A knowledge- and data-driven approach to constructing an overall performance proxy model for aircraft engines

Abstract

The present invention discloses a method for constructing a knowledge and data dual-driven agent model for overall performance of an aero-engine, which relates to the field of numerical simulation technology. The construction process includes the following steps: using Latin hypercube sampling to generate an initial sample set, performing simulation iterative solutions on the sample points in the initial sample set one by one to obtain a converged sample set and train IVTX and ETDX agent models, which are connected in series to form a knowledge and data dual-driven agent model KDSM. For non-converged sample points, the KDSM model is used to predict the initial guess value of the engine performance simulation model, and simulation iterative solutions are performed to supplement the converged sample set and update the training agent model until no new converged sample points are generated, thereby obtaining the final KDSM. The agent model mentioned in the present invention is suitable for performance calculation scenarios of aero-engines of different configurations, has high prediction accuracy, strong generalization ability, can assist in improving the computational efficiency and convergence of the engine performance simulation model, and is flexible in application.

Description

Technical Field

[0001] The present invention relates to the field of numerical simulation technology, and in particular to a method for constructing a knowledge- and data-driven overall performance proxy model for an aero-engine. Background Art

[0002] Aero-engine overall performance simulation utilizes numerical simulation technology to predict and analyze engine performance under different operating conditions based on the engine's configuration, design parameters, and operating principles. This is crucial for the design, optimization, and control of engines and aircraft.

[0003] Existing engine overall performance simulation models primarily utilize zero-dimensional component-level modeling, establishing a nonlinear mathematical model based on the principles of engine aerodynamics and thermodynamics. The engine's steady-state and transient performance are then obtained by iteratively solving the engine's common operating equations through algorithms such as Newton-Raphson. This model is characterized by high accuracy and conformity to the physical operating principles of the engine. However, its limitations lie in the frequent non-convergence of model iterations over a wide operating range. Furthermore, since performance solutions require repeated iterations of the model for flow calculations, its real-time performance is poor. With the continuous development of modern aero-engine design, the engine configuration and operating principles are becoming increasingly complex, further increasing the complexity of numerical simulation models. Convergence issues and non-real-time characteristics are becoming more pronounced, making it difficult to meet the requirements of overall engine performance design optimization and real-time performance feedback control.

[0004] In recent years, surrogate models have been widely used in various fields to approximate complex simulation models to reduce computational costs and improve optimization efficiency. In the field of overall aeroengine performance simulation, existing surrogate models primarily employ data-driven modeling approaches. These methods generate a sample database based on the overall performance simulation model and apply methods such as kriging, radial basis function interpolation, support vector regression, and artificial neural networks for fitting, regression, and feature learning, thereby constructing a surrogate model that closely approximates the overall performance simulation model. Given input parameters such as engine design parameters and operating conditions, the trained surrogate model alone can rapidly obtain key engine performance parameters without repeatedly iterating the overall performance simulation model, while avoiding model non-convergence issues and thus enabling further application in engine performance optimization and control design.

[0005] However, purely data-driven surrogate models have numerous shortcomings and limitations in practical applications. First, due to the numerous design variables and highly nonlinear characteristics of aircraft engines, existing overall performance surrogate models, while capable of achieving relatively accurate approximations in some cases, still leave much room for improvement in accuracy and generalization. Consequently, the credibility of engine performance and optimization design results predicted by such surrogate models is difficult to guarantee. Second, although the established surrogate models offer high real-time responsiveness, they do not improve the convergence of the overall engine performance simulation model itself. During the sample generation step for surrogate model training, the overall performance simulation model is required to converge across all sample spaces with feasible solutions. This often requires researchers to adjust the initial guess of the engine based on different sample spaces, significantly increasing the difficulty and workload of surrogate model training. Furthermore, purely data-driven surrogate models often suffer from the problem of not satisfying physical constraints. When faced with operating conditions or design parameter combinations that would prevent the actual engine from operating properly, the surrogate model will still output performance predictions without providing information on their physical feasibility. Finally, because the training and prediction computation time of surrogate models is positively correlated with the number of predictor variables, surrogate models struggle to demonstrate significant computational efficiency advantages when it comes to simultaneously predicting all detailed engine performance parameters. For example, a comprehensive engine performance simulation model provides comprehensive performance information from the overall engine to the component level, as well as over a hundred parameters such as cross-sectional thermal state. Consequently, surrogate models are often used only to predict key engine performance parameters, such as thrust and fuel consumption, further limiting the scope and application of traditional surrogate models in engine performance simulation.

