Online Identification Method, Device and Performance Parameter Perception Method of the State Space Model Matrix of Aero-Engines Based on Digital Twin

Abstract

The present invention discloses an online identification method for the state space model matrix of an aero-engine based on digital twin, belonging to the technical field of aero-engine control. The online identification method for the state space model matrix of the aero-engine in the present invention includes: S1, constructing a past and future row full-rank Hankel matrix according to the finite input-output data set of the auxiliary airborne model within a current finite time period; S2, constructing an optimal prediction output estimator; S3, obtaining a future f-step prediction output equation according to the optimal prediction output estimator, and then obtaining the state space model matrix of the aero-engine by solving the simultaneous equations of the estimated output values and the actual output values for the future k+1 and k+2 steps. The present invention also discloses an online identification device for the state space model matrix of an aero-engine based on digital twin and a method for perceiving the performance parameters of an aero-engine. The present invention can realize high-precision perception of the performance parameters of an aero-engine in real time online.

Description

Technical Field

[0001] The present invention belongs to the technical field of aeroengine control, and particularly relates to an online identification method and device for a state space model matrix of an aeroengine and a performance parameter perception method. Background Art

[0002] The airborne adaptive model is an important technical means for perceiving the performance parameters of modern aeroengines. It is indispensable in the intelligent control and health management system of aeroengines. In order to realize the performance parameter estimation of the whole life cycle of an aeroengine, the current airborne adaptive model mainly consists of an airborne model + a Kalman filter: The Kalman filter first estimates the degradation amount of health parameters according to the measurement deviation between the airborne model and the real engine, and then uses this degradation amount of health parameters to update the airborne model, so as to adapt to the performance degradation of engine components in the whole life cycle and realize the accurate estimation of working parameters and key performance parameters.

[0003] At present, from the perspective of real-time performance, designing an airborne adaptive model based on a linear Kalman filter is the preferred choice. However, the linear Kalman filter needs to know the linear state space model matrix of the current steady-state operating point of the engine at all times. The existing methods mainly solve the state space model matrix of the standard engine operating point by using methods such as the small perturbation method and the fitting method offline and then call it online. This method has a low envelope coverage rate and poor accuracy. A Chinese invention patent (Sheng Hanlin et al., A method and device for model-based aeroengine performance recovery control, CN202110958916.8) discloses an online linearization method for an airborne model based on frequency analysis. The nonlinear system is represented by a low-frequency dynamic state and a high-frequency dynamic state, and then the Taylor series expansion of the high-frequency state dynamics equation in deviation form is carried out to obtain a linearized state space model. The state matrix can be updated in real time online, but there are many high-dimensional inverse matrix calculations in the whole calculation process, with high complexity, and it needs to be recalculated every time during flight, resulting in serious waste of software / hardware computing resources. Therefore, there are still many technical deficiencies in the current engine performance parameter estimation method based on the airborne adaptive model. Summary of the Invention

[0004] The purpose of the present invention is to overcome the technical problem of insufficient performance perception accuracy existing in the existing airborne adaptive model based on a linear Kalman filter, and provide an online identification method for the state space model matrix of an aeroengine based on digital twin, which can perform high-precision online identification of the state space model matrix of an aeroengine, and further can realize high-precision perception of the performance parameters of the aeroengine in real time online.

[0005] The technical solution proposed by the present invention is specifically as follows:

[0006] Online identification method for the state space model matrix of an aeroengine based on digital twin. The control input quantity u and the health parameter excitation signal h are input online and in real time into an auxiliary airborne model identical to the host airborne model to obtain the corresponding output data of the auxiliary airborne model. And based on the input-output data of the auxiliary airborne model, the following method is used for online identification of the state space model matrix of the aeroengine:

[0007] S1. Construct past and future row full-rank Hankel matrices according to the finite input-output data set of the auxiliary airborne model within a current finite time period:

[0008]

[0009]

[0010] where X f is the Hankel matrix of future state variables; U p and U f are the Hankel matrices of past and future control variables respectively; Y p and Y f are the Hankel matrices of past and future output variables respectively; p represents the past time domain; f represents the future time domain; n x represents the dimension of the state variable x, n u represents the total dimension of the control variable u and the health parameter h, n y represents the dimension of the output variable y;

