Aircraft S-shaped ejection system design method based on machine learning

Abstract

The invention provides an aircraft S-shaped injection system design method based on machine learning, comprising: constructing a three-dimensional geometric model of an injection system, the model comprising a nozzle connection section, an S-shaped curved surface section and an outlet connection section; parametric modeling is carried out on the S-shaped curved surface section, and geometric parameters and control variables of the S-shaped curved surface section are defined; constructing a machine learning model and an aerodynamic performance optimization target, and optimizing geometric parameters of the S-shaped curved surface; and performing pneumatic and structure matching verification on the optimized S-shaped curved surface section, the spray pipe connecting section and the outlet connecting section.

Description

Technical Field

[0001] This invention belongs to the field of aircraft internal flow channel design technology, and specifically relates to a design method for an S-shaped ejector system for aircraft based on machine learning. Background Technology

[0002] With the rapid development of aerospace technology, the performance requirements of aircraft are increasing, and their exhaust systems, as key components, have a crucial impact on the overall performance of aircraft. Traditional aircraft exhaust system design methods mainly rely on empirical formulas, theoretical calculations, and wind tunnel tests. These methods have many limitations when dealing with the design of complex S-shaped ejector systems. Empirical formulas are often derived based on specific aircraft types and operating conditions, and their adaptability and accuracy are difficult to guarantee for exhaust system designs of new aircraft or under non-standard operating conditions. Although theoretical calculations can provide relatively accurate analysis results, solving the governing equations of systems like S-shaped ejector systems, which have complex shapes and hydrodynamic characteristics, requires a lot of time and computational resources, and it is difficult to quickly obtain optimized design solutions in practical engineering applications. While wind tunnel tests are an important means of verifying exhaust system performance, they are costly, time-consuming, and the test conditions differ from the actual flight environment, limiting the extrapolation of test results. In addition, traditional design methods are also insufficient when considering the multidisciplinary coupling problems of S-shaped ejector systems. The design of aircraft exhaust systems involves not only fluid mechanics but also structural mechanics and other disciplines. Traditional design methods often involve designing each discipline independently and then coordinating them. This approach makes it difficult to achieve overall system optimization and can easily lead to bottlenecks in the performance of certain disciplines, thus failing to fully realize the potential of each discipline.

[0003] In recent years, machine learning technology has been widely applied in various fields, and its powerful data processing and pattern identification capabilities have provided new ideas for solving complex engineering problems. In the field of aircraft exhaust system design, machine learning methods can establish a mapping relationship between input parameters and output performance by learning from a large number of known cases and corresponding performance data, thereby quickly predicting the performance of new design schemes and providing guidance for design optimization. However, applying machine learning technology to the design of aircraft S-shaped ejector systems also faces some challenges. On the one hand, exhaust systems involve numerous parameters, including geometric parameters and operating condition parameters. How to effectively select and process these parameters so that they can be accurately learned and identified by machine learning models is a problem that urgently needs to be solved. On the other hand, the accuracy and reliability of machine learning models require a large amount of high-quality data for verification and evaluation. In practical engineering applications, obtaining sufficient and representative exhaust system design data is not easy and requires interdisciplinary and cross-domain collaboration and data sharing.

[0004] In summary, traditional aircraft exhaust system design methods have many shortcomings when dealing with complex systems like S-shaped ejector systems. While machine learning technology has potential advantages, it still faces some challenges in practical applications. Therefore, developing a machine learning-based design method for aircraft S-shaped ejector systems can fully utilize the advantages of machine learning, overcome the limitations of traditional design methods, and provide an efficient, accurate, and optimized solution for aircraft exhaust system design. This has significant theoretical and engineering application value. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the inventors have conducted intensive research and provided a machine learning-based design method for S-shaped ejector systems of aircraft. This method enables rapid and optimized design of S-shaped ejector systems, improves the performance and reliability of aircraft exhaust systems, and promotes the development of aerospace technology.

