A method for quantitatively evaluating performance uncertainty of impeller mechanical blade profile geometry deviation

Abstract

This invention relates to the field of turbomachinery performance evaluation technology, and discloses a method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometric deviations. The method includes parametrically modeling geometric deviation characteristics and sampling to construct a sample; generating a three-dimensional blade simulation model based on the sample by performing deviation modeling on the original design blade profile; performing aerodynamic simulation calculations on the simulation model; extracting aerodynamic indicators using a condition control strategy to construct a mapping database; training a surrogate model using the database to establish mapping relationships; and finally using the surrogate model to perform aerodynamic performance inference to complete the evaluation. Furthermore, interpretable machine learning methods are combined to calculate contribution to identify geometric features highly sensitive to aerodynamic performance. This invention overcomes the limitation of two-dimensional evaluation in failing to consider radial flow effects by reconstructing a three-dimensional simulation model, and eliminates flow deviation operating errors by employing a condition control strategy, providing a quantitative decision-making basis for robust design, machining tolerance setting, and cleaning and maintenance schemes for turbomachinery.

Description

Technical Field

[0001] This invention relates to the field of turbomachinery performance evaluation technology, specifically a method for quantitatively evaluating the uncertainty of turbomachinery blade geometry deviation performance. Background Technology

[0002] Fans, compressors, and turbines are core components of various power and industrial energy conversion devices, and their aerodynamic performance directly affects the overall operating efficiency and reliability of the machine. Even well-designed turbomachinery can exhibit performance deficiencies during actual testing and use. The root cause lies in the uncertainties introduced into the design geometry of the turbomachinery at various stages of design, production, manufacturing, assembly, and even use. (See attached diagram) Figure 11 In the figure, the horizontal axis represents performance and the vertical axis represents probability distribution. Due to the influence of uncertainties, the actual performance of turbomachinery deviates from the design performance. In some cases, the performance data will be lower than the required performance limit, thus falling into the substandard performance range. Traditional deterministic design cannot fully consider many uncertainties, which increases the design iteration cycle of turbomachinery and the entire device, as well as the manufacturing and maintenance costs after production.

[0003] Existing methods for quantifying the aerodynamic performance uncertainty of airfoils are typically based on simulations using two-dimensional blade cascade models. These methods input error features such as the leading and trailing edge shape, thickness, and installation angle of the airfoil section when quantifying airfoil performance uncertainty. Based on these input error features, they obtain the aerodynamic performance dispersion through independent two-dimensional blade cascade simulation calculations.

[0004] This uncertainty analysis approach, which isolates the research object from the real operating environment, has objective limitations. It ignores the changes in the pressure gradient field caused by the geometric deviations of the studied blade profile in the real three-dimensional environment of turbomachinery, and therefore cannot consider the accompanying three-dimensional spatial integral effects such as radial redistribution of flow. These limitations mean that existing quantitative results cannot accurately reflect the dispersed impact of blade profile deviations on the overall macroscopic performance of three-dimensional blades, and cannot accurately identify sensitive deviation parameters that lead to performance degradation. Therefore, developing a quantitative assessment method for the uncertainty of geometric deviation performance based on the three-dimensional blade operating environment to accurately consider the influence of three-dimensional effects is a problem that needs to be solved in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a quantitative evaluation method for the performance uncertainty of turbomachinery blade geometric deviation. This method solves the problem that existing quantitative evaluation methods based on two-dimensional blade cascade models ignore the three-dimensional spatial integral effects such as radial redistribution of flow caused by geometric deviation in a real three-dimensional environment. As a result, the quantitative results cannot objectively reflect the impact of geometric deviation on the overall aerodynamic performance dispersion of the three-dimensional blade, and cannot identify the geometric features that are highly sensitive to aerodynamic performance and cause performance fluctuations.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] This invention provides a method for quantitatively evaluating the uncertainty of the performance of turbomachinery blade geometry deviation, employing the following technical solution:

[0008] A method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation includes the following steps:

[0009] The geometric deviation characteristics of the turbomachinery blade profile are parametrically modeled to extract geometric deviation parameters and sampled to construct a geometric deviation parameter sample. Based on the geometric deviation parameter sample, the original blade profile is deviated and modeled to generate a three-dimensional blade simulation model.

[0010] Aerodynamic simulation calculations are performed on the three-dimensional blade simulation model. The aerodynamic indices of the geometric deviation parameter samples under the same working conditions are extracted using the operating condition control strategy, and a mapping database of geometric deviation and aerodynamic indices is constructed.

[0011] A surrogate model is trained using the geometric deviation and aerodynamic index mapping database to establish a nonlinear mapping relationship between the geometric deviation parameters and the aerodynamic indexes.

[0012] The surrogate model is used to perform aerodynamic performance inference on randomly generated geometric deviation parameter samples to complete uncertainty quantification assessment. In addition, interpretable machine learning methods are used to calculate the contribution of each geometric deviation parameter in the randomly generated geometric deviation parameter samples to identify geometric features that are highly sensitive to aerodynamic performance.