[0006] In response to the above problems, there is an urgent need to improve the traditional aircraft engine overall performance proxy model and build a proxy model that can maintain high calculation accuracy, significantly improve the engine performance calculation speed, fully meet the engine physical constraints, improve model convergence, and have a wider range of applications. Summary of the Invention

[0007] To address the aforementioned issues and address the shortcomings and deficiencies of existing technologies, the present invention aims to provide a method for constructing a knowledge- and data-driven proxy model for overall aircraft engine performance. By incorporating fundamental engine physics and domain knowledge, this dual-driven approach is employed to train and apply the proxy model, thereby improving the predictive performance and application scope of the proxy model. Specifically, this method includes two proxy model training methods: predicting guess vectors and revising them; two proxy model application methods: "prediction-performance" and "prediction-iteration-performance"; and a program algorithm flow for training and applying the entire proxy model.

[0008] A method for constructing a knowledge- and data-driven proxy model for overall aircraft engine performance includes the following steps:

[0009] S1. Obtain the inlet total temperature and engine control law variables and the value range of the inlet total temperature and engine control law variables ;

[0010] S2, using Latin hypercube sampling method, in the range of Generate initial sample set , the initial sample set Sample points in Represents a value combination of each independent variable;

[0011] S3, based on the engine performance simulation model, the initial sample set Sample points in Perform simulation iterations one by one to obtain converged sample set A, converged sample set B and non-converged sample set The converged sample set A is composed of the converged sample point a, the converged sample set B is composed of the converged sample point b, and the non-converged sample set From the non-convergent sample points constitute;

[0012] S4, based on the initial training of the convergent sample set A The proxy model is initially trained based on the converged sample set B to obtain the initial Proxy Model, Initial Proxy Model and Initial The agent models are connected in series to form the initial knowledge data dual-driven agent model ;

[0013] S5. Using initial knowledge data dual-driven agent model Predicting the non-convergent sample set Non-converged sample points The initial value of the guess vector , and based on the engine performance simulation model, each non-convergent sample point is simulated and solved iteratively one by one, and the convergent sample set A is composed of the convergent sample point a, the convergent sample set B is composed of the convergent sample point b, and the non-convergent sample set From the non-convergent sample points Construct the convergent sample set A , convergent sample set B and non-convergent sample sets ;

[0014] S6, based on the convergent sample set A Update training to get updated Proxy model, based on convergent sample set B Update training to get updated Proxy Model, Update Proxy Models and Updates Agent model composition and update knowledge data dual-driven agent model ;

[0015] S7. Repeat S5-S6 until no new convergence sample point a or convergence sample point b appears in S5. Terminate the iteration and obtain the knowledge- and data-driven overall performance proxy model of the aero-engine.

[0016] Preferably, the initial value of the guess vector for the simulation iteration solution in S3 is ,According to different convergence situations, adopt corresponding processing strategies, as follows:

[0017] For the sample points where the calculation converges, the strategy is as follows:

[0018] S301. Record the values ​​of independent variable combinations , iterative convergence guess vector solution , the residual vector obtained by the initial iterative calculation of the simulation model And the guess vector correction ;

[0019] S302, will As input, As output, the combination is used for The convergence sample point a of the surrogate model training is in the form of , and add it to the converged sample set The convergent sample set A is in the middle;

[0020] S303, will As input, As output, the combination is used for The convergence sample point b of the agent model training is in the form of , and add it to the converged sample set The middle one is the convergent sample set B;

[0021] For sample points where the calculation does not converge, the strategy is as follows:

[0022] Add it to the non-convergent sample set , as the sample set for the next round of simulation iteration.

[0023] In S5, for the sample points where the calculation converges, the strategy is as follows:

[0024] S501, record the value of the independent variable combination IV i ', iterative convergence guess vector solution , the residual vector obtained by the initial iterative calculation of the simulation model And the guess vector correction ;

[0025] S502, will As input, As output, the combination is used for The new convergence sample point a for the agent model training is in the form of , and add it to the converged sample set That is, the convergent sample set A middle;

[0026] S503, will As input, As output, the combination is used for The new convergence sample point b for the agent model training is in the form of , and add it to the converged sample set That is, the convergent sample set B middle;

[0027] For sample points where the calculation does not converge, the strategy is as follows:

[0028] Add it to the new non-converged sample set , as the sample set for the next round of simulation iteration.

