine, and the output of the preliminary deep neural network model is the modified similarity conversion coefficient of the measurable parameter obtained in step S2;
[0019] S4. Perform an interpretative analysis on the preliminary deep neural network model, simplify the input features and neurons therein, and obtain the interpretable deep neural network model.
[0020] Preferably, the measurable parameters include: compressor outlet pressure P3, low-pressure turbine outlet temperature T5, low-pressure rotor speed N1, and high-pressure rotor speed N2.
[0021] Preferably, an integrated gradient algorithm is used to perform an interpretative analysis on the preliminary deep neural network model.
[0022] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:
[0023] (1) The present invention considers the actual flight data to correct the similarity conversion coefficient. Compared with the traditional solution of using a constant similarity conversion coefficient, the present invention solves the problem of inaccurate similarity conversion coefficient and improves the reliability and accuracy of the aero-engine health monitoring results.
[0024] (2) The present invention analyzes and simplifies the deep neural network model based on an interpretable method, and can use fewer software and hardware resources to obtain more accurate corrected similarity conversion coefficients in real time, thereby obtaining more accurate steady-state baseline prediction data. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 It is a cross-sectional definition diagram of a twin-shaft turbofan engine;
[0026] Figure 2 This is a schematic diagram of the health monitoring principle;
[0027] Figure 3 A schematic diagram of the structure of the health monitoring system established by the present invention;
[0028] Figure 4 It is the flow chart of conversion factor correction;
[0029] Figure 5 This is a preliminary schematic diagram of the deep neural network model structure;
[0030] Figure 6 Visualize the gradient integral of the characteristic variable;
[0031] Figure 7 Visualize the integral gradient of a neuron;
[0032] Figure 8(a) is a comparison of the flight data and the baseline model;
[0033] Figure 8(b) is a comparison of flight data and deep interpretable model;
[0034] Fig. 9 This is the fault monitoring diagram of channel A. DETAILED DESCRIPTION
[0035] The existing model-based aircraft engine health monitoring solution uses the steady-state baseline model as a virtual sensor to predict the measurable parameters of the aircraft engine, compares the steady-state baseline model prediction value with the actual steady-state measurement value, and uses the residual between the actual measurable parameter and the measurable parameter predicted by the steady-state baseline model to extract the health parameters for health monitoring. Under the premise that the steady-state baseline model is accurate, the residual can be used to analyze the degradation of the engine and perform sensor fault monitoring.
[0036] The existing steady-state baseline model uses a constant similarity conversion coefficient for similarity conversion to apply to the entire flight envelope. Although this method is simple and easy, it will inevitably lead to inaccurate similarity conversion coefficients, which in turn affects the accuracy of the steady-state baseline model prediction results.
[0037] In view of the above problems, the solution of the present invention is to correct the similarity conversion coefficient based on real flight data, and establish a deep neural network model covering the full envelope based on the corrected similarity conversion coefficient to generate accurate corrected similarity conversion coefficients in real time, and streamline the established deep neural network model through the method of interpretable analysis, so that accurate corrected similarity conversion coefficients can be obtained in real time with minimal software and hardware resources, thereby obtaining more accurate steady-state data prediction results.
[0038] The present invention specifically adopts the following technical solutions to solve the above technical problems:
[0039] A method for monitoring the health of an aircraft engine uses a steady-state baseline model to predict measurable parameters of the aircraft engine, and monitors the health of the aircraft engine according to the deviation between the predicted value and the measured value of the measurable parameter; the similarity conversion coefficient of the steady-state baseline model is a modified similarity conversion coefficient generated by an interpretable deep neural network model; the interpretable deep neural network model is established by the following method:
[0040] S1. Fitting a preliminary engine steady-state baseline model based on ground steady-state data;
[0041] S2. Calculate the modified similarity conversion coefficient of the measurable parameter according to the high-altitude steady-state data;
[0042] S3, establishing a preliminary deep neural network model, wherein the input features of the preliminary deep neural network model are the control variables and environmental variables of the aircraft engine, and the output of the preliminary deep neural network model is the modified similarity conversion coefficient of the measurable parameter obtained in step S2;
[0043] S4. Perform an interpretative analysis on the preliminary deep neural network model, simplify the input features and neurons therein, and obtain the interpretable deep neural network model.