[0013] By adopting the above technical solution, the time cost and computational power consumption of aerodynamic evaluation are reduced due to the combination of parametric modeling and surrogate models. The operating condition control strategy eliminates the interference of variable operating conditions on aerodynamic indicators, ensuring that performance degradation originates from geometric deviations. Combined with the interpretable machine learning method, the contribution of each geometric deviation parameter to performance fluctuations is quantified while realizing aerodynamic performance inference. Therefore, the solution achieves the effect of identifying highly sensitive deviation characteristics while balancing the computational efficiency and prediction accuracy of quantitative evaluation, providing guidance for the design and manufacturing tolerance formulation of turbomachinery.

[0014] Furthermore, the parametric modeling of the geometric deviation characteristics of the turbomachinery blade profile, extraction of geometric deviation parameters, and sampling to construct a geometric deviation parameter sample specifically includes:

[0015] Extract the geometric deviation features under the real operating environment as the geometric deviation features of the impeller blade profile, and perform mathematical parameterization modeling on the geometric deviation features of the impeller blade profile to construct a geometric deviation input parameter vector as the geometric deviation parameter;

[0016] The range of values ​​for each parameter in the geometric deviation input parameter vector is determined, and a geometric deviation parameter sample set containing multiple sets of geometric deviation parameter samples is constructed by sampling within the range of values ​​of each parameter using a random sampling method.

[0017] By adopting the above technical solution, the value range of each parameter is defined and sampled based on the real operating environment, ensuring that the generated geometric deviation parameter sample set has physical meaning and representativeness, and establishing the input data foundation for the subsequent training of the proxy model.

[0018] Furthermore, the step of performing deviation modeling on the original blade design based on the geometric deviation parameter samples to generate a three-dimensional blade simulation model specifically includes:

[0019] For each set of geometric deviation parameter samples in the geometric deviation parameter sample set, a deviation air profile is generated at the target air profile section of the original design air profile to produce a real air profile.

[0020] The actual blade profile is superimposed with the original design blade profile at the other blade heights except the target blade height section along the blade radial direction to reconstruct a three-dimensional blade geometric model, and the three-dimensional blade geometric model is output as the three-dimensional blade simulation model.

[0021] By adopting the above technical solution, the deviated blade profile is shaped for the target blade height section and stacked along the blade radial direction, thus realizing the accurate restoration of local geometric deviations and ensuring that the established three-dimensional blade simulation model can reflect the spatial morphological variation of the blade during service.

[0022] Furthermore, the geometric deviation characteristics under the actual operating environment include deviations in the actual geometric shape caused by fouling, and the geometric deviation input parameter vector includes the starting position and the intensity of fouling thickening.

[0023] During the stacking process along the blade radial direction, a smooth transition process is used in the vicinity of the target blade height section to form a smooth transition zone for geometric deviation, ensuring that the surface of the three-dimensional blade geometric model is smooth and continuous.

[0024] By adopting the above technical solution, the geometric abrupt change at the junction of the real airfoil and the original design airfoil is eliminated by the smooth transition processing, which avoids the three-dimensional mesh distortion caused by surface discontinuity during aerodynamic simulation calculation, and ensures the stability and convergence accuracy of the calculation process.

[0025] Furthermore, the aerodynamic simulation calculation of the three-dimensional blade simulation model, the extraction of aerodynamic indices of the geometric deviation parameter samples under the same operating conditions using a working condition control strategy, and the construction of a mapping database between geometric deviations and aerodynamic indices specifically include:

[0026] At the design speed, the original design airfoil is scanned under all operating conditions to obtain the highest efficiency point, and the benchmark operating condition parameters corresponding to the highest efficiency point are extracted.

[0027] For each set of geometric deviation parameter samples in the geometric deviation parameter sample set, CFD calculation is performed as the aerodynamic simulation calculation. While maintaining the same inlet boundary conditions as when the reference operating condition parameters were obtained, the outlet back pressure corresponding to each set of geometric deviation parameter samples is dynamically adjusted as the operating condition control strategy to force the actual aerodynamic parameters of each set of geometric deviation parameter samples to be equal to the reference operating condition parameters.

[0028] After meeting the same working conditions, the aerodynamic index corresponding to each set of geometric deviation parameter samples is extracted to form an aerodynamic index output vector. The geometric deviation input parameter vectors corresponding to all the geometric deviation parameter samples are combined with the aerodynamic index output vectors to construct the geometric deviation and aerodynamic index mapping database.

[0029] By adopting the above technical solution, the operating condition control strategy of dynamically adjusting the outlet back pressure to force the actual aerodynamic parameters to be equal to the reference operating condition parameters eliminates the operating error caused by the flow deviation under a single back pressure setting, ensures that all the aerodynamic index extractions are based on a consistent aerodynamic operating point, and improves the physical rigor of the correspondence between the geometric deviation and the input and output in the aerodynamic index mapping database.

[0030] Furthermore, the reference operating parameters include the reference value of the inlet relative Mach number and the reference value of the relative airflow angle; the aerodynamic indicators include isentropic efficiency and pressure ratio.

[0031] By adopting the above technical solution and selecting aerodynamic and thermodynamic parameters as the benchmark for determining operating conditions and evaluating performance, the evaluation method can be applied to the conventional design and assessment system of turbomachinery, thus enhancing the versatility of the method.

[0032] Furthermore, the step of training a surrogate model using the geometric deviation and aerodynamic index mapping database to establish a nonlinear mapping relationship between the geometric deviation parameters and the aerodynamic index specifically includes:

[0033] A fully connected neural network is selected as the surrogate model. The input layer of the fully connected neural network is set as the geometric deviation input parameter vector, and the output layer of the fully connected neural network is set as the aerodynamic index output vector.

