Adaptive Correction Method for Mainstream Fast Computational Models of Aero-engines Based on Measured Parameters

Abstract

This invention discloses an adaptive correction method for a mainstream rapid calculation model of an aero-engine based on measured parameters. The method includes: developing a rapid calculation method for mainstream key parameters based on measured parameters, using aerodynamics, thermodynamics, and neural network component modeling principles, combined with the thermal balance method; establishing a mainstream rapid calculation model for the main rotating components of the engine based on measured parameters; analyzing the influence of component performance on the residuals of the degradation equilibrium equation, and determining feasible solutions for component characteristic adjustment parameters used for matching and correcting the mainstream rapid calculation model; and designing an adaptive correction method based on OSELM-EKF for component characteristic adjustment parameter estimation and pressure ratio compensation, based on the mainstream rapid calculation model, so that the established model can more accurately and quickly reflect the current state of the engine during flight and provide more accurate mainstream parameters for the inlet and outlet of rotating components.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine modeling and simulation, specifically involving an adaptive correction method for a mainstream fast calculation model of aero-engines based on measured parameters. Background Technology

[0002] Nonlinear models of aero-engines are established based on the characteristics of components in the engine's design state. However, due to individual differences in the engine manufacturing process and the inevitable performance degradation during service, the rated engine model may fail to reflect the actual engine operating state. To more accurately reflect the current engine state during flight and provide more precise mainstream parameters, adaptive correction of the established mainstream fast calculation model is essential. For aero-engine models that do not use measured parameters, adaptive correction using the deviation between the actual engine and model output parameters via filtering algorithms is usually sufficient. However, for mainstream fast calculation models that use measured parameters, directly using the deviation between the actual engine and model output parameters for adaptive correction is not feasible because some output parameter values ​​are measured. During engine performance degradation, the thermodynamic parameters of components change, and the residuals of their equilibrium equations also change accordingly. Therefore, adaptive correction based on these residuals is of practical significance. Based on this idea, this invention establishes an adaptive correction method for mainstream fast calculation models based on OSELM-EKF and verifies its effectiveness. Summary of the Invention

[0003] The technical problem to be solved by this invention is to address the shortcomings of the prior art, based on measured parameters (low-pressure speed N). L High pressure and high speed N H Total fan outlet pressure P 22 The mainstream fast calculation model of low-pressure turbine outlet total pressure (P5) is used to design an OSELM-EKF adaptive scheme with self-learning function.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0005] Step A) Based on the measured parameters of the engine, and combined with the thermal balance method, we study the mainstream key parameter calculation methods based on the measured parameters, and establish the mainstream fast calculation model based on the measured parameters.

[0006] Step B) Analyze the impact of component performance on the residuals of the degradation equilibrium equation and determine the feasible solution of adjustment parameters for estimating component characteristic adjustment parameters.

[0007] Step C) Based on the mainstream fast calculation model based on measured parameters, an adaptive correction method for component characteristic adjustment parameter estimation and pressure ratio compensation based on OSELM-EKF is designed. The adaptive correction method is developed in combination with the model, and the effectiveness of the correction method is verified.

[0008] Furthermore, the specific steps of step A, which involves studying mainstream key parameter calculation methods based on measured engine parameters using the thermal balance method and establishing a mainstream fast calculation model based on measured parameters, are as follows:

[0009] Step A1) Replace the aerodynamic-thermodynamic model of the cold-end rotating component with a neural network model. The input parameter for the input layer of the fan component neural network model is selected as the low-pressure speed n. L Total inlet temperature T2, fan pressure ratio π fan The output parameters of the output layer are selected based on the total temperature T at the fan outlet. 22 The calculation process can be expressed by the following formula:

[0010] T 22 =f1(n L ,T2,π fan )

[0011] Where f1(·) is the network function for the fan component.

[0012] Input parameters for the neural network model of compressor components: high pressure speed n H Imported total temperature T 25 Import total pressure P 25 Compressor pressure ratio π comp The output parameters for the output layer are selected as follows: compressor outlet total temperature T3, compressor outlet flow rate W3, and interstage bleed air temperature T required for high and low pressure turbine airflow mixing calculations. 27 The calculation process can be expressed by the following formula:

[0013] [T 27 [,T3,W3]=f2(n H ,T 25 ,P 25 ,π comp )

[0014] Where f2(·) is the compressor component network function.

[0015] Step A2), calculate the combustion chamber outlet temperature T4 and the high-pressure turbine inlet temperature T using the heat balance method. 41 The process can be described in the following form:

[0016] (1) Calculate the compressor outlet enthalpy H3 using the compressor outlet section temperature T3:

[0017] H3 = f T2H (0,T3)

[0018] Among them, f T2H (·) is the temperature-enthalpy conversion function.