[0006] The technical solution provided by this invention is as follows: A machine learning-based design method for an S-shaped ejection system for aircraft includes the following steps: A three-dimensional geometric model of the ejection system is constructed, the model including a nozzle connection section, an S-shaped curved surface section and an exit connection section; Parametric modeling of the S-shaped surface segment is performed, defining its geometric parameters and control variables; Construct a machine learning model and aerodynamic performance optimization objectives to optimize the geometric parameters of the S-shaped surface; The optimized S-shaped curved surface section was aerodynamically and structurally matched with the nozzle connection section and the outlet connection section.

[0007] Furthermore, the nozzle connection section adopts a split modular design. The high-temperature section of the nozzle connection section, which connects to the tail nozzle of the aircraft, has an inner wall surface made of silicon carbide fiber reinforced ceramic matrix composite material and integrates a closed-loop active cooling channel. The transition section connecting the nozzle connection section and the S-shaped curved section is made of nickel-based high-temperature alloy casting. The detachable connection mechanism is used for detachable installation of the high-temperature section, allowing the high-temperature section to be replaced without stopping the engine.

[0008] Furthermore, the outlet connection section is an expansion channel with an expansion angle of 2°~10°, and its inner wall is made of high-temperature resistant composite material, which integrates an active cooling structure; the outlet section and the rear body skin of the aircraft adopt a conformal fusion design, the fusion zone length is ≥1.5 times the outlet diameter, and the curvature continuity is G2 level.

[0009] Furthermore, the parametric modeling of the S-shaped surface segment includes: defining the geometric features of the S-shaped surface segment using a non-uniform rational B-spline method, which consists of four curves, wherein: each spline curve has multiple control points, and the first and last control points coincide with the geometric centers of the nozzle connection section outlet section and the outlet connection section inlet section, respectively; the control point weight factor is dynamically adjusted based on aerodynamic performance sensitivity analysis, and the weight adjustment range is 0.7~1.3.

[0010] Furthermore, the step of parametrically modeling the S-shaped surface segment also includes setting geometric constraints: (i) Curvature continuity constraint: The rate of change of curvature between adjacent control points is ≤20%, and the radius of curvature is at its minimum value. R min satisfy:

[0011] in, D nozzle The nozzle connection section outlet diameter is [missing information]. L s-curve The axial length of the S-shaped curved surface segment; (ii) Flow channel cross-sectional area constraint: the rate of change of cross-sectional area along the airflow direction is ≤10% / mm.

[0012] Furthermore, the step of parametrically modeling the S-shaped surface segment also includes parameter coupling verification of the parametrically modeled NURBS parametric model, specifically: The NURBS parametric model interacts with the aerodynamic performance prediction model in real time. When the risk of boundary layer separation is detected, 5 to 10 control points in the high-risk area are automatically added to each curve. The structural strength was verified by finite element analysis, and the maximum equivalent stress of the S-shaped curved surface segment was constrained to be ≤ 70% of the allowable stress of the material.

[0013] Furthermore, in the step of constructing the machine learning model and the aerodynamic performance optimization objective, the generation of training data for the machine learning model is implemented in the following manner: (i) Through joint simulation of computational fluid dynamics (CFD) and thermal structure coupled analysis, a training set containing the following multi-dimensional data is generated: Aerodynamic performance data: total pressure recovery coefficient, exit Mach number distribution, secondary flow intensity, ejection efficiency; Thermodynamic data: wall heat flux density distribution, cooling channel heat transfer efficiency; Structural response data: thermal stress contour plot, local strain rate, vibration modal frequency; (ii) Based on multiphysics coupling data, sample the boundary layer flow field data of the S-shaped curved surface segment, including virtual samples under extreme conditions: An improved Latin hypercube sampling method is used to generate basic sample points in the NURBS parameter space; An active learning mechanism is introduced, and based on the uncertainty quantification results of the Gaussian process proxy model, sampling is intensified in the sensitive area of ​​the S-shaped surface segment with a curvature change rate > 12%. Data augmentation of the boundary layer separation field is performed by adversarial generative mesh to generate virtual samples under extreme conditions. (iii) Physical constraint embedding During the data annotation stage, conservation law constraints are enforced to automatically eliminate simulation results that violate mass conservation or momentum conservation. By removing topological features of the flow field through graph neural mesh, a rule base for the association of curvature, vorticity, and entropy production is established for the rationality verification of training data.