[0034] The geometric deviation and aerodynamic index mapping database is divided into a training set and a validation set, and the geometric deviation and aerodynamic index mapping database is used as a data source to input into the fully connected neural network for model training.

[0035] During the model training process, the K-fold cross-validation method is used to optimize and validate the hyperparameters of the fully connected neural network, and the optimal network hyperparameters with the highest prediction accuracy on the validation set are selected to establish the data-driven nonlinear mapping relationship.

[0036] By adopting the above technical solution, the fitting ability of the fully connected neural network is used to replace the computationally intensive CFD solution process, and the hyperparameters of the network are optimized by the K-fold cross-validation method, thereby suppressing the overfitting tendency of the fully connected neural network and constructing the surrogate model with generalization ability.

[0037] Furthermore, the step of using the surrogate model to infer aerodynamic performance from randomly generated geometric deviation parameter samples to complete the uncertainty quantification assessment specifically includes:

[0038] Given a geometric deviation probability distribution, a randomly generated sample of geometric deviation parameters containing each of the geometric deviation parameters is generated using the Monte Carlo method.

[0039] The randomly generated geometric deviation parameter samples are input into the fully connected neural network to obtain the aerodynamic performance prediction results corresponding to the randomly generated geometric deviation parameter samples;

[0040] The aerodynamic performance prediction results of all the randomly generated geometric deviation parameter samples are statistically analyzed and the mean of the aerodynamic performance prediction results is calculated. The design point performance of the original airfoil is compared with the mean of the aerodynamic performance prediction results. The performance degradation between the original airfoil and the randomly generated geometric deviation parameter samples is calculated to complete the uncertainty quantification assessment.

[0041] By adopting the above technical solution, and using the Monte Carlo method combined with the fully connected neural network to perform inference on multiple sets of data, it is possible to statistically analyze the probability distribution characteristics of aerodynamic performance under the combined effect of multidimensional random variables, and realize the quantitative measurement of the performance degradation.

[0042] Furthermore, the calculation of the contribution of each geometric deviation parameter in the randomly generated geometric deviation parameter sample using interpretable machine learning methods specifically includes:

[0043] The SHAP interpretable machine learning method is selected as the interpretable machine learning method. The aerodynamic performance prediction results of the fully connected neural network are mathematically decomposed, and the marginal contribution of each geometric deviation parameter input into the fully connected neural network is calculated as the contribution.

[0044] The SHAP interpretability machine learning method is used to decompose the aerodynamic performance prediction result corresponding to a single randomly generated geometric deviation parameter sample into the sum of the aerodynamic performance prediction result baseline value and the marginal contribution of each geometric deviation parameter to the aerodynamic performance prediction result.

[0045] By adopting the above technical solution and introducing the SHAP interpretable machine learning method to analyze the fully connected neural network, the internal quantitative composition of a single prediction result is clarified, and the decoupling and calculation of the influence of each geometric deviation parameter on the aerodynamic performance prediction result are realized.

[0046] Furthermore, the identification of geometric features highly sensitive to aerodynamic performance specifically includes:

[0047] After calculating the marginal contribution of each of the geometric deviation parameters, the global weight of the marginal contribution of each of the geometric deviation parameters in all the randomly generated geometric deviation parameter samples is statistically evaluated, and a bar chart containing the marginal contribution values ​​of the features is output. The highly sensitive geometric features that cause performance fluctuations are identified as the geometric features that are highly sensitive to aerodynamic performance, and a sensitivity analysis report containing the geometric features that are highly sensitive to aerodynamic performance is generated.

[0048] By adopting the above technical solution, and by statistically analyzing the global weights and outputting the bar chart, the geometric features that are highly sensitive to aerodynamic performance and cause performance fluctuations are identified, providing a quantitative basis for decision-making regarding modifying machining tolerances or developing cleaning and maintenance plans.

[0049] This invention provides a method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation. It has the following beneficial effects:

[0050] 1. This invention generates a real blade profile by performing a deviation blade profile shaping at the target blade height section of the original design blade profile, and then superimposes the real blade profile with the original design blade profile at other blade heights along the blade radial direction to reconstruct a three-dimensional blade simulation model. This overcomes the defect of the evaluation method based on two-dimensional blade cascades that cannot consider radial flow effects, thereby objectively reflecting the impact of the deviation of the real geometric shape caused by fouling on the overall performance dispersion of the blade, and improving the engineering applicability of the quantitative evaluation method for the performance uncertainty of turbomachinery blade profile geometric deviation.

[0051] 2. This invention performs aerodynamic simulation calculations on a three-dimensional blade simulation model and adopts a dynamic adjustment of the outlet back pressure as an operating condition control strategy to force the actual aerodynamic parameters of each set of geometric deviation parameter samples to be equal to the baseline operating condition parameters. This eliminates the operating error caused by flow deviation under a single back pressure setting, thereby ensuring that all aerodynamic index extractions are based on a consistent aerodynamic operating point. It objectively reflects the uncertainty caused solely by geometric deviation characteristics and the impact on the performance degradation of the original design blade profile, ensuring the physical rigor of the evaluation data.