[0044] In order to facilitate public understanding, the technical solution of the present invention is described in detail below through a specific embodiment and in conjunction with the accompanying drawings:
[0045] This embodiment takes a certain type of dual-shaft turbofan engine as the research object, and the cross-section of each component is as follows: Figure 1 As shown, the main gas path components include: air inlet, compressor, high-pressure compressor, combustion chamber, high-pressure turbine, low-pressure turbine, mixing chamber and nozzle.
[0046] The aircraft engine adopts a model-based health monitoring solution, the basic principle of which is as follows Figure 2 As shown, the predicted value y0 output by the steady-state baseline model is compared with the actual steady-state measurement value y. The residual between the actual parameter and the baseline model prediction value is used to extract health parameters for health monitoring. Under the premise of accurate model, the residual can be used to analyze the degradation of the engine and perform sensor fault monitoring.
[0047] It is difficult to distinguish engine degradation and model error in the residual. In order to avoid the influence of model error on health monitoring, the steady-state baseline model needs to be further modified. The basic architecture of the health monitoring system proposed in this invention is as follows: Figure 3 As shown, an interpretable deep neural network model is introduced on the basis of the classic residual estimation method to correct the similarity conversion coefficient C in the steady-state baseline model in real time, and the measurable parameter measurement value y0 is obtained based on the steady-state baseline model with the corrected similarity conversion coefficient. Finally, the corresponding aircraft engine health monitoring result is obtained by analyzing the residual between the measured value y and y0 of the measurable parameter of the real engine. The input of the deep network model is the control variable u and the environmental variable v, and the output includes the correction coefficient C of the compressor outlet pressure P3, the low-pressure turbine outlet temperature T5, the low-pressure rotor speed N1, and the high-pressure rotor speed N2.
[0048] The specific process of establishing the interpretable deep neural network model is as follows:
[0049] S1. Fitting a preliminary engine steady-state baseline model based on ground steady-state data;
[0050] The engine steady-state baseline model can be expressed as a fixed smooth curve with a certain trend, and is only related to its corresponding control variable. A single fixed mapping relationship is established according to the corresponding input and output parameters. Its function correspondence is:
[0051]
[0052] In the formula, q0, q1, ... q n are the coefficients of the n-order linear polynomial, where n represents the order of the polynomial, and w fb,cor represents the converted fuel flow, y0 represents the controlled conversion amount, where y0=[N cor ,T cor ,P cor ],N cor is the speed conversion value of high and low pressure rotors, T cor is the temperature conversion value of any cross section, P cor is the converted value of pressure at any cross section.
[0053] S2. Calculate the modified similarity conversion coefficient of the measurable parameter based on the high-altitude steady-state data:
[0054] The steady-state baseline model uses similar conversion to apply to the entire flight envelope. After the steady-state data is extracted, the conversion factor of the baseline model is corrected according to the high-altitude steady-state data. Figure 4 For the conversion factor correction process, the similar conversion factor correction formula for the speed of the high and low pressure rotors is:
[0055]
[0056] In the formula N is the speed measurement value of the high and low pressure rotors measured by the engine, T2 is the measured value of the engine compressor inlet temperature, and only the conversion coefficient c is an unknown number. The corrected conversion coefficient value can be calculated by reverse calculation based on the measured high-altitude steady-state data.
[0057] Similarly, the correction similarity conversion factor formula for temperature and pressure is:
[0058]
[0059] Where T 2,cor =273.15k,P 2,cor =101325Pa, T is the temperature value measured at any section of the engine, T2 is the temperature measured at the engine compressor inlet, P is the pressure value measured at any section of the engine, and P2 is the pressure measured at the engine compressor inlet.