[0014] Furthermore, in the step of constructing the machine learning model and the aerodynamic performance optimization objective, the machine learning model adopts a multimodal hybrid intelligent architecture, including: (i) Construct a "dual-channel-multi-task" hybrid learning framework, wherein: Channel 1: Extracts flow field features based on deep convolutional neural networks. The input is a high-resolution cloud map. An improved U-net architecture is used, with an encoder of 8 layers and a decoder of 6 layers. Channel 2: Based on random forest processing of structured parameters, the input includes: a) flight condition parameters, Mach number range, angle of attack range, sideslip angle range; b) geometric parameters: S-curve inlet size, S-curve outlet size, S-curve length, outlet spread angle; (ii) Multi-task learning design Main task: Predict the coordinate sequence of control points on an S-shaped surface, and output a 4×N dimensional vector, where N corresponds to the number of control points on each line; Auxiliary tasks: Predict flow field performance indicators, total pressure recovery coefficient, and outlet wall gas temperature; The optimization is achieved through a weighted multi-task loss function, with the main task having a weight of 0.6 and the auxiliary task having a weight of 0.4. (iii) Physical constraint embedding Add a conservation law constraint module to the output layer to force the following conditions to be met: i) mass conservation: inlet and outlet flow deviation is less than 0.5%; ii) momentum conservation: axial thrust coefficient error is less than 1.2%.

[0015] Furthermore, in the step of constructing the machine learning model and the aerodynamic performance optimization objective, the aerodynamic performance optimization objective includes: Aerodynamic performance targets: Total pressure recovery coefficient ≥ 0.95, ejection efficiency ≥ 14%; Structural performance target: Maximum thermal stress ≤ 60% of material yield strength.

[0016] Furthermore, in the step of constructing the machine learning model and the aerodynamic performance optimization objective, the step of implementing aerodynamic performance optimization of the S-shaped curved surface segment using a multi-objective optimization algorithm includes: (i) Constructing a hybrid intelligent optimization framework for multi-objective optimization algorithms First layer, global exploration layer An improved NSGA-III algorithm was used, with a population size of 100-200, a crossover probability of 0.6-0.8, and a mutation probability of 0.05-0.1. The optimization objectives include aerodynamic performance objectives and structural performance objectives; The second layer, the partial refining layer. A Gaussian process regression surrogate model is constructed based on the Pareto solution set, and a hypervolume improvement criterion is used to select candidate schemes. Topology optimization theory is introduced to rank the parameter sensitivity of the S-shaped surface control points and retain key variables with sensitivity ≥0.8. The third layer, the physical verification layer. The optimization results are verified in real time using a digital twin system, and the candidate solutions are matched with a multiphysics database for consistency, with an error threshold set at 5%. (ii) Establish a dynamic constraint handling mechanism Establish an adaptive constraint violation function:

[0017] Where n≥3, g i (x) represents inequality constraints, including the total pressure recovery coefficient g1(x) ≥ 0.95, the ejection efficiency g2(x) ≥ 14%, and the maximum thermal stress g3(x) ≤ 60% of the material's yield strength. i,max This represents the maximum permissible violation value for the i-th inequality constraint, used for normalization; m≥2, h i (x) represents equality constraints, including mass conservation constraint h1(x) and momentum conservation constraint h2(x). h j,ref CV(x) is the reference value for the j-th equality constraint, used for normalization; CV(x) is the total constraint violation value, used to evaluate the overall constraint satisfaction of design variable x.

