CN-122021262-A - Air compressor blade profile optimization design method based on condition generation countermeasure network-reinforcement learning and free deformation technology

CN122021262ACN 122021262 ACN122021262 ACN 122021262ACN-122021262-A

Abstract

The invention relates to the technical field of impeller machinery, in particular to a compressor blade profile optimization design method based on a condition generation countermeasure network-reinforcement learning and free deformation technology, which comprises the steps of setting a global parameter space, combining a proxy model in the global parameter space, generating countermeasure network and reinforcement learning, performing multi-objective collaborative active learning and intelligent searching, and generating an optimal basic blade profile meeting global aerodynamic performance requirements; the method comprises the steps of taking an optimal basic blade profile as input, identifying an interference sensitive area of shock wave incidence and a boundary layer based on flow field diagnosis, carrying out parameterization and high-degree-of-freedom deformation on a molded line of the interference sensitive area by adopting a free deformation technology so as to minimize shock wave loss and control flow separation into direct targets for optimization, arranging a non-smooth surface structure on a suction surface of the blade profile, wherein the non-smooth surface structure comprises grooves, and arranging the non-smooth surface structure on the suction surface of the optimized blade profile to obtain the designed blade profile. The invention improves the reliability and engineering applicability of the result.

Inventors

HAN JIANG
GUO ZHONGJIA
YANG XUDONG
HUANG JUNQI
HAN SHAOBING

Assignees

大连海事大学

Dates

Publication Date: 20260512
Application Date: 20260107

Claims (7)