[0060] S3. Establish a preliminary deep neural network model:
[0061] The preliminary deep neural network model structure established in this embodiment is as follows: Figure 5 As shown, it contains 1 input layer, and the input characteristic parameters include: height H, Mach number Ma, inlet inlet temperature T2, inlet inlet pressure P2, guide vane angle α and converted fuel flow w fb,cor ; The middle part is 4 layers of adjustable hidden layers, each layer initialized with 10 neurons; the last is the output layer, including the corrected similarity conversion coefficient C of the compressor outlet pressure P3, the low-pressure turbine outlet temperature T5, the low-pressure rotor speed N1, and the high-pressure rotor speed N2 P3 , C T5 , C N1 , C N2 .
[0062] Each hidden layer contains multiple neurons, each with its own weight w and bias b. The formula for calculating the input value of the kth neuron in the mth layer is as follows:
[0063]
[0064] Where f is the activation function; Z m-1 is the number of neurons in the m-1th layer; is the i-th neuron in the m-1-th layer; is the weight between neurons; The bias of the nth neuron in the mth layer.
[0065] The loss function consists of two parts. The first part is the residual between the output value and the expected value. The second part is the regularization loss caused by overfitting, which is controlled by the parameter θ. The loss function can be expressed as:
[0066]
[0067] Where w and b represent the weight value and bias parameter in the neural network respectively.
[0068] The preliminary deep neural network model is trained by using the high-altitude steady-state data collected in S2 and the calculated corrected similarity conversion coefficient to obtain the altitude H, Mach number Ma, inlet inlet temperature T2, inlet inlet pressure P2, guide vane angle α and converted fuel flow w fb,cor As input features, the modified similarity conversion coefficient C of compressor outlet pressure P3, low-pressure turbine outlet temperature T5, low-pressure rotor speed N1, and high-pressure rotor speed N2 is used. P3 , C T5 , C N1 , C N2 The output of the preliminary deep neural network model.
[0069] S4. Perform an interpretative analysis on the preliminary deep neural network model, simplify the input features and neurons therein, and obtain the interpretable deep neural network model:
[0070] The preliminary deep neural network model obtained by S3 can realize real-time correction of the similarity conversion coefficient in the steady-state baseline model, but considering that the preliminary deep neural network model occupies more software and hardware resources, it is not conducive to airborne real-time health monitoring, so it is necessary to further streamline it. The present invention further uses an explanatory analysis method to analyze the preliminary deep neural network model, eliminating some input features and neurons that have a small contribution to the output.
[0071] The present invention can adopt various existing neural network interpretive analysis methods. The present invention preferably uses the integral gradient algorithm to perform interpretive analysis on the preliminary deep neural network model, and interprets the features and neurons through gradient integral to understand the contribution to the output. The integral gradient algorithm is a prior art. For more details, please refer to the literature [Zeng Chunyan, Yan Kang, Wang Zhifeng, et al. A review of the research on the interpretability of deep learning models [J]. Computer Engineering and Applications, 2021, 57(08): 1-9.]. The gradient integral formula for the feature can be expressed as:
[0072]
[0073] Where δ is an integral variable from 0 to 1, f(δ·X) represents the model output after scaling the feature input X to δ·X, and then calculating the gradient relative to X, where X = [u, v].
[0074] Figure 6 This is a visualization diagram of the gradient integral of the input characteristic variable in this embodiment. It can be seen that T2 has little effect on the similarity conversion factor. This is related to the characteristics of the temperature sensor itself. The lag link causes inaccurate temperature measurement values, which directly affects the correlation between the conversion factor and the temperature measurement value. Therefore, the input characteristic variable T2 is eliminated.
[0075] The gradient integral formula for neurons can be expressed as:
[0076]
[0077] Where: f(x) is the output of the prediction model, x is the neuron input, and z is the neuron output.
[0078] Figure 7 This is a visualization diagram of the gradient integral of the neurons in this embodiment. The neuron with index 4 has hardly learned anything, and the neuron with the largest contribution is index 7, so the neuron with index 4 can be eliminated.