[0018] The design method for an S-shaped ejection system for an aircraft based on machine learning provided by the present invention has the following beneficial effects: (1) Improve design efficiency and accuracy: Traditional design methods rely on experience, theoretical calculations and wind tunnel tests, which have the problems of long cycle and low accuracy. This method can quickly process a large amount of data, establish the mapping relationship between input parameters and output performance, thereby quickly predicting the performance of new design schemes and providing guidance for design optimization. This not only greatly shortens the design cycle, but also improves the accuracy of design results.

[0019] (2) Optimize system performance: By learning from a large number of known design cases and corresponding performance data, this invention optimizes the geometry and hydrodynamic performance of the ejector system, realizes real-time optimization under high-speed conditions, improves the overall performance and reliability of the system, and achieves an ejection efficiency of ≥14%.

[0020] (3) Enhancing multidisciplinary collaborative optimization capabilities: Traditional design methods often adopt the approach of designing independently by each discipline and then coordinating them. Machine learning-based design methods can integrate multidisciplinary data and achieve overall system optimization by establishing mapping relationships between multidisciplinary parameters. Attached Figure Description

[0021] Figure 1 Schematic diagram of S-shaped ejection system; Figure 2 Parametric modeling of S-shaped surfaces.

[0022] 1 is the nozzle connection section, 2 is the S-shaped curved surface section, and 3 is the outlet connection section. Detailed Implementation

[0023] The features and advantages of the present invention will become clearer and more apparent from the following detailed description.

[0024] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0025] This invention provides a machine learning-based design method for an S-shaped ejection system for aircraft, comprising the following steps: Step 1: Construct a three-dimensional geometric model of the ejector system. The model includes nozzle connection section 1, S-shaped curved surface section 2, and exit connection section 3. See [link / details]. Figure 1 .

[0026] In this step, the nozzle connection section 1 adopts a split modular design. The high-temperature section of the nozzle connection section 1, which connects to the tail nozzle of the aircraft, has its inner wall surface made of silicon carbide fiber reinforced ceramic matrix composite material and integrates a closed-loop active cooling channel. The transition section connecting the nozzle connection section 1 and the S-shaped curved section 2 is made of nickel-based high-temperature alloy casting. The detachable connection mechanism is used for the detachable installation of the high-temperature section and includes a radial flange, circumferential positioning pin and quick-release latch, which allows the high-temperature section to be replaced without stopping the engine.

[0027] In this step, the outlet connection section 3 is an expansion channel with an expansion angle of 2°~10°. Its inner wall is made of high-temperature resistant composite material and it integrates an active cooling structure. The outlet section and the rear body skin of the aircraft adopt a conformal fusion design, with the fusion zone length ≥ 1.5 times the outlet diameter and curvature continuity level G2.

[0028] Step 2: Perform parametric modeling on the S-shaped surface segment, defining its geometric parameters and control variables.

[0029] In this step, the parametric modeling of the S-shaped surface segment is implemented through the following steps: (i) Parametric modeling stage The geometric features of the S-shaped surface segment are defined using the Non-Uniform Rational B-Spline (NURBS) method, consisting of four curves. Each spline curve has 10-15 control points, with the first and last control points coinciding with the geometric centers of the nozzle connection section's exit section and the nozzle connection section's inlet section, respectively. (See...) Figure 2 The control point weighting factor is dynamically adjusted based on aerodynamic performance sensitivity analysis, with an adjustment range of 0.7 to 1.3.

[0030] (ii) Definition of geometric constraints Curvature continuity constraint: The rate of change of curvature between adjacent control points (A1~Am, B1~Bn, C1~Cp, D1~Ds) is ≤20%, and the radius of curvature is at its minimum value. R min satisfy: (1) in, D nozzle The nozzle connection section outlet diameter is [missing information]. L s-curve This is the axial length of the S-shaped curved surface segment.

[0031] Flow channel cross-sectional area constraint: the rate of change of cross-sectional area along the airflow direction ≤ 10% / mm.