1. The method for optimizing the design of the blade profile of the air compressor based on the condition generation countermeasure network-reinforcement learning and free deformation technology is characterized by comprising the following steps of: S1, setting a global parameter space, combining a proxy model in the global parameter space, generating an countermeasure network and reinforcement learning, performing multi-objective collaborative active learning and intelligent searching, and generating an optimal basic leaf profile meeting global aerodynamic performance requirements; S2, identifying an interference sensitive area of shock wave incidence and a boundary layer based on flow field diagnosis by taking the optimal basic leaf profile as input, and carrying out parameterization and high-degree-of-freedom deformation on a molded line of the interference sensitive area by adopting a free deformation technology so as to minimize shock wave loss and control flow separation to be a direct target for optimization, thereby obtaining an optimized leaf profile; S3, a non-smooth surface structure is arranged on the suction surface of the blade profile, and the non-smooth surface structure comprises grooves; S4, setting the non-smooth surface structure on the suction surface of the optimized blade profile to obtain the designed blade profile.
2. The method for optimizing the design of a compressor blade profile based on a condition generation countermeasure network-reinforcement learning and free deformation technique according to claim 1, wherein S1 comprises: defining a geometric parameter space of the blade profile, wherein the geometric parameter space of the blade profile comprises a chord length, a bent angle, a thickness distribution control point, a mounting angle, an inlet geometric angle and an outlet geometric angle, and setting a value range and design constraint of each parameter; generating initial sample points in a geometric parameter space of the leaf type by utilizing Latin hypercube sampling or Sobol sequences, performing CFD simulation to obtain a total pressure loss coefficient and a pressure ratio performance index, and constructing an initial database; Training a first proxy model for rapidly predicting the performance of leaf type parameters by utilizing the initial database, wherein the first proxy model is formed by Gaussian process regression, gradient lifting trees, random forests and deep neural network weighting; Predicting the performance and uncertainty of a new sample point by using the trained first proxy model, selecting a point with high uncertainty or sensitivity to a target to perform complementary CFD simulation, adding new data into an initial database to update the proxy model, and performing iteration; Training a condition generation countermeasure network, wherein a generator of the condition generation countermeasure network receives a noise vector and a condition vector containing working conditions and performance target thresholds, outputs a leaf pattern parameter combination, and judges whether the leaf pattern parameter combination meets a condition constraint or not by a discriminator; Constructing a first reinforcement learning environment, wherein the first reinforcement learning environment takes a leaf type parameter combination as a state, parameter disturbance as action and pneumatic performance as rewards, adopts a depth deterministic strategy gradient or a near-end strategy optimization method to train an agent, and learns a parameter searching and optimizing strategy; candidate parameters generated by the trained condition generation countermeasure network and the first reinforcement learning environment are screened through the updated first proxy model, CFD verification is carried out on the optimal structure, a result is fed back to update the database, the proxy model and the intelligent network, and basic leaf pattern parameters with optimal performance are output after iteration convergence.
3. The method for optimizing the design of a compressor blade profile based on a condition generating countermeasure network-reinforcement learning and free deformation technique according to claim 2, wherein the total loss function of the condition generating countermeasure network is defined as: Wherein, L is the total loss function, L adv is the counterloss, L const is the constraint penalty, L sur is the proxy consistency loss, L div is the diversity loss, λ 1 is the weight coefficient of the constraint penalty, λ 2 is the weight coefficient of the proxy consistency loss, and λ 3 is the weight coefficient of the diversity loss.
4. The method for optimizing the design of a compressor blade profile based on a condition generation countermeasure network-reinforcement learning and free deformation technique according to claim 1, wherein S2 comprises: On the optimal basic blade profile, identifying a shock wave incidence and boundary layer interference sensitive area through a flow field CFD result, arranging control points in the interference sensitive area, and establishing a free deformation model based on the control points; Taking the offset and local curvature change of the control point as optimization variables, and taking the minimum shock wave intensity, the minimum separation area length and the maximum static pressure coefficient as objective functions; Searching in a local deformation parameter space by adopting a Bayesian optimization, genetic algorithm or reinforcement learning method based on the optimization variable and the objective function, generating a local geometric deformation trend by combining with GAN, and predicting and coupling with a proxy model to obtain a preferred deformation scheme; And carrying out CFD verification on the optimized deformation scheme, and outputting a leaf profile scheme with local sensitive area optimization after comprehensive evaluation to obtain an optimized leaf profile.
5. The method for optimizing the design of a compressor blade profile based on a condition generation countermeasure network-reinforcement learning and free deformation technique according to claim 1, wherein S3 comprises: Carrying out overall optimization design on the structural parameters of the grooves; And carrying out free deformation optimization on the groove subjected to global optimization design.
6. The method for optimizing the design of a compressor blade profile based on a condition generation countermeasure network-reinforcement learning and free deformation technology according to claim 5, wherein the global optimization of the structural parameters of the groove comprises: Defining structural parameters of a non-smooth surface groove, wherein the structural parameters of the non-smooth surface groove comprise groove width, depth, side wall angle, interval and radius of rounding, and setting the value range and design constraint of each structural parameter; sampling and CFD simulation are carried out in the structural parameter space of the non-smooth surface groove, and a groove structural parameter-performance database is constructed; Training a second agent model for predicting the performance of the groove structure based on the groove structure parameter-performance database, and establishing an active learning closed loop to improve the model precision; training a condition generating countermeasure network of the groove structure, wherein a generator of the condition generating countermeasure network of the groove structure is used for generating a groove parameter combination under given working conditions and performance targets; Constructing a second reinforcement learning environment, wherein the second reinforcement learning environment performs optimization search by taking the groove parameters as states to generate candidate groove parameters; And screening and CFD verification are carried out on the candidate groove parameters through a second agent model, and after iterative updating, the grooves after global optimization design are output.
7. The method for optimizing a compressor blade profile based on a condition generation countermeasure network-reinforcement learning and free deformation technique according to claim 5, wherein the optimization of free deformation of the groove after the global optimization is performed comprises: Taking a cross section defined by a groove after global optimization design as an initial geometry, performing shape parametrization on the initial geometry by using a free deformation technology, and simultaneously setting a groove control point and deformation constraint; setting an objective function with the aim of drag reduction and control separation by taking the deformation of a groove control point as an optimization variable; iterative adjustment is carried out on the groove control points by using a gradient method, bayesian optimization or reinforcement learning strategy, so that the cross-sectional shape of the groove is optimized; outputting the optimized groove section, the control point of the optimized groove section and the geometric configuration.