[0052] 3. This invention uses a fully connected neural network as a surrogate model to infer the aerodynamic performance of randomly generated geometric deviation parameter samples, and combines the SHAP interpretability machine learning method to calculate the marginal contribution of each geometric deviation parameter. While completing the uncertainty quantification assessment, it also calculates the global weight to locate the geometric features that are highly sensitive to aerodynamic performance and cause performance fluctuations. This achieves the decoupling and calculation of the influence of each geometric deviation parameter on the aerodynamic performance prediction results, and provides a quantitative decision-making basis for formulating cleaning and maintenance plans. Attached Figure Description

[0053] Figure 1 A flowchart illustrating a method for quantifying the uncertainty of performance of turbomachinery blade geometry deviation according to an embodiment of the present invention;

[0054] Figure 2 This is a cross-sectional comparison diagram of an unbiased original airfoil and an airfoil with fouling deviation according to an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of a three-dimensional rotor blade with a geometric deviation of fouling at a specific blade height position according to an embodiment of the present invention;

[0056] Figure 4 A schematic diagram illustrating the process of constructing a high-fidelity database of geometric deviations and aerodynamic performance according to an embodiment of the present invention;

[0057] Figure 5 A schematic diagram of a single-channel rotor blade geometric model for CFD calculation using input simulation software according to an embodiment of the present invention;

[0058] Figure 6 This is a histogram of aerodynamic performance dispersion based on a three-dimensional blade environment according to an embodiment of the present invention;

[0059] Figure 7 A graph showing the results of sensitivity analysis of geometric deviation parameters based on a three-dimensional blade environment according to an embodiment of the present invention;

[0060] Figure 8 This is a histogram of the dispersion of aerodynamic performance based on a traditional two-dimensional cascade model.

[0061] Figure 9 The figure shows the results of sensitivity analysis of geometric deviation parameters based on the traditional two-dimensional cascade model.

[0062] Figure 10 This is a comparison chart of the actual and predicted values ​​of isentropic efficiency according to an embodiment of the present invention;

[0063] Figure 11 This is a schematic diagram illustrating the probability distribution of turbomachinery performance degradation caused by uncertain factors. Detailed Implementation

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] See attached document Figure 1 This invention provides a method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation, comprising the following steps:

[0066] S10. Parametric modeling is performed on the geometric deviation characteristics of the impeller blade in the real operating environment, and parameter samples are extracted. Based on the parameter samples, the original blade design is shaped with deviation, and then a three-dimensional blade CFD simulation model containing the blade with geometric deviation is generated.

[0067] S20. Perform CFD calculations on the generated three-dimensional blade CFD simulation model. Ensure that all simulation samples operate under the same conditions by dynamically adjusting the outlet back pressure. Extract key aerodynamic indicators to construct a high-fidelity mapping database of geometric deviation and aerodynamic performance.

[0068] S30. Use a high-fidelity mapping database to train and optimize the surrogate model, and establish a nonlinear mapping relationship between the input geometric deviation parameters and the output aerodynamic performance indicators to replace CFD calculations.

[0069] S40. Using a trained surrogate model combined with the Monte Carlo method, aerodynamic performance is inferred from a large number of randomly generated geometric deviation inference samples to achieve uncertainty quantification. At the same time, an interpretable machine learning method is introduced to calculate the contribution of each geometric deviation parameter, thereby identifying the geometric features most sensitive to performance.

[0070] To enable those skilled in the art to better understand the technical solution and computational logic of the present invention, the key steps in the above method flow will be described in detail below with reference to the specific accompanying drawings.

[0071] See attached document Figure 2 and attached Figure 3 To accurately reproduce the geometric changes of the blade profile under real service conditions in three-dimensional space, step S10 involves parametric modeling of the geometric deviation characteristics of the turbomachinery blade profile in the real operating environment and extracting parameter samples. Based on the parameter samples, the original design blade profile is modeled with deviations, and then a three-dimensional blade CFD simulation model containing the geometric deviation blade profile is generated. This step specifically includes the following sub-steps:

[0072] S101. Determine the research object and geometric deviation type of the turbomachinery. The research object specifically includes blades in single-stage or multi-stage turbomachinery systems, such as fans, compressors, and turbine blades. The research object of the turbomachinery in this invention is not limited to axial-flow rotor blades, but also applies to stator blades, inlet guide vanes (IGVs), and energy conversion system components such as centrifugal compressors and radial turbines that contain subsonic, transonic, or supersonic blade profiles. Geometric deviations specifically include deviations from the designed geometry of the blade due to manufacturing tolerances, assembly errors, and environmental factors during actual service. Deviations in the actual geometry caused by environmental factors are specifically manifested as fouling, wear, or corrosion. Furthermore, geometric deviations can be equivalently replaced by features such as blade profile machining contour errors, installation angle errors, tip clearance variations, or blade surface roughness, and then parametrically modeled.

[0073] S102. Extract geometric deviation features from the actual operating environment and perform mathematical parameterization modeling on these features to construct a geometric deviation input parameter vector. In practice, geometric deviation parameterization modeling is typically based on blade profile design within the geometric deviation range, using NUMECA's AutoBlade module or custom-written blade profile design software or programs. Taking the geometric deviation caused by surface fouling during actual blade service as an example, refer to... Figure 2 Based on measurements of various cross-sections of real blades, it was found that fouling typically thickens the suction surface of the blade. Based on this physical phenomenon, selecting the starting position and intensity of fouling thickening as parametric features can effectively characterize the actual degree of geometric deviation. The geometric deviation input parameter vector for a single blade profile is defined as follows: It satisfies:

[0074] ;

[0075] in, Input parameter vector for geometric deviation, This is the starting point for the buildup of dirt. To increase strength. Location of initial thickening of scale buildup. Specifically, it refers to the normalized position along the chord length of the blade, and the increased strength. Specifically, it refers to the maximum relative thickness of the accumulated dirt.