[0011] S2. Construct an optimal prediction output estimator in the following form:

[0012]

[0013] where represents the future estimated output of the optimal prediction output estimator; the coefficient matrices and are obtained by solving the least squares problem of the future estimated output and the future actual output Y f : ; is the Frobenius 2-norm;

[0014] S3. Obtain the future f-step prediction output equation according to the optimal prediction output estimator:

[0015]

[0016] where m is the control time domain;

[0017] Then, by solving the following estimated output values for the next k+1 and k+2 steps and the actual output values y k+1 、y k+2 of the simultaneous equations, the state matrix A, input matrix B, output matrix C, and transfer matrix D of the aero-engine state space model are obtained:

[0018]

[0019] Preferably, the following method is used to solve the least squares problem:

[0020] Step 1: Transform the solution of the least squares problem into the orthogonal projection from the row space of matrix Y f to the matrix [W p U f T , that is:

[0021]

[0022] where " / " represents orthogonal projection;

[0023] Step 2: Use the QR decomposition method to decompose the following combined Hankel matrix:

[0024]

[0025] where is an upper triangular matrix, is an orthogonal matrix;

[0026] Step 3: Obtain the values of the coefficient matrices L w and L u through the following formula:

[0027]

[0028] where the superscript represents the generalized inverse.

[0029] Further preferably, the input-output data pair matrix R is updated in real time using the auxiliary airborne model in real time, and the updated matrix R is directly used to calculate the values of the coefficient matrices L w and L u according to the formula in Step 3.

[0030] Based on the same inventive concept, the following technical solutions can also be obtained:

[0031] An on-line identification device for the matrix of the aero-engine state space model based on digital twin, comprising:

[0032] ​An auxiliary airborne model, which is the same as the main airborne model, is used to obtain corresponding output data according to the control input quantity u and the health parameter excitation signal h input online in real time;

[0033] An online finite data-driven identification model is used to perform online identification of the state space model matrix of an aeroengine according to the input-output data of the auxiliary airborne model through the following method:

[0034] S1. Construct a past and future row full-rank Hankel matrix according to the finite input-output data set of the auxiliary airborne model within a current finite time period:

[0035]

[0036]

[0037]

[0038] where X f is the Hankel matrix of future state variables; U p and U f are the Hankel matrices of past and future control variables respectively; Y p and Y f are the Hankel matrices of past and future output variables respectively; p represents the past time domain; f represents the future time domain; n x represents the dimension of the state variable x, n u represents the total dimension of the control variable u and the health parameter h, n y represents the dimension of the output variable y;

[0039] S2. Construct an optimal prediction output estimator in the following form:

[0040]

[0041] where represents the future estimated output of the optimal prediction output estimator; the coefficient matrices and are obtained by solving the least squares problem of the future estimated output and the future actual output Y f : obtained; is the Frobenius 2-norm;

[0042] S3. Obtain a future f-step prediction output equation according to the optimal prediction output estimator:

[0043]

[0044] where m is the control time domain;

[0045] Then, by solving the following estimated output values for the next k+1 and k+2 steps and the actual output value y k+1 , y k+2 , the state matrix A, input matrix B, output matrix C, and transfer matrix D of the aero-engine state space model are obtained by solving the simultaneous equations:

[0046]

[0047] Preferably, the following method is used to solve the least squares problem:

[0048] Step 1: Transform the solution of the least squares problem into the orthogonal projection from the row space of matrix Y f to the matrix [W p U f T , that is:

[0049]

[0050] where " / " represents orthogonal projection;

[0051] Step 2: Decompose the following combined Hankel matrix using the QR decomposition method:

[0052]

[0053] where is an upper triangular matrix, is an orthogonal matrix;

[0054] Step 3: Obtain the values of the coefficient matrices L w and L u using the following formula:

[0055]

[0056] where the superscript represents the generalized inverse.

[0057] Further preferably, the input-output data pair matrix R is updated in real time using the auxiliary airborne model in real time, and the updated matrix R is directly used to calculate the values of the coefficient matrices L w and L u using the formula in Step 3.