[0019] (2) Calculate the combustion chamber outlet enthalpy H4 and the high-pressure turbine inlet enthalpy H after cold air mixing. 41 :

[0020]

[0021] Among them, the low calorific value H of fuel oil μ =42900KJ / kg, η B For combustion efficiency, a -导冷气 W3 is the proportion of air drawn into the high-pressure turbine guide vane in the compressor outlet flow rate, W4 is the compressor outlet air flow rate, and W5 is the combustion chamber outlet air flow rate. 41 W is the airflow rate at the high-pressure turbine inlet after cold air mixing. fb Main fuel flow rate.

[0022] (3) Utilizing the enthalpy H4 at the combustion chamber outlet and the enthalpy H at the high-pressure turbine inlet 41 Calculate the combustion chamber outlet temperature T4 and the high-pressure turbine inlet temperature T. 41 :

[0023]

[0024] Among them, f4 and f 41 The air-fuel ratio at the combustion chamber outlet section and the high-pressure turbine inlet section, f H2T (·) is the enthalpy-temperature conversion function.

[0025] Step A3) Establish a mainstream fast calculation model based on measured parameters, using the measured speed n at high and low pressure. Hr n Lr Measured total pressure P at the fan outlet 22r Measured total pressure at turbine outlet P 5r With support, the simplified state-space expression of the LPV model in mainstream fast computation models based on measured parameters can be described by the following formula:

[0026]

[0027] Among them, state variables Control quantity u = [W fb A8] T Output The subscript r represents the measured parameter.

[0028] Based on the measured parameters of the current engine, the pressure ratio parameters of each rotating component under this condition can be obtained by combining the above formula. Running the neural network model of the fan and compressor components can yield the main parameters of each mainstream component in front of the combustion chamber. Then, the combustion chamber outlet temperature T4 and the high-pressure turbine inlet temperature T are calculated using the heat balance method. 41 This allows you to obtain the main parameters of the engine as a whole.

[0029] Furthermore, the specific steps in step B, which analyzes the impact of component performance on the residuals of the degradation equilibrium equation and determines the feasible solution of the adjustment parameters used for component characteristic adjustment parameter estimation, are as follows:

[0030] The residuals of the equilibrium equation are described by equation (1), and equation (2) describes the deviation between the main parameters calculated by the model and the measured parameters:

[0031]

[0032]

[0033] Among them, J H and J L These represent the rotational inertia of the high-pressure and low-pressure rotor shafts, respectively, in N. HT N LT The power outputs of the high-pressure and low-pressure turbines are respectively, N F N C and N ex These are the power consumption of the fan, compressor, and related accessories, in W. g P represents the gas flow rate at each cross-section. s For the static pressure at each section, W cool This refers to the induced air flow rate.

[0034] When the adjustment parameters for engine component characteristics degrade, their actual measured parameters also change. However, when these actual measured parameters are input into the mainstream fast calculation model, the model assumes the component is in a healthy state. Therefore, the mainstream parameters calculated based on this model cause changes in the residuals of the balance equation and their deviation from the measured parameters, with the overall trend of the residuals increasing. Thus, it is feasible to use these residual changes to estimate the component characteristic adjustment parameters.

[0035] Furthermore, step C, based on the mainstream fast calculation model based on measured parameters, designs an adaptive correction method for component characteristic adjustment parameter estimation and pressure ratio compensation based on OSELM-EKF. The specific steps for verifying the effectiveness of the adaptive correction method, combined with model development, are as follows:

[0036] Step C1): When the performance of engine components degrades, the component characteristic adjustment parameters are estimated using an extended Kalman filter (EKF) based on the change in residuals. The system state variable at time k is selected as x. k=[SE1 SW1 SW2 SE3] T SE i SW is the efficiency coefficient of the rotating component. i Let u be the flow coefficient of the rotating component, and let u be the control variable of the system at time k. k =[W fb A8] T The measured value y at time k k =[e1 e2 e3 e4 e5 e6 e7] T This represents the residuals of the equilibrium equation.

[0037] Step C2) involves cross-selecting different degradation modes and degrees of performance parameters at different state points. Without updating the network parameters, multiple independent OSELM offline networks are used to train the training set, and the output weights are assigned according to Euclidean distance. After weighted summation, the final output is the compensation matrix for the pressure ratio caused by performance degradation.