[0079] The simplified deep neural network model after removing the input feature variable T2 and the neuron with index 4 is trained again, and finally the interpretable deep neural network model is obtained.
[0080] Figures 8(a) and 8(b) show the comparison results of the prediction effect of the baseline model based on the deep interpretable model established by the present invention and the traditional baseline model on the test set. The horizontal axis is the number of steady-state samples, and the vertical axis is the true value and predicted value of the measurement parameter. As can be seen from Figure 8(a), the traditional baseline model has high accuracy in most cases, but the error is large in local states; as can be seen from Figure 8(b), the baseline model based on the deep interpretable model of the present invention is basically consistent with the real data, indicating that the method proposed in this patent has good prediction accuracy.
[0081] The baseline model prediction value acts as a virtual sensor. When the error of 5 consecutive steady-state values exceeds the threshold of 3%, the sensor is judged to be faulty. The measurement parameters used include two channels, A and B. Under no fault conditions, the two sensors A and B are almost completely consistent. When one of the channels fails, such as Fig. 9 As shown in the figure, the measured value of channel B is very close to the predicted value of the method of the present invention, but there is a large deviation from the value of channel A, which can accurately locate the fault in channel A. The method of deep interpretable model can realize sensor fault diagnosis, which shows that the method proposed in the present invention has a good use effect.
Claims
1. A method for monitoring the health of an aircraft engine, using a steady-state baseline model to predict measurable parameters of the aircraft engine, and monitoring the health of the aircraft engine according to the deviation between the predicted value and the measured value of the measurable parameter; characterized in that: The similarity conversion coefficient of the steady-state baseline model is a modified similarity conversion coefficient generated by an interpretable deep neural network model; the interpretable deep neural network model is established by the following method: S1. Fitting a preliminary engine steady-state baseline model based on ground steady-state data; S2. Calculate the modified similarity conversion coefficient of the measurable parameter according to the high-altitude steady-state data; S3, establishing a preliminary deep neural network model, wherein the input features of the preliminary deep neural network model are the control variables and environmental variables of the aircraft engine, and the output of the preliminary deep neural network model is the modified similarity conversion coefficient of the measurable parameter obtained in step S2; S4. Perform an interpretative analysis on the preliminary deep neural network model, simplify the input features and neurons therein, and obtain the interpretable deep neural network model.
2. The aircraft engine health monitoring method according to claim 1, characterized in that: The measurable parameters include: compressor outlet pressure P3, low-pressure turbine outlet temperature T5, low-pressure rotor speed N1, and high-pressure rotor speed N2.
3. The aircraft engine health monitoring method according to claim 1, characterized in that: The preliminary deep neural network model was interpreted using the integrated gradient algorithm.
4. An aircraft engine health monitoring device, comprising a steady-state baseline model for predicting a measurable parameter of the aircraft engine, and a health monitoring module for monitoring the health status of the aircraft engine according to the deviation between the predicted value and the measured value of the measurable parameter; characterized in that: The similarity conversion coefficient of the steady-state baseline model is a modified similarity conversion coefficient generated by an interpretable deep neural network model; the interpretable deep neural network model is established by the following method: S1. Fitting a preliminary engine steady-state baseline model based on ground steady-state data; S2. Calculate the modified similarity conversion coefficient of the measurable parameter according to the high-altitude steady-state data; S3, establishing a preliminary deep neural network model, wherein the input features of the preliminary deep neural network model are the control variables and environmental variables of the aircraft engine, and the output of the preliminary deep neural network model is the modified similarity conversion coefficient of the measurable parameter obtained in step S2; S4. Perform an interpretative analysis on the preliminary deep neural network model, simplify the input features and neurons therein, and obtain the interpretable deep neural network model.
5. The aircraft engine health monitoring device according to claim 4, characterized in that: The measurable parameters include: compressor outlet pressure P3, low-pressure turbine outlet temperature T5, low-pressure rotor speed N1, and high-pressure rotor speed N2.
6. The aircraft engine health monitoring device according to claim 4, characterized in that: The preliminary deep neural network model was interpreted using the integrated gradient algorithm.