[0032] (iii) Parameter coupling verification The NURBS parametric model interacts with the aerodynamic performance prediction model in real time. When the risk of boundary layer separation is detected, 5 to 10 control points in the high-risk area are automatically added to each curve. The structural strength was verified by finite element analysis, and the maximum equivalent stress of the S-shaped curved surface segment was constrained to be ≤ 70% of the allowable stress of the material.

[0033] Step 3: Construct a machine learning model and aerodynamic performance optimization objectives to optimize the geometric parameters of the S-shaped surface.

[0034] In this step, the training data for the machine learning model is generated in the following manner: (i) Acquiring multiphysics coupling data By using computational fluid dynamics (CFD) and thermal structure coupled analysis for joint simulation, a training set containing the following multi-dimensional data is generated; Aerodynamic performance data: total pressure recovery coefficient, exit Mach number distribution, secondary flow intensity, ejection efficiency; Thermodynamic data: wall heat flux density distribution, cooling channel heat transfer efficiency; Structural response data: thermal stress contour plot, local strain rate, vibration modal frequency; (ii) Based on the multiphysics coupling data, the boundary layer flow field data of the S-shaped curved surface segment is sampled. The adaptive intelligent sampling strategy is as follows: The first stage employs improved Latin hypercube sampling to generate basic sample points within the NURBS parameter space; The second stage introduces an active learning mechanism, which, based on the uncertainty quantification results of the Gaussian process proxy model, densifies sampling in sensitive areas of the S-shaped surface segment with a curvature change rate > 12%. The third stage uses adversarial generative meshes to augment the boundary layer separation field, generating virtual samples under extreme conditions. (iii) Physical constraint embedding During the data annotation phase, enforce conservation law constraints to address violations of quality conservation (residuals > 1e). -4 ) or conservation of momentum (slaughter > 5e -3 The simulation results are automatically discarded. By removing topological features of the flow field through graph neural mesh, a rule base for the association of curvature, vorticity, and entropy production is established for the rationality verification of training data.

[0035] In this step, the machine learning model adopts a multimodal hybrid intelligent architecture, specifically including: (i) Model Architecture Construct a hybrid learning framework of "dual-channel-multi-task", in which Channel 1: Extracts flow field features based on deep convolutional neural networks (CNN). The input is a high-resolution cloud map (including pressure, temperature, and velocity gradient fields). It adopts an improved U-net architecture with an encoder depth of 8 layers and a decoder depth of 6 layers. Channel 2: Based on random forest processing of structured parameters, the input includes: a) flight condition parameters, Mach number range, angle of attack range, sideslip angle range; b) geometric parameters: S-curve inlet size, S-curve outlet size, S-curve length, outlet spread angle.

[0036] (ii) Multi-task learning design Main task: Predict the coordinate sequence of control points on an S-shaped surface, and output a 4×N dimensional vector (N corresponds to the number of control points on each vector).

[0037] Auxiliary tasks: predict flow field performance indicators, total pressure recovery coefficient, and outlet wall airflow temperature.

[0038] The optimization is achieved through a multi-task loss function with a weighted average. The main task has a weight of 0.6, and the auxiliary task has a weight of 0.4.

[0039] (iii) Physical constraint embedding Add a conservation law constraint module to the output layer to force the following conditions to be met: a) mass conservation: inlet and outlet flow deviation is less than 0.5%; b) momentum conservation: axial thrust coefficient error is less than 1.2%.

[0040] In this step, the aerodynamic performance optimization objectives include: Aerodynamic performance targets: Total pressure recovery coefficient ≥ 0.95, ejection efficiency ≥ 14%; Structural performance target: Maximum thermal stress ≤ 60% of material yield strength.

[0041] In this step, a multi-objective optimization algorithm is used to optimize the aerodynamic performance of the S-shaped curved surface segment. The multi-objective optimization algorithm adopts a hybrid intelligent optimization framework and achieves collaborative optimization through the following steps: (a) Layered optimization architecture (i) First layer, global exploration layer An improved NSGA-III algorithm was used, with a population size of 100-200, a crossover probability of 0.6-0.8, and a mutation probability of 0.05-0.1. The optimization targets include aerodynamic performance targets (total pressure recovery coefficient ≥ 0.95, ejection efficiency ≥ 14%) and structural performance targets (maximum thermal stress ≤ 60% of material yield strength).