[0038] Assuming a total of p OSELM offline networks are trained, at the k-th step of performance estimation, the pressure ratio compensation matrix obtained from each network is used. The residuals e of each equilibrium equation at this point are obtained. i =[e i1 e i2 e i3 e i4 e i5 e i6 e i7 ] T The subscript i represents the i-th OSELM offline network, which is the Euclidean distance to the ideal residual equilibrium point. weight λ i The assignment conditions are shown in the following formula:

[0039]

[0040] Step C3) updates the model with the pressure ratio compensation matrix obtained by weighted summation of multiple independent OSELM offline network outputs. Under the condition of performance parameter degradation, the engine state-space model expression can be written in the following form:

[0041]

[0042] in, This is the compensation matrix for the pressure ratio caused by performance degradation.

[0043] Step C4) Design a network self-learning module for the OSELM-EKF adaptive correction scheme to update the network topology parameters when the neural network accuracy is insufficient. The network topology parameters are updated based on the residuals of the model's output balance equations. If the residuals exceed a threshold, the network topology parameters are updated. This process can be described by the following formula:

[0044]

[0045]

[0046] Where, β k and β k-1 P represents the output weight update values ​​obtained at the current time step and the previous time step, respectively. k and P k-1 h represents the gain terms at the current time and the previous time, respectively. k This is the output matrix of the hidden layer at the current moment.

[0047] During the OSELM network iteration update process, the network parameters are iteratively updated based on the data blocks obtained at each time step until the training ends. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the OSELM-EKF adaptive correction scheme with self-learning capabilities;

[0049] Figure 2 These are test results of the neural network model for the fan component, where (a) shows the time consumption check of the neural network model, and (b) shows the T... 22 Accuracy simulation verification diagram;

[0050] Figure 3 These are test results of the neural network model for compressor components, where (a) shows the time consumption check of the neural network model, (b) shows the T3 accuracy simulation verification diagram, and (c) shows the T... 27 Accuracy simulation verification diagram, (d) is the accuracy simulation verification diagram of W3;

[0051] Figure 4 This is a flowchart of the mainstream fast computing model;

[0052] Figure 5 The following is a comparison of simulation results of the OSELM-EKF adaptive correction scheme tracking when performance degrades: (a) is the estimated gas path performance parameters, (b) is the fan outlet section, (c) is the compressor outlet section, (d) is the combustion chamber outlet section, (e) is the high-pressure turbine outlet section, and (f) is the low-pressure turbine outlet section. Detailed Implementation

[0053] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0054] The idea behind this invention is to introduce measured engine parameters and establish a mainstream fast calculation model based on measured parameters using the thermal balance method. At the same time, taking the established mainstream fast calculation model of the thermal balance method as the research object, an adaptive correction method based on OSELM-EKF is designed.

[0055] The specific embodiments of the present invention take a dual-rotor turbofan engine as the research object. Figure 1 This is a flowchart illustrating the OSELM-EKF adaptive correction scheme with self-learning capabilities. The design of this adaptive correction scheme includes the following steps:

[0056] Step A) Based on the measured parameters of the engine, and combined with the thermal balance method, we study the mainstream key parameter calculation methods based on the measured parameters, and establish the mainstream fast calculation model based on the measured parameters.

[0057] Step B) Analyze the impact of component performance on the residuals of the degradation equilibrium equation and determine the feasible solution of adjustment parameters for estimating component characteristic adjustment parameters.

[0058] Step C) Based on the mainstream fast calculation model based on measured parameters, an adaptive correction method for component characteristic adjustment parameter estimation and pressure ratio compensation based on OSELM-EKF is designed. The adaptive correction method is developed in combination with the model, and the effectiveness of the correction method is verified.

[0059] The detailed steps of step A) are as follows:

[0060] Step A1) Replace the aerodynamic thermodynamic model of the cold-end rotating component with a neural network model.

[0061] Taking a twin-rotor turbofan engine as the research object, the definitions of each section are as follows:

[0062] Table 1 Definitions of various sections of a turbofan engine

[0063]

[0064] The fan components only need to provide the total outlet temperature T. 22 and total export pressure P 22 Two parameters, simultaneously the total outlet pressure P 22 The following formula can be used for direct calculation. Therefore, to minimize the complexity of the neural network of the fan component, the output parameter only needs to be selected as the total temperature T at the fan outlet. 22 .

[0065] P 22 =P2*π fan

[0066] Where, πfan This represents the fan pressure ratio.

[0067] The influence of fan components on the total outlet temperature T 22 Correlation analysis was performed on the parameters, and the following input parameters were selected: low-pressure speed n L Total inlet temperature T2, fan pressure ratio π fan Different flight state points are selected, and relevant parameters are obtained through a component-level model. This leads to the establishment of a dataset of "input parameters - output parameters" for the fan component. A neural network method is then used to train the training set, resulting in an offline neural network model for the fan component. The calculation process can be represented by the following formula:

[0068] T 22 =f1(n L ,T2,π fan )

[0069] Where f1(·) is the network function for the fan component.