[0058] ​A method for perceiving the performance parameters of an aeroengine based on digital twin uses an on-board host model to perceive the performance parameters of the aeroengine; meanwhile, according to the measurement deviation between the on-board host model and the real engine, a linear Kalman filter is used to estimate the degradation amount of the health parameters of the aeroengine, and the estimated degradation amount of the health parameters is used to update the on-board host model; the aeroengine state space model matrix used by the linear Kalman filter is obtained by the following method:

[0059] Step 1: During the first flight, use the method described in any one of claims 1 to 3 to obtain the aeroengine state space model matrix and provide it to the linear Kalman filter; meanwhile, cluster the input parameters of each working point of the aeroengine, and use the class centers of the obtained classes and their corresponding aeroengine state space model matrices as samples to construct a sample database.

[0060] Step 2: After the first flight, define the envelope area with frequent flights during the full life cycle of the engine as the high flight area, and the envelope area with fewer flights as the low flight area. Divide the samples in the current sample database into high flight area and low flight area, and use the input parameters of the high flight area samples in the sample database as the training input, and the corresponding aeroengine state space model matrix as the expected output to train the deep neural network to obtain the deep neural network model of the high flight area.

[0061] Step 3: During the next online flight, judge whether the minimum distance between the input parameters of the current working point of the aeroengine and the input parameters of each high flight area sample in the current sample database exceeds a preset threshold. If so, use the method described in any one of claims 1 to 3 to obtain the aeroengine state space model matrix and provide it to the linear Kalman filter; otherwise, input the input parameters of the current working point of the aeroengine into the deep neural network model of the high flight area, and provide the aeroengine state space model matrix output by the deep neural network model to the linear Kalman filter; meanwhile, cluster the input parameters of the current working point of the aeroengine, and use the obtained class center and its corresponding state space model matrix as new samples to update the sample database.

[0062] Step 4: After the flight, use the input parameters of the high flight area samples in the updated sample database as the training input, and the corresponding aeroengine state space model matrix as the expected output to retrain the deep neural network model offline to obtain the updated deep neural network model of the high flight area; Step 5: During the next flight, go to Step 3.

[0063] Preferably, the Gaussian mixture model clustering method is used for the clustering.

[0064] Preferably, the distance between the input parameters of the current operating point of the aero-engine and the input parameters of each high-flight region sample in the current sample database is the Mahalanobis distance.

[0065] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0066] The online identification method of the aero-engine state space model matrix based on digital twin proposed by the present invention can perform online high-precision identification of the aero-engine state space model matrix, solve the problem that it is difficult to obtain an accurate linear state space model within the full envelope range existing in the airborne adaptive model, and can realize online real-time high-precision perception of the performance parameters of the aero-engine;

[0067] The present invention further combines the above online identification method of the aero-engine state space model matrix based on digital twin with a data-driven deep neural network. Through offline / online learning of the engine state space model matrix, the real-time performance of calculating the state space model matrix in the high-flight envelope region of the engine is improved, and waste of computing resources is avoided, thus achieving a good coordination and unity of high precision and high real-time performance. Description of the Drawings

[0068] Figure 1 It is a schematic block diagram of the implementation principle of the aero-engine performance parameter perception method of the present invention;

[0069] Figure 2 It is a schematic diagram of the high / low flight region division principle. Detailed Embodiments

[0070] Aiming at the technical problem of insufficient performance perception accuracy existing in the existing airborne adaptive model based on a linear Kalman filter, the solution idea of the present invention is based on digital twin technology, and realizes high-precision online identification of the aero-engine state space model matrix through the limited input-output data set of the auxiliary airborne model, and further realizes online real-time high-precision perception of the performance parameters of the aero-engine.

[0071] For the convenience of public understanding, the technical solution of the present invention will be described in detail below with reference to the drawings:

[0072] As Figure 1As shown, on the basis of the original airborne adaptive model architecture based on a linear Kalman filter, the present invention adds an auxiliary airborne model. This auxiliary airborne model is the same as the main airborne model and serves to simulate the operating state of a standard engine. It can provide the output data of the standard engine in real time according to the same control input as the real engine. When a health parameter input excitation signal is given simultaneously, the output data corresponding to both the control input quantity u and the health parameter excitation signal h as inputs can be obtained. Then, based on the input (control input quantity u and health parameter excitation signal h) of the auxiliary airborne model collected over a period of time and the corresponding dataset of the auxiliary airborne model output, an online finite data-driven identification model is designed to perform online identification on this finite input-output dataset to obtain the engine state space model matrices (state matrix A, input matrix B, output matrix C, transfer matrix D): First, based on the finite input-output dataset, past and future full row rank Hankel matrices are constructed, and an optimal prediction output estimator similar in form to the state space model is designed. Then, the engine state space model matrices are solved according to the coefficient matrices of the optimal prediction output estimator. Finally, the linear Kalman filter estimates the engine health parameters according to this state space model matrix to update the main airborne model, thereby realizing the accurate estimation of the true values of performance parameters such as the thrust and surge margin of the engine throughout the full life cycle of the main airborne model.

[0073] The basic implementation principle of the online finite data-driven identification model proposed by the present invention is as follows:

[0074] Near a certain steady state point of the engine, a persistent excitation input signal u k (k = 1, 2, …, N) is continuously applied to the system, and the input / output (I / O) dataset D(U, Y ct ) with a sampling number of N is constructed into past and future full row rank Hankel matrices to obtain:

[0075]

[0076] Among them, X a,f is the Hankel matrix of future state variables; U p and U f are the Hankel matrices of past and future control variables respectively; Y p and Y f are the Hankel matrices of past and future output variables respectively; p represents the past time domain; f represents the future time domain, and generally p = f; x a is the augmented state vector of the high / low pressure shaft speed of the engine and all health parameter combinations, n x represents the dimension of x a , n u represents the dimension of the control variable u, n yRepresents the dimension of the output variable y (such as the fan speed Nf and the core speed Nc, the total pressure Pt17 at the inlet of the bypass duct, the total temperature Tt25, the total pressure Pt25 and the outlet total temperature Tt30, the static pressure Ps30 at the inlet of the high-pressure compressor, the total temperature Tt45 at the outlet of the high-pressure turbine, etc.).

[0077] The state-space model expression of the engine steady-state point is as follows:

[0078]

[0079] Among them, the matrices A, B, C, and D are the state matrix A, the input matrix B, the output matrix C, and the transfer matrix D respectively; k is the discrete time.

[0080] Therefore, according to the linear state-space model, Y f can be deduced from X f and U f The linear relationship is as follows:

[0081]

[0082] Among them, Φ x is the extended observable matrix; H u is a lower triangular Toeplitz matrix.

[0083] Then, according to the expression form of Y f in the above formula, a future optimal prediction output estimator is designed based on the past and future Hankel matrices, and the specific form is as follows:

[0084]

[0085] Among them, and respectively represent the online data-driven prediction estimator coefficient matrices of the past I / O data and the future input data; W p is the combination of the past input / output Hankel matrix, defined as

[0086]

[0087] Next, by constructing a least-squares problem regarding the future estimated output and the future actual output Y f to solve the values of L w and L u :

[0088]

[0089] Among them, is the Frobenius 2-norm.

[0090] According to matrix theory, the solution to the above equation can be understood as the matrix Y f The row space of the matrix [W p U f ] T The orthogonal projection is:

[0091]

[0092] Here, “ / ” indicates orthographic projection.

[0093] The following combined Hankel matrix is decomposed using the QR decomposition method

[0094]

[0095] in, is an upper triangular matrix, is an orthogonal matrix.

[0096] So, L w and L u The values are:

[0097]

[0098] The superscript represents the generalized inverse.

[0099] The prediction output equation for the next f steps is:

[0100]

[0101] in, m is the control time domain.

[0102] Therefore, the estimated output values for the next k+1 and k+2 steps are The actual output value y k+1 ,y k+2 By combining the equations of , we can get:

[0103]

[0104] According to this equation, the C matrix can be solved: Cx k+1 =L u (1:n y ,:)w p , and then, solve this equation according to the solved C matrix to obtain the A matrix: CAx k+1 =L u (n y +1:2n y ,:)w p ; D = L u (1:n y ,1:nu ) According to the solved C matrix, solve this equation to obtain the B matrix: CB = L u (n y +1:2n y ,1:2n u ).