[0042] (ii) Second layer, partial refining layer A Gaussian process regression surrogate model is constructed based on the Pareto solution set, and a hypervolume improvement criterion is used to select candidate schemes. Topology optimization theory is introduced to rank the parameter sensitivity of the S-shaped surface control points and retain key variables with sensitivity ≥0.8.

[0043] (iii) The third layer, the physical verification layer The optimization results are verified in real time through a digital twin system, and the candidate solutions are matched with the above multiphysics database for consistency, with the error threshold set at 5%.

[0044] (b) Dynamic constraint handling mechanism Establish an adaptive constraint violation function: (2) Where n≥3, g i (x) represents inequality constraints, including the total pressure recovery coefficient g1(x) ≥ 0.95, the ejection efficiency g2(x) ≥ 14%, and the maximum thermal stress g3(x) ≤ 60% of the material's yield strength. i,max This represents the maximum permissible violation value for the i-th inequality constraint, used for normalization; m≥2, h i (x) represents equality constraints, including mass conservation constraint h1(x) and momentum conservation constraint h2(x). h j,ref CV(x) is the reference value for the j-th equality constraint, used for normalization; CV(x) is the total constraint violation value, used to evaluate the overall constraint satisfaction of design variable x.

[0045] Step 4: Perform aerodynamic and structural matching verification on the optimized S-shaped curved surface section, nozzle connection section, and outlet connection section.

[0046] The present invention also provides an S-shaped ejector system, which is applied to the exhaust system of a high-speed aircraft, comprising: The nozzle connection section has its inlet connected to the aircraft's exhaust system and its outlet connected to the S-shaped curved surface. The shape and size of the nozzle connection section are optimized according to the aircraft's exhaust requirements to ensure smooth exhaust and reduce flow loss. The shape and size of the S-shaped surface are determined by machine learning using the method described above. This can effectively reduce the volume and weight of the system while ensuring aerodynamic performance. Furthermore, the curvature change of the S-shaped surface is smooth, avoiding airflow separation and pressure loss caused by abrupt flow changes. The outlet connection section has its inlet connected to the S-shaped curved surface and its outlet connected to the external environment. The shape and size of the outlet connection section are optimized according to the shape and size of the S-shaped curved surface and the external environmental conditions to ensure that the exhaust can be smoothly discharged and quickly diffused, reducing interference with the flow field around the aircraft.

[0047] The entire system is optimized through the methods described above, enabling good matching and collaborative working capabilities between the nozzle connection section, the S-curve surface, and the outlet connection section. This improves the overall aerodynamic performance and structural strength of the system, while reducing system complexity and manufacturing costs. It can meet the exhaust requirements of the aircraft under different flight conditions, thereby improving the performance and reliability of the aircraft.

[0048] The present invention has been described in detail above with reference to specific embodiments and exemplary examples; however, these descriptions should not be construed as limiting the present invention. Those skilled in the art will understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and embodiments of the present invention without departing from the spirit and scope of the invention, and all such modifications and improvements fall within the scope of the present invention. The scope of protection of the present invention is defined by the appended claims.

[0049] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A design method for an S-shaped ejection system for an aircraft based on machine learning, characterized in that, Includes the following steps: A three-dimensional geometric model of the ejection system is constructed, the model including a nozzle connection section, an S-shaped curved surface section and an exit connection section; Parametric modeling of the S-shaped surface segment is performed, defining its geometric parameters and control variables; Construct a machine learning model and aerodynamic performance optimization objectives to optimize the geometric parameters of the S-shaped surface; The optimized S-shaped curved surface section was aerodynamically and structurally matched with the nozzle connection section and the outlet connection section.