[0070] Using an empirical formula for selecting the number of neurons, the number of neurons in the input layer of the neural network was set to 3, and the number of neurons in the output layer to 1. The neural network was trained using the obtained "input parameter-output parameter" dataset of the fan component. The large dataset was divided into training data, validation data, and test data, with proportions of 70%, 15%, and 15%, respectively. The Levenberg-Marquardt optimization algorithm, which is suitable for medium-sized networks, was selected for training the training set. The output of the neural network model was checked against the output of the component-level model. Once the accuracy requirements were met, training was terminated, and an offline neural network model of the fan component was built using mainstream fast computation models.

[0071] After the offline neural network model of the fan component was established, its time consumption and accuracy were checked. The thermodynamic model and the neural network model of the fan component were each run 50,000 times, and their time consumption was compared as follows: Figure 2 As shown in (a), the total time taken for the thermodynamic model of the fan component is approximately 0.0780s, while the time taken for the neural network model is approximately 0.0470s, saving approximately 40% of the time.

[0072] The accuracy of the output parameters of the neural network model for the fan component was tested at a ground point. Initial main fuel flow rate W fb =0.8kg / s, final main fuel flow rate W fb =1.2 kg / s, simulation steps 1000, step size 0.025 s, the initial fuel quantity jumps to the final fuel quantity at 12.5 s, the thermodynamic model and neural network model of the fan component are simulated and verified, and the simulation results are as follows. Figure 2As shown in (b), the output error between the neural network model and the thermodynamic model of the fan component is small, with a maximum error of only 0.48%, proving that the neural network model of the fan component has high accuracy and can replace the thermodynamic model of the fan component.

[0073] For compressor components, the more important output parameters include: compressor outlet total temperature T3, compressor outlet total pressure P3, compressor outlet flow rate W3, and interstage bleed air temperature T required for high and low pressure turbine airflow mixing calculations. 27 Similarly, the compressor outlet total pressure P3 can be directly calculated using the following formula, therefore only the remaining three output parameters need to be selected. Based on the internal air path calculation of the compressor components, a correlation analysis of the output parameters is performed, and the selected input parameters are as follows: high pressure speed n. H Imported total temperature T 25 Import total pressure P 25 Compressor pressure ratio π comp .

[0074] P3 = P 25 *π comp

[0075] The neural network calculation process for compressor components can be represented by the following formula:

[0076] [T 27 [,T3,W3]=f2(n H ,T 25 ,P 25 ,π comp )

[0077] Where f2(·) is the compressor component network function.

[0078] The "input parameter-output parameter" dataset of the compressor component was obtained using the same method. The number of neurons in the input layer of the neural network was set to 4, and the number of neurons in the output layer was set to 3.

[0079] After the offline neural network model of the compressor components was established, its time consumption and accuracy were checked. The thermodynamic model and the neural network model of the compressor components were each run 50,000 times, and their time consumption was compared as follows: Figure 3 As shown in (a), the total time taken by the thermodynamic model of the compressor components is approximately 0.1090 s, while the time taken by the neural network model is approximately 0.0310 s, which saves about 70% of the time compared to the component thermodynamic model.

[0080] The accuracy of the output parameters of the compressor component's neural network was tested at a ground point. The initial main fuel flow rate Wfb = 1.5 kg / s, the intermediate main fuel flow rate Wfb = 2.4 kg / s, and the final main fuel flow rate Wfb = 1.8 kg / s. The simulation consisted of 1000 steps with a step size of 0.025 s. From 5 to 10 s, the fuel flow rate linearly increased from the initial to the intermediate level; from 10 to 15 s, it remained at the intermediate level; and from 15 to 20 s, it linearly decreased from the intermediate level to the final level. The output parameters of the compressor component's thermodynamic model and neural network model were verified through simulation. The simulation results are as follows: Figure 3 As shown in (b) to (d), the output errors of the compressor component neural network model and the component thermodynamic model are relatively small. The maximum errors of the three output parameters are 0.46%, 0.28%, and 0.80%, respectively, which proves that the compressor component neural network model has high accuracy and can replace the compressor component thermodynamic model.

[0081] In summary, replacing the thermodynamic models of fan and compressor components with neural network models further improves their real-time performance with minimal impact on accuracy. Therefore, using component neural network models to replace component thermodynamic models in mainstream fast computation models is of positive significance.