[0029] Preferably, the knowledge and data dual-driven aircraft engine overall performance agent model is the latest updated The KDSM proxy model is constructed by connecting the proxy model with the ETDX proxy model.

[0030] The preferred application process method of the knowledge and data dual-driven aircraft engine overall performance agent model is as follows:

[0031] Step 1: Given the engine control law variables and the independent variable value combination of the inlet total temperature ;

[0032] Step 2: As The input of the proxy model is used to predict the guess value vector ;

[0033] Step 3: Based on the guess vector , call the engine performance simulation model to perform a flow calculation and calculate the error vector ;

[0034] Step 4: As The input of the proxy model is used to predict the correction of the guess vector , thus obtaining Output of the proxy model ;

[0035] Step 5: Based on Output of the proxy model , Used to call engine performance simulation models for performance prediction or iterative solution.

[0036] Preferably, the The specific contents used to call the engine performance simulation model for performance prediction or iterative solution are as follows:

[0037] Prediction-Performance: Based on output volume The engine matching working state is directly predicted based on its value, and the engine performance simulation model is called to perform a flow path calculation to predict the engine performance.

[0038] Prediction-Iteration-Performance: Output As the initial value of the iterative solution of the engine performance model simulation, iterative calculation is performed, and the convergence The determined engine matching operating conditions calculate accurate engine performance.

[0039] In summary, the adaptive cycle aircraft engine configuration and simulation model of the present invention are applicable to aircraft engines of various configurations. Compared with traditional aircraft engine overall performance proxy models, the present invention has the following advantages:

[0040] 1. The knowledge- and data-driven agent model training and application method proposed in this invention has low training difficulty and cost, and its prediction results are consistent with engine performance principles. It can significantly improve the model's prediction accuracy and generalization ability, and is applicable to a variety of complex working conditions and aircraft engines of different configurations.

[0041] 2. The proposed proxy model accurately predicts engine state guess vectors, significantly reducing the number of simulation model iterations and improving model convergence. Furthermore, its application method provides a wide range of engine performance information, from the overall to the component level, including flow path thermal parameters, component operating status, and performance parameters. This supports detailed engine operating state assessment and optimization constraints, expanding the application scope of the proxy model to meet more complex and diverse engineering needs.

[0042] 3. The proposed proxy model offers flexible and extensive application methods, including "prediction-performance" and "prediction-iteration-performance." The "prediction-performance" method boasts extremely fast computational speed and is suitable for rapid design optimization of engine performance where a certain degree of error is acceptable, as well as for scenarios requiring high engine performance computational response. The "prediction-iteration-performance" method significantly reduces computational cost while maintaining the same accuracy as traditional simulation models, making it suitable for high-precision optimization and convergence assistance for iterative solutions of engines with a wide design range.

[0043] The technical method of the present invention is further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The flow chart of the knowledge and data dual-driven aircraft engine overall performance agent model training process of the present invention;

[0045] Figure 2 This is a flowchart of the application of the knowledge and data dual-driven aviation engine overall performance proxy model of the present invention;

[0046] Figure 3 This is a schematic structural diagram of the adaptive cycle engine of the present invention;

[0047] Figure 4 This is a comparison chart of relative errors in thrust prediction of the proxy model of the present invention;

[0048] Figure 5 This is a comparison chart of relative errors in fuel consumption rate prediction using the proxy model of the present invention;

[0049] Figure 6 It is the prediction accuracy of the guess value vector of the proxy model of the present invention.

[0050] Reference numerals:

[0051] 1. Front fan; 2. Rear fan; 3. Mode selector valve; 4. Front duct ejector; 5. Outer duct nozzle; 6. Rear duct ejector; 7. Main nozzle; 8. Core engine drive fan; 9. Compressor; 10. Main combustion chamber; 11. High-pressure turbine; 12. Low-pressure turbine; 13. Afterburner; 14. First duct; 15. Second duct; 16. Third duct. DETAILED DESCRIPTION

[0052] The technical method of the present invention is further described below through the accompanying drawings and embodiments. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and values ​​described in these embodiments do not limit the scope of this application.