2. The design method for an S-shaped ejection system for an aircraft based on machine learning according to claim 1, characterized in that, The nozzle connection section adopts a split modular design. The high-temperature section of the nozzle connection section, which connects to the tail nozzle of the aircraft, has its inner wall surface made of silicon carbide fiber reinforced ceramic matrix composite material and integrates a closed-loop active cooling channel. The transition section connecting the nozzle connection section and the S-shaped curved section is made of nickel-based high-temperature alloy casting. The detachable connection mechanism is used for the detachable installation of the high-temperature section, allowing the high-temperature section to be replaced without stopping the engine.

3. The design method for an S-shaped ejection system for an aircraft based on machine learning according to claim 1, characterized in that, The outlet connection section is an expansion channel with an expansion angle of 2°~10°. Its inner wall is made of high-temperature resistant composite material and has an integrated active cooling structure. The outlet section and the rear body skin of the aircraft adopt a conformal fusion design with a fusion zone length ≥ 1.5 times the outlet diameter and a curvature continuity level of G2.

4. The design method for an S-shaped ejection system for an aircraft based on machine learning according to claim 1, characterized in that, The parametric modeling of the S-shaped surface segment includes: defining the geometric features of the S-shaped surface segment using a non-uniform rational B-spline method, which consists of four curves, wherein: each spline curve has multiple control points, and the first and last control points coincide with the geometric centers of the nozzle connection section exit section and the outlet connection section inlet section, respectively; the control point weight factor is dynamically adjusted based on aerodynamic performance sensitivity analysis, and the weight adjustment range is 0.7~1.

3.

5. The design method for an S-shaped ejection system for aircraft based on machine learning according to claim 1, characterized in that, The step of parametrically modeling the S-shaped surface segment also includes setting geometric constraints: (i) Curvature continuity constraint: The rate of change of curvature between adjacent control points is ≤20%, and the radius of curvature is at its minimum value. R min satisfy: in, D nozzle The nozzle connection section outlet diameter is [missing information]. L s-curve The axial length of the S-shaped curved surface segment; (ii) Flow channel cross-sectional area constraint: the rate of change of cross-sectional area along the airflow direction is ≤10% / mm.

6. The design method for an S-shaped ejection system for an aircraft based on machine learning according to claim 1, characterized in that, The step of parametrically modeling the S-shaped surface segment also includes parametric coupling verification of the NURBS parametric model after parametric modeling, specifically: The NURBS parametric model interacts with the aerodynamic performance prediction model in real time. When the risk of boundary layer separation is detected, 5 to 10 control points in the high-risk area are automatically added to each curve. The structural strength was verified by finite element analysis, and the maximum equivalent stress of the S-shaped curved surface segment was constrained to be ≤ 70% of the allowable stress of the material.

7. The design method for an S-shaped ejection system for an aircraft based on machine learning according to claim 1, characterized in that, In the steps of constructing the machine learning model and the aerodynamic performance optimization objective, the training data for the machine learning model is generated in the following manner: (i) Through joint simulation of computational fluid dynamics and thermal structure coupled analysis, a training set containing the following multi-dimensional data is generated: Aerodynamic performance data: total pressure recovery coefficient, exit Mach number distribution, secondary flow intensity, ejection efficiency; Thermodynamic data: wall heat flux density distribution, cooling channel heat transfer efficiency; Structural response data: thermal stress contour plot, local strain rate, vibration modal frequency; (ii) Based on multiphysics coupling data, sample the boundary layer flow field data of the S-shaped curved surface segment, including virtual samples under extreme conditions: An improved Latin hypercube sampling method is used to generate basic sample points in the NURBS parameter space; An active learning mechanism is introduced, and based on the uncertainty quantification results of the Gaussian process proxy model, sampling is intensified in the sensitive area of ​​the S-shaped surface segment with a curvature change rate > 12%. Data augmentation of the boundary layer separation field is performed by adversarial generative mesh to generate virtual samples under extreme conditions. (iii) Physical constraint embedding During the data annotation stage, conservation law constraints are enforced to automatically eliminate simulation results that violate mass conservation or momentum conservation. By removing topological features of the flow field through graph neural mesh, a rule base for the association of curvature, vorticity, and entropy production is established for the rationality verification of training data.