[0082] Step A2), calculate the combustion chamber outlet temperature T4 and the high-pressure turbine inlet temperature T using the heat balance method. 41 .

[0083] The idea behind the heat balance method is to utilize the heat balance within the combustion chamber and the compressor outlet temperature to calculate the enthalpy at the high-pressure turbine inlet, thereby determining the high-pressure turbine inlet temperature. The specific calculation process is as follows:

[0084] (1) Calculate the compressor outlet enthalpy H3 using the compressor outlet section temperature T3:

[0085] H3 = f T2H (0,T3)

[0086] Among them, f T2H (·) is the temperature-enthalpy conversion function.

[0087] (2) Calculate the combustion chamber outlet enthalpy H4 and the high-pressure turbine inlet enthalpy H after cold air mixing. 41 :

[0088]

[0089] Among them, the low calorific value H of fuel oil μ =42900KJ / kg, η B For combustion efficiency, a -导冷气W3 is the proportion of air drawn into the high-pressure turbine guide vane in the compressor outlet flow rate, W4 is the compressor outlet air flow rate, and W5 is the combustion chamber outlet air flow rate. 41 W is the airflow rate at the high-pressure turbine inlet after cold air mixing. fb Main fuel flow rate.

[0090] (3) Utilizing the enthalpy H4 at the combustion chamber outlet and the enthalpy H at the high-pressure turbine inlet 41 Calculate the combustion chamber outlet temperature T4 and the high-pressure turbine inlet temperature T. 41 :

[0091]

[0092] Among them, f4 and f 41 The air-fuel ratio at the combustion chamber outlet section and the high-pressure turbine inlet section, f H2T (·) is the enthalpy-temperature conversion function.

[0093] Step A3) Establish a mainstream fast calculation model based on measured parameters.

[0094] The state-space model of the engine at a certain steady-state point (x0, u0, y0) is as follows:

[0095]

[0096] Δy=CΔx+DΔu

[0097] The parameters of the state-space model are chosen as follows: state variable x = [n L n H ] T Control quantity u = [W fb A8] T Output quantity y = [π fan π comp π HTurb π LTurb ] T .

[0098] Where, n L and n H For high and low voltage rotor speeds, W fb Fuel supply to the main combustion chamber, A8 is the throat area of ​​the tailpipe, π fan For fan voltage ratio, π comp π is the compressor pressure ratio. HTurb For the high-pressure turbine pressure ratio, π LTurb This refers to the pressure drop ratio of the low-pressure turbine.

[0099] Based on the engine similarity criterion, the state variable model covering the flight envelope is used with similarity-normalized parameters. The parameter similarity normalization is as follows:

[0100]

[0101]

[0102] Wherein, the subscript ds represents the engine design point parameters, and the engine state variable model expressed using similarity normalized parameters is as follows:

[0103]

[0104]

[0105] To solve for matrices A, B, C, and D, a combination of the small perturbation method and the fitting method can be used. Next, by changing the inlet temperature T2 of the engine core and the engine throat area A8, a large number of state-space models are combined into an LPV model at each steady-state point.

[0106]

[0107] Different throat areas A8 and different high-pressure speeds n of the engine H The state-space model below forms an LPV model of speed and pressure ratio. Then, using similarity conversion theory, this model is extended within the full envelope, describing the engine's LPV model as follows:

[0108]

[0109] Where x is the state variable, y is the output variable, and the subscript cor indicates the similarity conversion to ground point parameters.

[0110] Measured speed n at high and low pressure Hr n Lr With this support, it's no longer necessary to use A and B matrices to calculate high and low pressure speeds, thus eliminating the need to store A and B matrices; simultaneously, since the measured value of the total fan outlet pressure P is available... 22r Measured total pressure at turbine outlet P 5r Support for fan voltage ratio π fan and low-pressure turbine pressure ratio π GTurb There is no need to use C and D matrices for calculations, therefore the output quantity Only the compressor pressure ratio π needs to be selected. comp and the high-pressure turbine pressure ratio π HTurb With two parameters, matrices C and D only need to store half of the data.

[0111] From the above analysis, it can be seen that the simplified state-space expression of the LPV model in the mainstream fast calculation model based on measured parameters can be described by the following formula:

[0112]

[0113] Among them, state variables Control quantity u = [W fb A8] T Output

[0114] Since there is no need to store matrices A and B, and the required storage is... Compared to the C and D matrices of the LPV model, the matrix has reduced the amount of data by 1 / 2. Therefore, the amount of coefficient matrix that this mainstream fast calculation model needs to store is reduced by 2 / 3 compared to the LPV model.