Description

Air compressor blade profile optimization design method based on condition generation countermeasure network-reinforcement learning and free deformation technology Technical Field The invention relates to the technical field of impeller machinery, in particular to a compressor blade profile optimization design method based on a condition generation countermeasure network-reinforcement learning and free deformation technology. Background The air compressor is used as a core component of the aeroengine, and the aerodynamic performance of the air compressor directly influences the overall efficiency and thrust output of the engine. In a variable cycle engine, a compressor needs to adapt to various complex working conditions such as subsonic cruise, high thrust acceleration, cross/supersonic flight and the like, and extremely high requirements are provided for blade profile design. Conventional compressor blade profile designs are based primarily on empirical optimization or low dimensional parameter tuning, with optimization objectives generally limited to a single operating condition or narrow operating range. However, in actual operation of a variable cycle engine, the compressor needs to maintain high aerodynamic efficiency, stability margin, and stall resistance under wide operating conditions. Particularly under the condition of cross/supersonic velocity, the problems of flow instability, efficiency reduction, surge and the like can be caused by strong interference of shock waves and a boundary layer, boundary layer separation and formation of an end wall three-dimensional vortex structure. Therefore, how to realize the optimization design of the blade profile under the conditions of high Reynolds number and high pressure ratio, so that the blade profile has good aerodynamic performance and stability in a wide working condition range, and is a key problem to be solved urgently in the current aeroengine field. At present, an optimal design method of the compressor blade profile mainly depends on empirical modification, parametric modeling and low-fidelity numerical simulation. Although these methods can achieve localized performance improvements under certain conditions, there are significant shortcomings in the wide-range operating-mode adaptations required for variable cycle engines. For example, conventional design methods typically only provide limited adjustments to fundamental geometric parameters such as chord length, bend angle, thickness, and mounting angle, and fail to adequately account for the effects of shock wave incidence location, local flow field non-uniformity, and multi-dimensional flow-structure coupling effects. The problems of over-strong shock waves, early flow separation and the like of the blade profile under the cross/supersonic speed working condition are caused, so that the stability margin and the efficiency of the air compressor are reduced. In addition, the existing optimization method has limited design space, is difficult to efficiently explore high-dimensional parameter combinations, and has low optimization efficiency. Therefore, a new blade profile design method capable of combining high-precision numerical simulation and intelligent optimization technology is needed to solve the shortcomings of the traditional design in the aspects of wide working condition adaptability, shock wave control, flow stability and the like. Disclosure of Invention According to the technical problems of limited design space and low optimization efficiency, the traditional compressor blade profile design method is provided, wherein the defects are caused in the aspects of wide working condition adaptability, shock wave control and flow stability, and the compressor blade profile optimization design method based on the condition generation countermeasure network-reinforcement learning and free deformation technology is provided. The invention mainly utilizes an intelligent design framework of multistage layered synergy to decompose the optimization of the compressor blade profile into three stages of basic blade profile global parameterization intelligent optimization, local flow field sensitive area geometric free deformation and non-smooth surface microstructure integrated optimization, and integrates an integrated agent model, a condition generation countermeasure network (cGAN), reinforcement learning and free deformation (FFD) technology, thereby achieving the effects of realizing progressive refinement optimization from macroscopic pneumatic profile to local flow control, effectively widening design space and improving optimization efficiency. The invention adopts the following technical means: A compressor blade profile optimization design method based on a condition generation countermeasure network-reinforcement learning and free deformation technology comprises the following steps: S1, setting a global parameter space, combining a proxy model in the global parameter space, generating