8. The design method for an S-shaped ejection system for an aircraft based on machine learning according to claim 1, characterized in that, In the steps of constructing the machine learning model and the aerodynamic performance optimization objective, the machine learning model adopts a multimodal hybrid intelligent architecture, including: (i) Construct a "dual-channel-multi-task" hybrid learning framework, wherein: Channel 1: Extracts flow field features based on deep convolutional neural networks. The input is a high-resolution cloud map. An improved U-net architecture is used, with an encoder of 8 layers and a decoder of 6 layers. Channel 2: Based on random forest processing of structured parameters, the input includes: a) flight condition parameters, Mach number range, angle of attack range, sideslip angle range; b) geometric parameters: S-curve inlet size, S-curve outlet size, S-curve length, outlet spread angle; (ii) Multi-task learning design Main task: Predict the coordinate sequence of control points on an S-shaped surface, and output a 4×N dimensional vector, where N corresponds to the number of control points on each line; Auxiliary tasks: Predict flow field performance indicators, total pressure recovery coefficient, and outlet wall gas temperature; The optimization is achieved through a weighted multi-task loss function, with the main task having a weight of 0.6 and the auxiliary task having a weight of 0.

4. (iii) Physical constraint embedding Add a conservation law constraint module to the output layer to force the following conditions to be met: i) mass conservation: inlet and outlet flow deviation is less than 0.5%; ii) momentum conservation: axial thrust coefficient error is less than 1.2%.

9. The design method for an S-shaped ejection system for an aircraft based on machine learning according to claim 1, characterized in that, In the steps of constructing the machine learning model and the aerodynamic performance optimization objective, the aerodynamic performance optimization objective includes: Aerodynamic performance targets: Total pressure recovery coefficient ≥ 0.95, ejection efficiency ≥ 14%; Structural performance target: Maximum thermal stress ≤ 60% of material yield strength.

10. The design method for an S-shaped ejection system for an aircraft based on machine learning according to claim 1, characterized in that, In the steps of constructing the machine learning model and the aerodynamic performance optimization objective, the step of implementing aerodynamic performance optimization of the S-shaped curved surface segment using a multi-objective optimization algorithm includes: (i) Constructing a hybrid intelligent optimization framework for multi-objective optimization algorithms First layer, global exploration layer An improved NSGA-III algorithm was used, with a population size of 100-200, a crossover probability of 0.6-0.8, and a mutation probability of 0.05-0.

1. The optimization objectives include aerodynamic performance objectives and structural performance objectives; The second layer, the partial refining layer. A Gaussian process regression surrogate model is constructed based on the Pareto solution set, and a hypervolume improvement criterion is used to select candidate schemes. Topology optimization theory is introduced to rank the parameter sensitivity of the S-shaped surface control points and retain key variables with sensitivity ≥0.

8. The third layer, the physical verification layer. The optimization results are verified in real time using a digital twin system, and the candidate solutions are matched with a multiphysics database for consistency, with an error threshold set at 5%. (ii) Establish a dynamic constraint handling mechanism Establish an adaptive constraint violation function: Where n≥3, g i (x) represents inequality constraints, including the total pressure recovery coefficient g1(x) ≥ 0.95, the ejection efficiency g2(x) ≥ 14%, and the maximum thermal stress g3(x) ≤ 60% of the material's yield strength. i,max This represents the maximum permissible violation value for the i-th inequality constraint, used for normalization; m≥2, h i (x) represents equality constraints, including mass conservation constraint h1(x) and momentum conservation constraint h2(x). h j,ref CV(x) is the reference value for the j-th equality constraint, used for normalization; CV(x) is the total constraint violation value, used to evaluate the overall constraint satisfaction of design variable x.

Images