[0115] Based on the measured parameters of the current engine, the pressure ratio parameters of each rotating component under this state can be obtained by combining the state-space model expression. Running the neural network model of the fan and compressor components yields the main parameters of the combustion chamber components. Then, using the thermal balance method, the main parameters of the entire engine can be obtained. The calculation process is as follows: Figure 4 As shown.

[0116] At this point, the mainstream fast calculation model has been established, which can provide mainstream parameters more quickly based on the current measured parameters of the engine.

[0117] The detailed steps of step B) are as follows:

[0118] The residuals of the equilibrium equation are described by equation (1), and equation (2) describes the deviation between the main parameters calculated by the model and the measured parameters:

[0119]

[0120]

[0121] Among them, J H and J L These represent the rotational inertia of the high-pressure and low-pressure rotor shafts, respectively, in N. HT N LT The power outputs of the high-pressure and low-pressure turbines are respectively, N F N C and N ex These are the power consumption of the fan, compressor, and related accessories, in W. g P represents the gas flow rate at each cross-section. s For the static pressure at each section, W cool This refers to the induced air flow rate.

[0122] At the ground point, when the characteristic adjustment parameters of the eight components of the engine degrade by 3% respectively, the changes in the residuals of the balance equation and the deviations between the mainstream parameters calculated by the model and the measured parameters are shown in Table 2.

[0123] Table 2. Changes in the residuals of the equilibrium equation when the component characteristic adjustment parameters degrade.

[0124]

[0125] As shown in Table 2, when the adjustment parameters of engine component characteristics degrade, their actual measured parameters also change. When input into the mainstream fast calculation model, it is assumed to be in a healthy state at this time. The mainstream parameters calculated by this model will cause changes in the residuals of the balance equation and their deviation from the measured parameters, with the overall trend of the residuals increasing. Therefore, it is feasible to use the changes in these residuals to estimate the adjustment parameters of component characteristics.

[0126] The detailed steps of step C) are as follows:

[0127] Step C1): When the performance of engine components degrades, the component characteristic adjustment parameters are estimated by using the change in residuals through extended Kalman filtering (EKF).

[0128] Define the component characteristic adjustment parameter that characterizes engine performance degradation as the efficiency coefficient SE of the rotating component. i and flow coefficient SW i as follows:

[0129]

[0130] Where, η i and W i These are the actual values ​​for component efficiency and flow rate. and W i * These are the rated values ​​for component efficiency and flow rate. The design values ​​will be used below. The subscript i represents the number of the rotating component.

[0131] Accordingly, the degradation amount of the component characteristic adjustment parameters is defined as follows:

[0132]

[0133] The calculation process for the extended Kalman filter is as follows:

[0134] (1) Initial update

[0135] (2) Time update equation

[0136] x k|k-1 =f(x) k-1 ,u k-1 )

[0137] P k|k-1 =A k P k-1 A k T +Q

[0138] (3) Measurement update equation

[0139] x k =x k|k-1 +K k [y k -g(x k|k-1 ,u k )]

[0140] P k =(IK k C k )P k|k-1

[0141] K k =P k|k-1 C k T (C k P k|k-1 C k T +R) -1

[0142] Among them, K k This is called Kalman gain, A k C k The Jacobian matrix is ​​calculated using the following formula:

[0143]

[0144] P k Let x be the state variable k The covariance.

[0145] (4) Let k = k + 1, and repeat step (2) until the filtering process ends.

[0146] Select the system state variable x at time k k =[SE1 SW1 SW2 SE3] T For health parameters, the control variable u of the system at time k k =[W fb A8] T Select y k =[e1 e2 e3 e4 e5 e6 e7] T , where e1 to e7 are the residuals of each equilibrium equation in the following formula.

[0147]

[0148] e2=(W g43 -W cool ) / W g4 -1

[0149] e3=(W g5 -W cool ) / W g44-1

[0150] e4 = W g9 / W g7 -1

[0151] e5 = P s16 / P s6 -1

[0152] e6=T 22 / T 22r -1

[0153] e7 = T5 / T 5r -1

[0154] The discrete form of the engine's nonlinear model and the Jacobian matrix A in the calculation process of the extended Kalman filter. k C k It can be written in the following form:

[0155] x k+1 =f(x) k ,u k )+ω k

[0156] y k =g(x k ,u k υ k )

[0157]

[0158] Where, ω k For the system noise, υ k This refers to the measurement noise of the system.

[0159] Based on the calculation process of the extended Kalman filter, the component characteristic adjustment parameters can be estimated using the residuals of the balance equation, and let Δh = [SE1 SW1 SW2 SE3]. T .