[0053] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0054] Technologies, systems, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0055] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0056] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0057] Compared to current problems that lack sufficient mathematical and physical theoretical support or are extremely difficult and costly to solve in a single attempt, the problem of solving the overall performance of aircraft engines actually has relatively well-developed and mature physical laws and scientific theories, and its performance simulation model has strong real-time performance when executing flow calculations in a single attempt. The main challenge is the need to iteratively solve a set of nonlinear co-working equations, which makes convergence significantly dependent on the initial guess value; at the same time, the need to repeatedly call flow calculations greatly increases the amount of calculation, resulting in a significant decrease in solution efficiency. Traditional engine overall performance proxy models are purely data-driven and have many limitations. Similar to physical information neural networks, the solution of the present invention introduces the basic physical laws of the engine and the advantages of domain knowledge, and adopts a knowledge- and data-driven approach to train and apply the proxy model, thereby improving the predictive performance and applicability of the engine overall performance proxy model.

[0058] First, unlike traditional approaches that directly predict engine performance using a proxy model of overall engine performance, this technical solution first predicts the engine state guess vector and then uses the engine simulation model algorithm to call the flow path calculation once or very rarely, thereby indirectly achieving a solution for quickly and accurately predicting engine performance.

[0059] On the other hand, Based on the proxy model, we further constructed and trained The proxy model predicts the correction value of the guess value vector based on the residual vector of the common working balance equation obtained by the single flow path calculation of the engine, thereby simulating the process of iterative correction of the guess value through the Newton-Raphson and other solution algorithms in the overall performance simulation model of the engine, thereby further improving the prediction accuracy of the guess value vector.

[0060] Example 1

[0061] A method for constructing a knowledge- and data-driven proxy model for overall aircraft engine performance includes the following steps:

[0062] S1. Obtain the inlet total temperature and engine control law variables and the value range of the inlet total temperature and engine control law variables The inlet total temperature and engine control law variables are the independent variables of the knowledge and data dual-driven aviation engine overall performance proxy model. The independent variables are the important design parameters that affect the engine's joint working matching. The value range of the inlet total temperature and engine control law variables is The upper and lower bounds of the values ​​of each independent variable within the design allowable range.

[0063] S2, using Latin hypercube sampling method, in the range of Generate initial sample set , the initial sample set Sample points in Represents a value combination of each independent variable;

[0064] S3, based on the engine performance simulation model, the initial sample set Sample points in Perform simulation iterations one by one to obtain converged sample set A, converged sample set B and non-converged sample set The converged sample set A is composed of the converged sample point a, the converged sample set B is composed of the converged sample point b, and the non-converged sample set From the non-convergent sample points constitute;

[0065] Furthermore, the initial value of the guess vector for the simulation iteration solution in S3 is ,According to different convergence situations, adopt corresponding processing strategies, as follows:

[0066] For the sample points where the calculation converges, the strategy is as follows:

[0067] S301. Record the values ​​of independent variable combinations , iterative convergence guess vector solution , the residual vector obtained by the initial iterative calculation of the simulation model And the guess vector correction ;

[0068] S302, will As input, As output, the combination is used for The convergence sample point a of the surrogate model training is in the form of , and add it to the converged sample set The convergent sample set A is in the middle;

[0069] S303, will As input, As output, the combination is used for The convergence sample point b of the agent model training is in the form of , and add it to the converged sample set The middle one is the convergent sample set B;

[0070] For sample points where the calculation does not converge, the strategy is as follows:

[0071] Add it to the non-convergent sample set , as the sample set for the next round of simulation iteration.

[0072] S4, based on the initial training of the convergent sample set A The proxy model is initially trained based on the converged sample set B to obtain the initial Proxy Model, Initial Proxy Model and Initial The agent models are connected in series to form the initial knowledge data dual-driven agent model , where the proxy model construction method used in training can be selected from one of Kriging, radial basis function interpolation, support vector regression, artificial neural network, etc.

[0073] initial Proxy Model and Initial The agent models are connected in series to form the initial knowledge data dual-driven agent model In fact, it is a series combination at the application level, such as Figure 2 As shown in the application flow chart, the data IV of KDSM is actually the input of IVTX. IVTX outputs X and calculates E, which is used as the input of ETDX, and outputs DX. The final X+DX is the output of KDSM.

[0074] That is, the input of KDSM is IV and the output is X+DX, but in fact the internal prediction is completed by using two proxy models in series.