[0076] S103. Determine the value range of each parameter and sample within the parameter space to construct a geometric deviation parameter sample set. Based on the measured blade fouling characteristic data, determine the starting position of fouling thickening. With increased strength The range of values. Using Latin hypercube sampling or Monte Carlo sampling methods, samples are drawn within the defined range. A set of random simulation samples, set The value of is 40, and all extracted simulation samples are collected into a geometric deviation parameter sample set. For the specific sampling process of the Latin hypercube sampling method or the Monte Carlo sampling method, those skilled in the art can directly implement it based on the preset number of simulation samples and conventional random sampling algorithms. The specific algorithm execution logic is well-known in the field and will not be elaborated here.

[0077] S104. Based on simulation samples from the geometric deviation parameter sample set, the original design airfoil is subjected to deviation modeling, and then superimposed to generate a 3D blade CFD simulation model containing the geometrically deviated airfoil. For any set of simulation samples in the geometric deviation parameter sample set, deviation airfoil modeling is implemented at a specific blade height section of the original design airfoil without deviation, generating a real airfoil with corresponding fouling thickening characteristics. When implementing deviation airfoil modeling, the airfoil geometry at other blade heights, except for the specific blade height section and its adjacent range, remains unchanged. The real airfoil containing geometric deviation and the original airfoil at the remaining blade heights are superimposed along the blade radial direction to reconstruct a 3D blade geometric model. To ensure the surface smoothness and continuity of the generated 3D blade geometric model, a smooth transition processing is adopted in the adjacent range of the specific blade height section, referring to... Figure 3 Taking the fouling deviation at 50% blade height as an example, the actual generated CFD simulation model, compared with the original blade, not only produces geometric deviation at the specific blade height section corresponding to the blade shape of the research object, but also has a certain geometric deviation in the vicinity of the specific blade height section, thus forming a smooth transition zone of geometric deviation. Finally, the reconstructed three-dimensional blade geometric model is output as a three-dimensional blade CFD simulation model with geometric deviation blade shape.

[0078] See attached document Figure 4 and attached Figure 5 After completing the parametric reconstruction of the three-dimensional blade geometry model, in order to obtain flow field and aerodynamic performance data for subsequent model training and analysis, step S20 is performed to conduct CFD calculations on the generated three-dimensional blade CFD simulation model. A dynamic adjustment of the outlet back pressure control strategy ensures that all simulation samples operate under the same conditions, and key aerodynamic indicators are extracted to construct a high-fidelity mapping database of geometric deviations and aerodynamic performance. This step specifically includes the following sub-steps:

[0079] S201. Based on the generated 3D blade CFD simulation model containing geometric deviations, refer to... Figure 5 A three-dimensional fluid dynamics computational domain was established and computational settings were configured. A single-channel or multi-channel CFD simulation computational domain was established, employing the Spalart-Allmaras turbulence model to capture boundary layer flow. Simultaneously, the height of the first layer of meshes at the wall was set to satisfy the dimensionless wall distance. The boundary conditions are set to satisfy the solution requirements of the turbulence model for viscous sublayer flow. The specific mesh generation method and the execution logic for mesh independence verification in the three-dimensional fluid dynamics computational domain can be directly implemented by those skilled in the art based on conventional fluid dynamics simulation specifications and fundamental principles of computational fluid dynamics. The mesh generation and independence verification methods are well-known techniques in this field and will not be elaborated upon here.

[0080] S202. Obtain the baseline operating parameters of the unbiased original three-dimensional blade at the design speed. Perform CFD simulation on the unbiased original three-dimensional blade at the rated design speed. While keeping the inlet total temperature and total pressure constant, perform a full-condition scan by gradually adjusting the outlet back pressure to obtain the highest efficiency point. Extract the baseline value of the inlet relative Mach number corresponding to this highest efficiency point. relative airflow angle reference value .

[0081] S203. Utilizing a dynamic backpressure adjustment control strategy, perform identical operating condition control calculations on all 3D blade CFD simulation models containing geometrically deviated airfoils. For each set of simulation samples in the geometric deviation parameter sample set, perform CFD calculations. While maintaining the same inlet boundary conditions as when obtaining the baseline operating condition parameters, dynamically adjust the outlet backpressure of the current simulation sample to force the actual aerodynamic parameters of the current simulation sample to equal the baseline parameters. In real turbomachinery operation, airfoil geometric deviations change the flow channel cross-sectional area and cause changes in internal flow blockage, leading to a shift in the actual operating point of the blade on the characteristic diagram. To prevent interference from operating condition uncertainties, it is necessary to control the inlet parameters consistently, thereby ensuring that the aerodynamic performance evaluation of all geometrically deviated simulation samples is performed under the same aerodynamic load. For the first... The simulation sample satisfies the following operating control conditions:

[0082] ;

[0083] ;

[0084] in, For the first The import relative Mach number of a simulation sample, This is the import relative Mach number benchmark value. For the first The inlet relative airflow angle of a simulated sample. This is the reference value for the relative airflow angle. Let be the simulation sample number in the geometric deviation parameter sample set, and The value range is from 1 to positive integers, This represents the total number of simulation samples. Furthermore, in other embodiments, mass flow rate and other parameters can be selected as the consistency control target for operating conditions, depending on the specific research needs of turbomachinery. This consistency can be achieved through dynamic adjustment to ensure that the constraint parameters such as mass flow rate remain consistent across all simulation samples.