[0160] Step C2) cross-selects component characteristic adjustment parameters under different degradation modes and degrees of performance parameters at different state points to establish a network parameter training set. Without updating the network parameters, multiple independent OSELM offline networks are used to train the training set, and the output weights are assigned according to Euclidean distance. After weighted summation, the final output is the compensation matrix of the pressure ratio caused by performance degradation.

[0161] Assuming training samples Where x i For the input layer, y iFor the desired output, R is the set of real numbers, N is the total number of samples, the hidden layer activation function is g(x), the number of hidden layer nodes is L, and N0 partial samples are selected from the total training samples. Randomly select input weights ω and hidden layer biases b, and calculate the initial output matrix H0 of the hidden layer:

[0162]

[0163] Calculate its initial output weight β0:

[0164]

[0165]

[0166] Where P0 represents the initial value of the gain term during the algorithm iteration process, and Y0 represents the expected output in the initial training data block.

[0167] During the initialization process, OSELM uses initial samples to obtain initial network parameters.

[0168] Assuming a total of p OSELM offline networks are trained, the performance estimation is based on the pressure ratio compensation matrix obtained from each network. The residuals e of each equilibrium equation at this point are obtained. i =[e i1 e i2 e i3 e i4 e i5 e i6 e i7 ] T The subscript i represents the i-th OSELM offline network, which is the Euclidean distance to the ideal residual equilibrium point. Different OSELM network output weights λ i The assignment conditions are shown in the following formula:

[0169]

[0170] Step C3) updates the pressure ratio compensation matrix obtained by weighted summation of multiple independent OSELM offline network outputs to the mainstream fast calculation model of the engine. Under the condition of performance parameter degradation, the engine state-space model expression can be written in the following form:

[0171]

[0172] in, This is the compensation matrix for the pressure ratio caused by performance degradation.

[0173] Step C4) Design a network self-learning module for the offline OSELM network. Judge the network topology parameters by checking the residuals of the equilibrium equations output by the model. If the residuals are greater than a threshold, update the network topology parameters. The update process is as follows:

[0174] Let the training sample set transmitted in the k-th step be:

[0175]

[0176] Where, N j This represents the number of samples contained in the j-th data block.

[0177] Calculate the output matrix h of the hidden layer k Let the expected output at step k be y. k The expressions for both are as follows:

[0178]

[0179]

[0180] According to the recursive least squares algorithm, the output weights at step k are updated as follows:

[0181]

[0182]

[0183] Where, β k-1 P represents the updated output weight value obtained at the previous sample time step. k-1 This is the gain term from the previous time step.

[0184] During the OSELM network iteration update process, the network parameters are iteratively updated based on the data blocks obtained at each time step until the training ends.

[0185] The accuracy of neural network training is represented by RMSE (Root Mean Square Error), and its calculation formula is shown below:

[0186]

[0187] Among them, y net|i y is the output value obtained by the neural network. real|i This is the actual output value.

[0188] Based on the principles of weight self-updating and network self-learning, an OSELM-EKF adaptive correction scheme with self-learning capabilities was established. At ground point (H=0km, Ma=0, W…),… fb=2.6kg / s), the mainstream fast calculation model of the thermal balance method was simulated and verified before and after the addition of an adaptive correction scheme when the performance degraded. The simulation time was 20s, the engine sampling step size was 0.025s, and the component characteristic adjustment parameters were all 1 at the beginning of the simulation. At the 5th second, the performance degradation simulation was added, with the engine fan efficiency degrading by 3% and the high-pressure turbine efficiency degrading by 1%. The simulation results are as follows. Figure 5 As shown.

[0189] The RMSE values ​​of each OSELM network before and after the simulation performance degradation were analyzed and calculated. The results show that the RMSE value of the neural network after self-learning is smaller than that of the initial value, which proves that the designed network self-learning module has a certain role in ensuring the accuracy of the neural network.

[0190] In summary, the designed OSELM-EKF adaptive correction scheme with self-learning function can better reflect the actual working state of the engine and update the neural network topology parameters, which is of positive significance for the adaptive correction of mainstream fast calculation models.