[0075] S5. Using initial knowledge data dual-driven agent model Predicting the non-convergent sample set Non-converged sample points The initial value of the guess vector Based on the engine performance simulation model, each non-convergent sample point is simulated and iterated again to obtain new convergent sample point a, convergent sample point b and non-convergent sample point , update the converged sample set A, converged sample set B and non-converged sample set Get the convergent sample set A , convergent sample set B and non-convergent sample sets ;

[0076] In S5, for the sample points where the calculation converges, the strategy is as follows:

[0077] S501, record the value of the independent variable combination IV i ', iterative convergence guess vector solution , the residual vector obtained by the initial iterative calculation of the simulation model And the guess vector correction ;

[0078] S502, will As input, As output, the combination is used for The new convergence sample point a for the agent model training is in the form of , and add it to the converged sample set That is, the convergent sample set A middle;

[0079] S503, will As input, As output, the combination is used for The new convergence sample point b for the agent model training is in the form of , and add it to the converged sample set That is, the convergent sample set B middle;

[0080] S503, will As input, As output, the combination is used for Convergence sample points for surrogate model training , and add it to the converged sample set That is, the convergent sample set B middle.

[0081] For sample points where the calculation does not converge, the strategy is as follows:

[0082] Add it to the new non-converged sample set , as the sample set for the next round of simulation iteration.

[0083] S6, based on the convergent sample set A Update training to get updated Proxy model, based on convergent sample set B Update training to get updated Proxy Model, Update Proxy Models and Updates Agent model composition and update knowledge data dual-driven agent model ;

[0084] S7. Repeat S5-S6 until no new convergence sample point a or convergence sample point b appears in S5. Terminate the iteration and obtain the knowledge- and data-driven overall performance proxy model of the aero-engine.

[0085] Furthermore, the knowledge and data-driven overall performance proxy model of aero-engines is the latest updated The KDSM proxy model is constructed by connecting the proxy model with the ETDX proxy model.

[0086] Furthermore, the application process method of the knowledge and data-driven aircraft engine overall performance agent model is as follows:

[0087] Step 1: Given the engine control law variables and the independent variable value combination of the inlet total temperature ;

[0088] Step 2: As The input of the proxy model is used to predict the guess value vector ;

[0089] Step 3: Based on the guess vector , call the engine performance simulation model to perform a flow calculation and calculate the error vector ;

[0090] Step 4: As The input of the proxy model is used to predict the correction of the guess vector , thus obtaining Output of the proxy model ;

[0091] Step 5: Based on Output of the proxy model , Used to call engine performance simulation models for performance prediction or iterative solution.

[0092] Furthermore, the The specific contents used to call the engine performance simulation model for performance prediction or iterative solution are as follows:

[0093] Prediction-Performance: Based on output volume The engine matching working state is directly predicted based on its value, and the engine performance simulation model is called to perform a flow path calculation to predict the engine performance.

[0094] Prediction-Iteration-Performance: Output As the initial value of the iterative solution of the engine performance model simulation, iterative calculation is performed, and the convergence The determined engine matching operating conditions calculate accurate engine performance.

[0095] The proxy model mentioned in the present invention has high prediction accuracy, strong generalization ability, low training difficulty and cost, and the prediction results are consistent with the engine performance principles. It can also assist in improving the computational efficiency and convergence of the engine performance simulation model. It is flexible in application and can be applied to performance calculation scenarios of aircraft engines of different configurations.

[0096] Traditional methods using surrogate models to directly predict key engine performance often struggle to achieve ideal accuracy, primarily due to the inherently nonlinear nature of the engine performance problem. In practice, the process from engine design parameters to engine performance can be decoupled into two steps: first, iteratively solving for a guess vector based on the engine design parameters that ensures the engine meets common operating conditions; second, calculating engine performance based on this guess vector.

[0097] The present invention proposes The surrogate model essentially takes on the fitting replacement of the first step, specifically using The proxy model predicts the initial guess value and The proxy model corrects the prediction results of the guess vector. On the one hand, the first step after decoupling has a lower degree of nonlinearity; on the other hand, The proxy model simulates the Newton-Raphson and other solution algorithms, and can use the information provided by the error vector to reasonably correct the guess vector. Therefore, the KDSM proxy model can achieve extremely high accuracy in predicting the guess vector. Using its predicted guess vector as the iterative initial value of the simulation model in the "prediction-iteration-performance" application method can significantly improve the convergence of the simulation model and reduce the number of iterations. For the second step, the solution of the present invention makes full use of the existing simulation model algorithm and directly performs a flow path calculation based on the guess vector after prediction correction. The engine performance error obtained is only derived from the guess vector prediction error in the first step. Therefore, The “prediction-performance” application method of the surrogate model can also achieve ideal engine performance prediction accuracy.