[0085] S204. Extract the key aerodynamic parameters of each simulation sample under the same operating conditions, and combine them to construct a high-fidelity mapping database. After meeting the above-mentioned operating conditions under the same conditions, extract the... The isentropic efficiency corresponding to each simulation sample With pressure ratio As a key aerodynamic indicator, in other specific application scenarios, the key aerodynamic indicator can be replaced by or additionally supplemented with macroscopic and microscopic flow field characteristics such as static pressure ratio, axial aerodynamic parameter distribution, radial aerodynamic parameter distribution, or circumferential aerodynamic parameter distribution, and constitute the first... Aerodynamic performance output vector of each simulation sample It satisfies:

[0086] ;

[0087] in, For the first The aerodynamic performance output vector of each simulation sample. For the first The isentropic efficiency of a simulated sample. For the first The pressure ratio of each simulated sample. (Refer to...) Figure 4 For all 3D blade CFD simulation models containing geometric deviations, the sample numbers defined above are used. Traverse sequentially, starting from number 1, number 2, and so on. Until the total number of simulation samples is reached For each deviation geometric sample, the corresponding efficiency and pressure ratio data are extracted. After completing all calculations and data extraction, the geometric deviation input parameter vectors and aerodynamic performance output vectors corresponding to all simulation samples are combined one-to-one to finally construct a complete high-fidelity mapping database of geometric deviations and aerodynamic performance.

[0088] Since directly performing CFD calculations on a large number of geometric deviation simulation samples would consume significant computational resources, to improve the efficiency of uncertainty assessment, after obtaining the aforementioned high-fidelity database, step S30 is performed to train and optimize the surrogate model using the high-fidelity mapping database, establishing a nonlinear mapping relationship between input geometric deviation parameters and output aerodynamic performance indicators to replace CFD calculations. This step specifically includes the following sub-steps:

[0089] S301. Determine the surrogate model architecture for the alternative computational fluid dynamics simulation, and establish the correspondence between the model's input and output layers. A fully connected neural network is selected as the surrogate model to establish the data fitting structure. Depending on the requirements of the actual engineering application, the surrogate model construction method can also be a Kriging model, multinomial chaotic expansion, or support vector machine. The input layer of the surrogate model is set as the geometric deviation input parameter vector. Specifically, it includes geometric features such as the starting location and intensity of fouling thickening; the output layer of the surrogate model is set as the aerodynamic performance output vector. Specifically, this includes aerodynamic indicators such as isentropic efficiency and pressure ratio. Regarding the construction of the basic structure of a fully connected neural network, such as the number of network layers and neurons, those skilled in the art can directly implement it using conventional machine learning frameworks based on preset fitting accuracy requirements. The basic network construction methods are well-known technologies in this field and will not be elaborated upon here.

[0090] S302. The model is trained using the constructed high-fidelity mapping database to establish a data-driven nonlinear mapping relationship. The high-fidelity mapping database of geometric deviations and aerodynamic performance constructed in the above steps is divided into a training set and a validation set. The high-fidelity mapping database is used as the data source and input into a fully connected neural network for training. During training, the neural network performs forward propagation and nonlinear activation operations on the input geometric deviation features through the weights and biases of multiple layers of neurons, thereby approximating the real physical laws in fluid dynamics simulation and ultimately establishing a nonlinear mapping relationship that replaces the three-dimensional computational fluid dynamics solver. The mapping relationship satisfies:

[0091] ;

[0092] in, This is the aerodynamic performance output vector. Input parameter vector for geometric deviation, This is a non-linear mapping relationship.

[0093] S303. K-fold cross-validation is used to optimize the hyperparameters of the surrogate model, ensuring its generalization ability and prediction accuracy in the multidimensional geometric bias parameter space. During the training of the surrogate model, K-fold cross-validation is used to optimize and verify the network hyperparameters. Specifically, the pre-divided training data is divided into equal parts... Given a set of disjoint subsets, in each iteration, one subset is selected as the validation set, and the remaining subsets are used as the validation set. A subset is used as the training set to perform... The process of model training and error evaluation. The number of folds in the cross-validation is a positive integer greater than 1. Taking 5-fold cross-validation as an example, through multiple data folding and iterations, the optimal network hyperparameters with the highest prediction accuracy on the validation set are selected, thereby eliminating the overfitting risk caused by a single data partition and improving the surrogate model's ability to generalize predictions of unknown geometric biases. The specific optimization process for updating the internal weights of the surrogate model through backpropagation based on the loss function can be directly implemented by those skilled in the art using conventional gradient descent algorithms. The underlying parameter update mechanism of the model is a well-known technique in this field and will not be elaborated upon here.