[0105] Design a linear Kalman filter using the obtained A, B, C, and D matrices. Estimate the engine health parameters based on the residuals r of the measurement parameters (y m ) and the airborne model estimation parameters (y e ) to achieve the full-life cycle adaptability of the host airborne model to the engine and obtain performance parameter values such as thrust and surge margin required for engine control.

[0106] To adapt to the changes in the non-linear working state of the engine and avoid high-dimensional matrix calculations at each moment, the present invention further updates the R matrix using the newly obtained input-output data at each moment, and then directly calculates the coefficient matrices L w and L u values using the updated matrix R.

[0107] Generally, during the full-life cycle of the engine within the flight envelope area, some areas are flown frequently while some areas are flown less. Especially in the frequently flown areas, during each flight, the engine state space model matrix will be repeatedly calculated based on the online finite data-driven identification model. To avoid wasting computing resources due to repeated calculation of the state space model matrix at the same working point during each flight, the present invention further combines the above-mentioned online identification method of the aero-engine state space model matrix based on digital twin with a data-driven deep neural network. Through offline / online learning of the engine state space model matrix, the real-time performance of the calculation of the engine state space model matrix in the high flight envelope area is improved, and computing resource waste is avoided.

[0108] As Figure 1 shown, during the process of perceiving the aero-engine performance parameters, the present invention uses the host airborne model to perceive the aero-engine performance parameters; at the same time, based on the measurement deviation between the host airborne model and the real engine, uses a linear Kalman filter to estimate the degradation amount of the aero-engine health parameters, and uses the estimated degradation amount of the health parameters to update the host airborne model; the aero-engine state space model matrix used by the linear Kalman filter is obtained by the following method:

[0109] Step 1: During the initial flight, use the online finite data-driven identification model to obtain the aero-engine state space model matrix and provide it to the linear Kalman filter. At the same time, cluster the input parameters of each working point of the aero-engine, and construct a sample database with the class centers of the obtained classes and their corresponding aero-engine state space model matrices as samples.

[0110] Step 2: After the initial flight, define the envelope area with frequent flights during the entire life cycle of the engine as the high-flight area, and the envelope area with fewer flights as the low-flight area. Divide the samples in the current sample database into high-flight areas and low-flight areas (the division principle can be referred to Figure 2 ), and use the input parameters of the high-flight area samples in the sample database as the training input, and the corresponding aero-engine state space model matrix as the expected output to train the deep neural network to obtain the deep neural network model of the high-flight area.

[0111] Step 3: During the next online flight, determine whether the minimum distance between the input parameters of the current working point of the aero-engine and the input parameters of each high-flight area sample in the current sample database exceeds the preset threshold. If so, use the online finite data-driven identification model to obtain the aero-engine state space model matrix and provide it to the linear Kalman filter; otherwise, input the input parameters of the current working point of the aero-engine into the deep neural network model of the high-flight area, and provide the aero-engine state space model matrix output by the deep neural network model to the linear Kalman filter. At the same time, cluster the input parameters of the current working point of the aero-engine, and use the obtained class center and its corresponding state space model matrix as new samples to update the sample database.

[0112] Determining whether the minimum distance between the input parameters of the current working point of the aero-engine and the input parameters of each high-flight area sample in the current sample database exceeds the preset threshold is essentially to determine whether the current deep neural network model of the high-flight area is applicable to the current working point. The distance can adopt measurement methods such as Euclidean distance, Mahalanobis distance, Manhattan distance, etc. In this embodiment, the Mahalanobis distance is adopted. Various existing clustering methods can be used for clustering the data. In this embodiment, the Gaussian mixture model clustering method is adopted.

[0113] As Figure 1 shown, if the judgment is negative, switch 1 and 2 are both switched to the 1 state, and environmental and control input quantities u such as altitude (H), Mach number (Ma), environmental temperature (T0) / pressure (P0) are input into the deep neural network model, and the state space model matrix is extracted and given to the linear Kalman filter estimator.