[0098] Compared to traditional approaches that directly predict engine performance, the proxy model proposed in this paper focuses on accurately predicting the engine state guess vector, while actual engine performance calculations are still based on using the engine simulation model to perform flow path calculations. This approach more fully leverages the existing foundation of overall engine performance simulation models. By simply predicting the engine state guess vector, a variety of engine performance information, from the overall to the component level, can be indirectly obtained, in addition to key engine performance. This includes thermal parameters of various engine flow path interfaces, the operating status of each component, and performance parameters. This not only facilitates detailed assessment of the rationality of the engine's operating state, but also effectively constrains key parameters during the optimization process.

[0099] Furthermore, the precisely predicted engine state guess vector approximates the converged guess vector obtained by iteratively solving the actual engine. Therefore, when used as the initial value for the iterative solution of the engine performance simulation model, it can significantly improve the model's convergence and reduce the number of iterations, thereby significantly reducing the computational complexity of the engine performance simulation model and enhancing its real-time performance. During the surrogate model training steps S5 and S6, the trained surrogate model is fully utilized to improve the convergence of the engine performance simulation model and provide more accurate initial iterative guess vectors for non-converged sample points to enhance convergence. This process further supplements and updates the converged sample set and surrogate model, resolving the issue of incomplete sample point coverage due to the simulation model's inherent lack of convergence. Furthermore, automated iteration and precise automatic adjustment of the initial guess vector significantly reduce the difficulty and workload of surrogate model training.

[0100] In summary, the proxy model proposed in this paper, by leveraging both knowledge and data, fully leverages the established physical laws and scientific theories for solving overall engine performance, ensuring that the proxy model's performance predictions adhere to engine performance principles. It demonstrates significant advantages in terms of accuracy, comprehensiveness, and improved convergence of engine performance predictions, enabling more flexible and extensive application methods and scenarios. The "prediction-performance" application method boasts extremely fast computational speed, albeit with some error. This makes it suitable for engine performance design optimization applications that can tolerate non-optimal solutions, allowing for rapid initial optimization. Furthermore, this method can also be applied to open-loop control feedback requiring rapid engine performance response or other scenarios requiring high computational response. The "prediction-iteration-performance" application method boasts fast computational speed and no prediction error. Applying this method to engine performance design optimization significantly reduces computational costs without compromising optimization accuracy. Furthermore, this method is applicable to iterative engine performance solutions across a wide design range, improving the convergence of simulation models. The proxy model of this invention not only improves the accuracy and computational efficiency of engine performance predictions but also expands its scope of application, meeting diverse engineering needs and possessing broad practical value.

[0101] Example 2

[0102] In order to facilitate understanding of the above technical solutions of the present invention, the advantages of the above technical solutions of the present invention are further described below through specific embodiments.

[0103] Figure 3 A structural schematic diagram of an adaptive cycle aircraft engine is given, showing its main components and duct structure. The adaptive cycle engine is a further development of the variable cycle engine concept, with more powerful cycle adjustment capabilities and mission adaptability. However, compared with traditional engine configurations, the adaptive cycle engine contains more and more complex component types, and more factors affecting the adjustable mechanisms need to be considered. Therefore, the computational complexity of its overall performance simulation model is significantly improved, and the input and output parameter dimensions of the proxy model to be constructed are also increased accordingly. If this high-dimensional proxy model can achieve good results on the adaptive cycle engine, then the application of this proxy model in other engine configurations of comparable or simpler complexity is also expected to achieve excellent performance, thereby effectively verifying the wide applicability and reliability of the technical solution proposed in the present invention.