[0094] See attached document Figure 6 and attached Figure 7 After successfully constructing and optimizing the surrogate model, a large amount of data inference and statistical analysis can be performed based on this model. Step S40 involves using the trained surrogate model combined with the Monte Carlo method to infer aerodynamic performance from a large number of randomly generated geometric deviation inference samples to quantify uncertainty. Simultaneously, interpretable machine learning methods are introduced to calculate the contribution of each geometric deviation parameter, thereby identifying geometric features highly sensitive to performance. This step specifically includes the following sub-steps:

[0095] S401. Using the trained surrogate model combined with the Monte Carlo method, aerodynamic performance is inferred from a large number of randomly generated geometric deviation inference samples. Under a given geometric deviation probability distribution, the Monte Carlo method is used to randomly sample and generate... A number of inference samples, set The value is set to 10000, generating a geometric deviation parameter inference sample that includes features such as the starting location and intensity of scale thickening. This... A randomly generated geometric deviation parameter inference sample is input into the optimized surrogate model. Using the established nonlinear mapping relationship, the aerodynamic performance prediction result corresponding to each inference sample is inferred. The established nonlinear mapping avoids the time-consuming direct three-dimensional fluid dynamics simulation calculation, satisfying the Monte Carlo method's requirement for a large sample size while ensuring accuracy. For generating a large number of random samples following a probability distribution, those skilled in the art can directly implement this using conventional random number generation functions. The specific probability distribution sampling process is a well-known technique in the field and will not be elaborated here.

[0096] S402. Statistical inference of aerodynamic performance data and calculation of distribution characteristics to quantify uncertainty. Extract aerodynamic performance inference results from all inference samples and plot frequency distribution histograms of performance parameters such as isentropic efficiency and pressure ratio (e.g., Figure 6 (As shown). Taking isentropic efficiency as an example, this single-dimensional performance index is used as the overall performance prediction value. Calculate the mean of the performance prediction values It satisfies:

[0097] ;

[0098] in, The mean of the performance predictions. To infer the sample size, For the first The overall performance prediction value of the inferred sample. To infer the sample number, its value ranges from 1 to... Positive integers. By comparing the mean and performance dispersion of the design point performance under no geometric deviation conditions with the predicted performance under actual deviation conditions, and calculating the difference between the two, we can obtain results such as... Figure 6 The isentropic efficiency degradation shown Pressure ratio degradation By extracting the aforementioned performance degradation, a quantitative assessment of the uncertainties in the actual operating environment of the three-dimensional blade is completed.

[0099] S403. The SHAP interpretability machine learning method is introduced for sensitivity analysis to calculate the marginal contribution of each geometric deviation parameter to the overall performance prediction value. To analyze the geometric factors leading to performance dispersion and degradation, global sensitivity analysis methods such as the SHAP method or the Sobol exponent method are introduced to mathematically decompose the single prediction results of the surrogate model. This embodiment specifically uses the SHAP method as an example. The SHAP method is based on the Shapley value theory in cooperative game theory. By calculating the expected marginal contribution of each feature under different feature subset combinations, it achieves the summative decomposition of the nonlinear model prediction results. The marginal contribution of each geometric deviation parameter in the input surrogate model to the aerodynamic performance prediction deviation is calculated, and the first... Overall performance prediction value of each inferred sample Benchmark value decomposed into overall performance prediction value With the The geometric feature parameter for the th geometric feature parameter Marginal contribution of the overall performance prediction value of each inferred sample The sum of them. It satisfies:

[0100] ;

[0101] in, For the first The overall performance prediction value of the inferred sample. This serves as the baseline value for the overall performance prediction. For geometric parameters, dimension For the first The geometric feature parameter for the th geometric feature parameter Marginal contribution of the overall performance prediction value of each inferred sample These are the numbers of the geometric feature parameters, with values ​​ranging from 1 to... Positive integers. In the evaluation of geometric deviations caused by fouling on the blade surface in this invention, the geometric parameter dimension... The value is 2.

[0102] S404. Identify highly sensitive regions based on the global marginal contribution of each geometric feature parameter, and output sensitivity assessment results to guide design. By statistically evaluating the global weight of the marginal contribution of each geometric parameter in all inferred samples, the global impact of parameters such as thickening location and intensity on aerodynamic performance uncertainties is quantified. Figure 7The system outputs a bar chart containing the Feature Marginal Contribution (SHAP) values, which helps locate highly sensitive geometric parameters causing performance fluctuations. Combining these identified highly sensitive geometric regions, such as blade leading edges or tips, a sensitivity analysis report is generated to guide robust design of turbomachinery targeting these highly sensitive areas, or to guide the development of more reasonable cleaning and maintenance schedules. For specific engineering implementation plans based on local structural optimization or cleaning and maintenance schedules for highly sensitive areas, those skilled in the art can directly implement them using conventional mechanical design and manufacturing specifications and equipment maintenance manuals. The specific engineering application and transformation process is well-known in the field and will not be elaborated upon here.

[0103] Specific application examples:

[0104] To better understand the technical solution of this invention, the following detailed description is provided in conjunction with specific application scenarios and accompanying drawings. This specific application embodiment is built upon a performance uncertainty quantification evaluation system for fouling and contamination of transonic compressor rotor blades. This system runs on a high-performance computing node and is responsible for acquiring three-dimensional blade geometric deviation data and performing aerodynamic performance inference and sensitive feature identification.