[0191] It should be noted that the above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations and substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive correction method for mainstream fast calculation models of aero-engines based on measured parameters, characterized in that, Includes the following steps: A neural network model is used to replace the aerodynamic and thermodynamic model of the cold-end rotating component, and the neural network model is trained using a dataset consisting of the input and output parameters of the cold-end rotating component. Based on the measured parameters of the current engine and combined with the local component equilibrium equations constructed from the engine state-space model, the pressure ratio parameters of each rotating component under this state are obtained. The trained neural network model is run to obtain the main parameters of the cold-end rotating component, and the combustion chamber outlet temperature and high-pressure turbine inlet temperature are calculated using the thermal balance method. The main parameters of the engine, including the inlet and outlet of the hot-end rotating component, are obtained, and a fast calculation model of the engine's main parameters is established. When the performance of engine components degrades, the change in the residual of the equilibrium equation is used to estimate the component characteristic adjustment parameters through extended Kalman filtering; different state points are selected to obtain the component characteristic adjustment parameter values ​​under different degradation modes and different degrees of performance parameters to form a training set; The initial network parameters of the OSELM offline network are obtained by training multiple independent OSELM offline networks using the training set without updating the network parameters. The pressure ratio compensation matrix obtained by weighted summation of multiple independent OSELM offline network outputs is updated to the mainstream fast calculation model of the engine. A network self-learning module is designed to update the network topology parameters when the accuracy of the OSELM offline network is insufficient. The residuals of the balance equations output by the engine state space model are used to judge the network topology parameters. If the residuals of the balance equations output by the engine state space model are greater than a threshold, the OSELM offline network topology parameters are updated. Among them, based on aerodynamic thermodynamics, the combustion chamber outlet temperature T4 and the high-pressure turbine inlet temperature T are calculated using the heat balance method. 41 The process can be described in the following form: (1) Calculate the compressor outlet enthalpy H3 using the compressor outlet section temperature T3: H3=f T2H (0,T3) Among them, f T2H (·) represents the temperature-enthalpy conversion function; (2) Calculate the combustion chamber outlet enthalpy H4 and the high-pressure turbine inlet enthalpy H after cold air mixing. 41 : Among them, the low calorific value H of fuel oil μ =42900KJ / kg, η B For combustion efficiency, a -导冷气 W3 is the proportion of air drawn into the high-pressure turbine guide vane in the compressor outlet flow rate, W4 is the compressor outlet air flow rate, and W5 is the combustion chamber outlet air flow rate. 41 W is the airflow rate at the high-pressure turbine inlet after cold air mixing. fb Main fuel flow; (3) Utilizing the enthalpy H4 at the combustion chamber outlet and the enthalpy H at the high-pressure turbine inlet 41 Calculate the combustion chamber outlet temperature T4 and the high-pressure turbine inlet temperature T. 41 : Among them, f4 and f 41 The air-fuel ratio at the combustion chamber outlet section and the high-pressure turbine inlet section, f H2T (·) is the enthalpy-temperature conversion function.

2. The method for constructing a mainstream rapid calculation model of aero-engines based on measured parameters according to claim 1, characterized in that, The input parameters for the input layer of the neural network model for the fan component are selected as low-pressure speed n. L Total inlet temperature T2, fan pressure ratio π fan The output parameters of the output layer are selected based on the total temperature T at the fan outlet. 22 The calculation process is expressed by the following formula: T 22 =f1(n L ,T2,π fan ) Where f1(·) is the network function for the fan component.

3. The method for constructing a mainstream fast calculation model of aero-engines based on measured parameters according to claim 1, characterized in that, Input parameters for the neural network model of compressor components: high pressure speed n H Imported total temperature T 25 Import total pressure P 25 Compressor pressure ratio π comp The output parameters for the output layer are selected as follows: compressor outlet total temperature T3, compressor outlet flow rate W3, and interstage bleed air temperature T required for high and low pressure turbine airflow mixing calculations. 27 The calculation process is expressed by the following formula: [T 27 ,T3,W3]=f2(n H ,T 25 ,P 25 ,π comp ) Where f2(·) is the compressor component network function.

4. The method for constructing a mainstream fast calculation model of aero-engines based on measured parameters according to claim 1, characterized in that, The updated state-space expression for the engine state-space model is: Among them, state variables Control quantity u = [W fb A8] T Output n Hr n Lr These are the measured speed values ​​for high and low pressure applications, π comp π HTurb These are the compressor pressure ratio and the high-pressure turbine pressure ratio, respectively, W fb Fuel supply to the main combustion chamber, A8 is the throat area of ​​the tailpipe, n H For low-pressure rotor speed, This is the compensation matrix for the pressure ratio caused by performance degradation.

5. The method for constructing a mainstream rapid calculation model of aero-engines based on measured parameters according to claim 1, characterized in that, The process of updating network topology parameters is described by the following formula: Where, β k and β k-1 P represents the output weight update values ​​obtained at the current time step and the previous time step, respectively. k and P k-1 h represents the gain terms at the current time and the previous time, respectively. k This is the output matrix of the hidden layer at the current time. During the OSELM network iteration update process, the network parameters are iteratively updated based on the data blocks obtained at each time step until the training ends.

Images