[0104] After screening, the important design parameters that affect the matching of adaptive cycle engines include: total temperature of engine inlet , rear fan with adjustable stator blade angle , the core engine drives the fan with adjustable stator blade angle , compressor adjustable stator blade angle , low-pressure turbine guide vane angle , the percentage change of the main nozzle throat area A8, the percentage change of the front duct ejector area and the percentage change of the ejector area after the duct The value ranges of each variable are shown in Table 1:

[0105] Table 1 Value ranges of key design parameters of adaptive cycle engines

[0106]

[0107] Since the adaptive cycle engine can select the opening and closing state of the valve by switching the mode, it has two operating modes: three-bypass mode and dual-bypass mode. In these two modes, there are significant differences in the working matching and performance between the components. Therefore, this embodiment carries out training and testing of the proxy model for different working modes. First, the Latin hypercube sampling method is used to generate an initial sample set within the parameter value range shown in Table 1. Subsequently, the simulation model based on the overall performance of the engine is calculated and generated in the dual-bypass and triple-bypass modes respectively. After the preliminary proxy model is trained and generated, the sample set is supplemented and the training proxy model is updated by iteration, and finally the proxy models under the two working modes are obtained. During the entire training process, the proxy model corresponding to each mode contains approximately 6,000 sample points in the convergent sample set, and the total training time is approximately 10 minutes. Overall, the training time is reasonable and efficient for generating high-precision proxy models.

[0108] In terms of proxy model performance evaluation, this embodiment mainly compares the proxy models trained by the solution of the present invention. "Prediction-Performance" Application Method and Traditional Agent Model The performance of the prediction accuracy of the key performance parameters of the engine. The test sample points also use the Latin hypercube sampling method to generate the test sample set within the parameter value range shown in Table 1, and the number of sample points is 2000. and proxy models The training adopts the Kriging model construction method, and the training sample points are consistent.

[0109] Figure 4 and Figure 5 The distribution diagrams of the relative errors of the two proxy models in predicting the thrust and fuel consumption rate of the adaptive cycle engine in different modes are given respectively. As can be seen from the figure, the traditional proxy model The prediction effect is not good: in the three-introduction mode, the maximum relative error of thrust is about 4%, the maximum relative error of fuel consumption is about 1.5%, and the error distribution is relatively dispersed, with a high average relative error level. The "prediction-performance" application method, whether it is thrust or fuel consumption rate, in the two operating modes of the adaptive cycle engine, its maximum relative error of the prediction is controlled within 0.5%, showing extremely high prediction accuracy, which can meet the prediction scenarios of most engine overall performance.

[0110] Figure 6 The accuracy comparison of two prediction methods is shown: one is to use only in The proxy model predicts the guess vector; the other one adds in The proxy model modifies the guess vector. The results show that the introduction of After the proxy model is modified, the normalized root mean square error of each component of the guess vector is significantly reduced, and the maximum normalized root mean square error is reduced to below 0.0065, thus ensuring The surrogate model can maintain stable high accuracy when predicting the key performance of the engine. In addition, the comparison results of the two prediction methods further prove that The agent model uses The necessity and importance of the surrogate model to correct the prediction results of the guess vector

[0111] At the same time, to verify The "prediction-iteration-performance" application method of the proxy model has an improvement effect on the convergence and computational efficiency of the performance simulation model. This embodiment compares the different performances of the traditional simulation model and the proxy model "prediction-iteration-performance" solution method under the same test sample set. The specific comparison results are shown in Table 2. For the two working modes of the adaptive intercooled cycle engine, compared with the solution method of the traditional simulation model, the "prediction-iteration-performance" application method can reduce the number of model calls by more than 80%, that is, the simulation calculation speed is increased by about 5 times. In addition, in the sample point test within a wide design range, the "prediction-iteration-performance" application method also significantly improved the convergence of the model. In the intercooled mode and the dual external mode, the convergence rate percentage relative to the traditional simulation model increased by 63.9% and 22.5%, respectively. The results show that The "prediction-iteration-performance" application method of the surrogate model has a significant effect in improving the convergence and computational efficiency of the performance simulation model.

[0112] Table 2 Comparison of the “prediction-iteration-performance” performance of the proxy model and the traditional simulation model solution

[0113]

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary technicians in this field should understand that they can still modify or replace the technical method of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical method to deviate from the spirit and scope of the technical method of the present invention.