[0114] If the judgment is yes, both switches 1 and 2 are turned to the 2 state, and the state space model matrix is solved by the online finite data-driven identification model;

[0115] Meanwhile, Gaussian mixture model clustering is performed on environmental and control input quantities such as altitude (H), Mach number (Ma), environmental temperature (T0) / pressure (P0), etc., and the obtained class centers and their corresponding state space model matrices are used as new samples to update the sample database, obtaining an updated sample database; Step 4, after the flight ends, using the input parameters of the high-flight region samples in the updated sample database as training inputs and the corresponding aeroengine state space model matrix as the desired output, the deep neural network model is retrained offline to obtain an updated deep neural network model for the high-flight region;

[0116] In the offline state, switch 3 is turned to the offline state, and the data samples in the updated sample database are divided into high / low flight regions, which can be divided manually or by software; then the high-flight region samples are used to retrain the deep neural network model offline, thereby obtaining an updated deep neural network model for the high-flight region;

[0117] Step 5, when flying again, go to Step 3.

[0118] Through the above method, during the cyclic flight process of the engine throughout its entire life cycle, the sample database is continuously updated. Finally, the division of the high / low flight regions tends to be stable, and the deep neural network model for the high-flight region also tends to be stable, enabling the rapid extraction of the state space model matrix of the high-flight region system.

Claims

1. An online identification method for the state space model matrix of an aeroengine based on digital twin, characterized in that, The control input quantity u and the health parameter excitation signal h are input online and in real time into an auxiliary aircraft-borne model identical to the main aircraft-borne model to obtain the corresponding output data of the auxiliary aircraft-borne model; and according to the input-output data of the auxiliary aircraft-borne model, the following method is used for online identification of the state space model matrix of the aeroengine: S1. Construct past and future row full-rank Hankel matrices according to the finite input-output data set of the auxiliary aircraft-borne model within a current finite time period: Among them, X f is the Hankel matrix of the future state variables; U p and U f are the Hankel matrices of the past and future control variables respectively; Y p and Y f are the Hankel matrices of the past and future output variables respectively; p represents the past time domain; f represents the future time domain; n x represents the dimension of the state variable x, n u represents the total dimension of the control variable u and the health parameter h, n y represents the dimension of the output variable y; S2. Construct an optimal prediction output estimator in the following form: Among them, represents the future estimated output of the optimal prediction output estimator; the coefficient matrices and are obtained by solving the least squares problem of the future estimated output and the future actual output Y f : are obtained; is the Frobenius 2-norm; S3. Obtain a future f-step prediction output equation according to the optimal prediction output estimator: Among them, m is the control time domain; Then, by solving the following equations for the estimated output values at the next k+1 and k+2 steps and the actual output values y k+1 , y k+2 , the state matrix A, input matrix B, output matrix C, and transfer matrix D of the aero-engine state space model are obtained:

2. The online identification method for the state space model matrix of an aero-engine based on digital twin according to claim 1, wherein Use the following method to solve the least squares problem: Step 1. Transform the solution of the least squares problem into the orthogonal projection from the row space of matrix Y f to the matrix [W p U f T , that is:​ where " / " represents orthogonal projection; Step 2. Decompose the following combined Hankel matrix by using the QR decomposition method: Among them, is an upper triangular matrix, is an orthogonal matrix; Step 3. Obtain the coefficient matrix L through the following formula w and the value of L u : where the superscript denotes the generalized inverse.

3. The online identification method of the state space model matrix of an aeroengine based on digital twin according to claim 2, wherein, The input-output data pair matrix R updated in real time using the auxiliary airborne model is updated in real time, and the coefficient matrix L is directly calculated according to the formula in step 3 using the updated matrix R w and L u values.

4. An online identification device for the state space model matrix of an aero-engine based on digital twin, characterized in that, including: An auxiliary aircraft-borne model, which is identical to the main aircraft-borne model and is used to obtain corresponding output data according to the control input quantity u and the health parameter excitation signal h input online and in real time; An online finite data-driven identification model, which is used for online identification of the state space model matrix of the aeroengine according to the input-output data of the auxiliary aircraft-borne model by the following method: S1. Construct past and future row full-rank Hankel matrices according to the finite input-output data set of the auxiliary aircraft-borne model within a current finite time period: Among them, X f is the Hankel matrix of the future state variables; U p and U f are the Hankel matrices of the past and future control variables respectively; Y p and Y f are the Hankel matrices of the past and future output variables respectively; p represents the past time domain; f represents the future time domain; n x represents the dimension of the state variable x, n u represents the total dimension of the control variable u and the health parameter h, n y represents the dimension of the output variable y; S2. Construct an optimal prediction output estimator in the following form: Among them, represents the future estimated output of the optimal prediction output estimator; the coefficient matrices and are obtained by solving the least squares problem of the future estimated output and the future actual output Y f : are obtained; is the Frobenius 2-norm; S3. Obtain a future f-step prediction output equation according to the optimal prediction output estimator: Among them, m is the control time domain; Then, by solving the following estimated output values for the next k+1 and k+2 steps and the actual output values y k+1 , y k+2 , the state matrix A, input matrix B, output matrix C, and transfer matrix D of the aero-engine state space model are obtained by solving the simultaneous equations:

5. The on-line identification device for the state space model matrix of an aero-engine based on digital twin according to claim 4, wherein Use the following method to solve the least squares problem: Step 1. Transform the solution of the least squares problem into the orthogonal projection of the row space of matrix Y f onto the matrix [W p U f T , that is:​ where " / " represents orthogonal projection; Step 2. Decompose the following combined Hankel matrix by using the QR decomposition method: Among them, is an upper triangular matrix, is an orthogonal matrix; Step 3. Obtain the coefficient matrix L through the following formula w and the value of L u as follows: where the superscript denotes the generalized inverse.

6. The online identification device for the state space model matrix of an aero-engine based on digital twin according to claim 5, characterized in that The input-output data pair matrix R updated in real time using the auxiliary airborne model is updated in real time, and the coefficient matrix L is directly calculated according to the formula in step 3 using the updated matrix R w and L u values 7. A method for perceiving the performance parameters of an aero-engine based on digital twin, characterized in that Use the online identification method for the state space model matrix of the aeroengine based on digital twin according to any one of claims 1 to 3, and use the main aircraft-borne model to perceive the performance parameters of the aeroengine; meanwhile, according to the measurement deviation between the main aircraft-borne model and the real engine, use a linear Kalman filter to estimate the degradation amount of the health parameters of the aeroengine, and use the estimated degradation amount of the health parameters to update the main aircraft-borne model; the state space model matrix of the aeroengine used by the linear Kalman filter is obtained by the following method: Step 1. During the initial flight, use the online identification method for the state space model matrix of the aeroengine based on digital twin to obtain the state space model matrix of the aeroengine and provide it to the linear Kalman filter; meanwhile, cluster the input parameters of each working point of the aeroengine, and construct a sample database with the class centers of the obtained classes and the corresponding state space model matrices of the aeroengine as samples; Step 2: After the initial flight, define the envelope area with frequent flights during the entire life cycle of the engine as the high-flight area, and the envelope area with fewer flights as the low-flight area. Divide the samples in the current sample database into high-flight and low-flight areas, and use the input parameters of the high-flight area samples in the sample database as the training input, and the corresponding aero-engine state space model matrix as the expected output to train the deep neural network, and obtain the deep neural network model for the high-flight area; Step 3: During the next online flight, determine whether the minimum distance between the input parameters of the current operating point of the aero-engine and the input parameters of each high-flight area sample in the current sample database exceeds the preset threshold. If so, use the online identification method of the aero-engine state space model matrix based on digital twin to obtain the aero-engine state space model matrix and provide it to the linear Kalman filter; Otherwise, input the input parameters of the current operating point of the aero-engine into the deep neural network model of the high-flight area, and provide the aero-engine state space model matrix output by the deep neural network model to the linear Kalman filter; Meanwhile, cluster the input parameters of the current operating point of the aero-engine, and use the obtained class centers and their corresponding state space model matrices as new samples to update the sample database; Step 4: After the flight, use the input parameters of the high-flight area samples in the updated sample database as the training input, and the corresponding aero-engine state space model matrix as the expected output to retrain the deep neural network model offline, and obtain the updated deep neural network model for the high-flight area; Step 5: During the next flight, go to Step 3.

8. The method for perceiving the performance parameters of an aero-engine based on digital twin according to claim 7, wherein Use the Gaussian mixture model clustering method for the clustering.

9. The method for perceiving the performance parameters of an aero-engine based on digital twin according to claim 7, characterized in that, The distance between the input parameters of the current operating point of the aero-engine and the input parameters of each high-flight area sample in the current sample database is the Mahalanobis distance.

Images