Claims

1. A method for constructing a knowledge and data-driven overall performance proxy model for an aircraft engine, characterized by: The construction method includes the following steps: S1. Obtain the inlet total temperature and engine control law variables and the value range of the inlet total temperature and engine control law variables ; S2, using Latin hypercube sampling method, in the range of Generate initial sample set , the initial sample set Sample points in Represents a value combination of each independent variable; S3, based on the engine performance simulation model, the initial sample set Sample points in Perform simulation iterations one by one to obtain converged sample set A, converged sample set B and non-converged sample set The converged sample set A is composed of the converged sample point a, the converged sample set B is composed of the converged sample point b, and the non-converged sample set From the non-convergent sample points constitute; S4, based on the initial training of the convergent sample set A The proxy model is initially trained based on the converged sample set B to obtain the initial Proxy Model, Initial Proxy Model and Initial The agent models are connected in series to form the initial knowledge data dual-driven agent model ; S5. Using initial knowledge data dual-driven agent model Predicting the non-convergent sample set Non-converged sample points The initial value of the guess vector Based on the engine performance simulation model, each non-convergent sample point is simulated and iterated again to obtain new convergent sample point a, convergent sample point b and non-convergent sample point , update the converged sample set A, converged sample set B and non-converged sample set Get the convergent sample set A , convergent sample set B and non-convergent sample sets ; S6, based on the convergent sample set A Update training to get updated Proxy model, based on convergent sample set B Update training to get updated Proxy Model, Update Proxy Models and Updates Agent model composition and update knowledge data dual-driven agent model ; S7. Repeat S5-S6 until no new convergence sample point a or convergence sample point b appears in S5. Terminate the iteration and obtain the knowledge- and data-driven overall performance proxy model of the aero-engine.

2. The method for constructing a knowledge and data dual-driven aircraft engine overall performance proxy model according to claim 1 is characterized in that: The initial value of the guess vector for simulation iterative solution in S3 ,According to different convergence situations, adopt corresponding processing strategies, as follows: For the sample points where the calculation converges, the strategy is as follows: S301. Record the values ​​of independent variable combinations , iterative convergence guess vector solution , the residual vector obtained by the initial iterative calculation of the simulation model And the guess vector correction ; S302, will As input, As output, the combination is used for The convergence sample point a of the surrogate model training is in the form of , and add it to the converged sample set The convergent sample set A is in the middle; S303, will As input, As output, the combination is used for The convergence sample point b of the agent model training is in the form of , and add it to the converged sample set The middle one is the convergent sample set B; For sample points where the calculation does not converge, the strategy is as follows: Add it to the non-convergent sample set , as the sample set for the next round of simulation iteration.

3. The method for constructing a knowledge and data dual-driven aircraft engine overall performance proxy model according to claim 1 is characterized in that: In S5, for the sample points where the calculation converges, the strategy is as follows: S501, record the value of the independent variable combination IV i ', iterative convergence guess vector solution , the residual vector obtained by the initial iterative calculation of the simulation model And the guess vector correction ; S502, will As input, As output, the combination is used for The new convergence sample point a for the agent model training is in the form of , and add it to the converged sample set That is, the convergent sample set A middle; S503, will As input, As output, the combination is used for The new convergence sample point b for the agent model training is in the form of , and add it to the converged sample set That is, the convergent sample set B middle; For sample points where the calculation does not converge, the strategy is as follows: Add it to the new non-converged sample set , as the sample set for the next round of simulation iteration.

4. The method for constructing a knowledge and data dual-driven aircraft engine overall performance proxy model according to claim 1, characterized in that: The knowledge and data-driven overall performance proxy model of aero-engines is the latest updated one. The KDSM proxy model is constructed by connecting the proxy model with the ETDX proxy model.

5. The method for constructing a knowledge and data dual-driven aircraft engine overall performance proxy model according to claim 1 is characterized in that: The application process and method of the knowledge- and data-driven aircraft engine overall performance proxy model are as follows: Step 1: Given the engine control law variables and the independent variable value combination of the inlet total temperature ; Step 2: As The input of the proxy model is used to predict the guess value vector ; Step 3: Based on the guess vector , call the engine performance simulation model to perform a flow calculation and calculate the error vector ; Step 4: As The input of the proxy model is used to predict the correction of the guess vector , thus obtaining Output of the proxy model ; Step 5: Based on Output of the proxy model , Used to call engine performance simulation models for performance prediction or iterative solution.

6. The method for constructing a knowledge and data dual-driven aircraft engine overall performance proxy model according to claim 5, characterized in that: described The specific contents used to call the engine performance simulation model for performance prediction or iterative solution are as follows: Prediction-Performance: Based on output volume Directly predict the engine matching working state based on its value, call the engine performance simulation model to perform a flow path calculation, and thus predict the engine performance; Prediction-Iteration-Performance: Output As the initial value of the iterative solution of the engine performance model simulation, iterative calculation is performed, and the convergence The determined engine matching operating conditions calculate accurate engine performance.

Images