[0105] In the stage of geometric deviation parametric modeling and computational fluid dynamics simulation, the modeling module needs to extract the fouling characteristics of the blades during actual service to calculate the geometric deviation input parameter vector. The parameters were set as follows: The compressor rotor was set to thicken with fouling at 50% blade height, and the starting position of fouling thickening in the extracted sample was obtained. The thickness is 0.3, and the strength is increased by 0.3. The value is 0.04. The modeling module substitutes the input parameter vector into the formula to calculate the geometric deviation. :

[0106] ;

[0107] The numerical calculation process is as follows:

[0108] ;

[0109] The calculation results show that the geometric deviation input parameter vector obtained by the system for Similarly, when the preset operating condition control conditions are met, namely the inlet relative Mach number reference value... The relative airflow angle reference value is 1.2. After reaching 55°, the simulation module extracts the key aerodynamic parameters of the corresponding simulation sample. (The remaining text appears to be incomplete and requires further context.) isentropic efficiency of a simulated sample It is 0.875, the first Pressure ratio of a simulated sample The value is 1.58. The simulation module substitutes the formula to calculate the... Aerodynamic performance output vector of each simulation sample :

[0110] ;

[0111] The numerical calculation process is as follows:

[0112] ;

[0113] Calculation results show that, under the same working conditions, the system obtains the first... Aerodynamic performance output vector of each simulation sample for The modeling and simulation module establishes a high-fidelity database by mapping various geometric deviation parameters to aerodynamic performance indicators in sequence, thus training the neural network.

[0114] The uncertainty quantification inference module performs Monte Carlo random sampling and aerodynamic performance inference steps to generate the mean of the performance prediction values. The parameters are set as follows: Preset number of inference samples. Given 10000, obtain 10000 inference samples and infer the first inference sample using the surrogate model. Overall performance prediction value of each inferred sample The sum The value is 8820. The uncertainty quantification inference module substitutes the values ​​into the formula to calculate the mean of the performance predictions. :

[0115] ;

[0116] The numerical calculation process is as follows:

[0117] ;

[0118] The calculation results show that, after introducing geometric bias uncertainty, the mean of the performance prediction value is... The value is 0.882. This value is lower than the unbiased baseline efficiency of 0.890, indicating the isentropic efficiency degradation. With a value of 0.008, the system achieves a quantitative assessment of macroscopic performance degradation.

[0119] The interpretability sensitivity analysis module performs prediction result decomposition and marginal contribution assessment steps to output decomposed constraints for overall performance degradation. Parameter settings are as follows:

[0120] Extract the currently inferred first Overall performance prediction value of each inferred sample The baseline value is 0.875, representing the preset overall performance prediction. 0.890; set the geometric parameter dimension. The value is 2. This sets the number of the geometric feature parameter. Taking values ​​of 1 and 2 respectively, we obtain the first geometric feature parameter for the first... Marginal contribution of the overall performance prediction value of each inferred sample The value is -0.005, the second geometric feature parameter is related to the first... Marginal contribution of the overall performance prediction value of each inferred sample The value is -0.010. The interpretability sensitivity analysis module is used to verify the formula. Overall performance prediction value of each inferred sample :

[0121] ;

[0122] The numerical calculation process is as follows:

[0123] ;

[0124] The calculation results show that the formula equation holds true. The interpretability sensitivity analysis module reads each marginal contribution parameter and decomposes the overall performance degradation value, enabling the control system to perform targeted robust optimization design for highly sensitive geometric regions.

[0125] To verify the accuracy of the aforementioned fully connected neural network surrogate model in replacing time-consuming 3D fluid dynamics simulations, the system compared and output the predicted distribution data of the validation set simulation samples. (See attached...) Figure 10 , Figure 10 The horizontal axis represents the true value of isentropic efficiency in computational fluid dynamics simulation, and the direction of the horizontal axis represents the progression of actual performance testing. The vertical axis represents the predicted value of isentropic efficiency from the surrogate model. Due to the lack of error interference, the ideal baseline in the figure exhibits a diagonal distribution along the isentropic efficiency values. The predicted sample points representing the validation set of this invention, established based on a data-driven nonlinear mapping relationship, show a trend of conforming to and converging near the ideal baseline in their predicted isentropic efficiency values. This result verifies that the surrogate model established in this embodiment possesses the ability to capture the nonlinear mapping relationship between geometric deviations and aerodynamic performance.

[0126] Based on the completed surrogate model inference, to further verify the effectiveness of the three-dimensional environment assessment method of this invention in capturing real performance degradation, the performance dispersion under different dimensional models was systematically compared. (See appendix) Figure 6 and attached Figure 8 In both figures, the vertical axis represents the frequency distribution, and the horizontal axis represents the aerodynamic performance parameter values ​​under different evaluation models. Specifically, Figure 8The histogram obtained by the traditional two-dimensional leaf cascade method has the total pressure loss coefficient on the horizontal axis (the figure further indicates the difference between the pressure loss coefficient at the design point and the mean, i.e., the degradation of the total pressure loss coefficient). Because this method lacks a constraint mechanism for three-dimensional spatial integral effects such as radial flow redistribution, its performance dispersion distribution in the chart deviates from the actual fluctuation range. Figure 6 The histogram representing the three-dimensional real blade evaluation method of this invention is shown in the figure. The horizontal axis represents isentropic efficiency and pressure ratio. Because this method preserves the complete three-dimensional flow field and radial flow coupling mechanism, its distribution exhibits an approximately normal distribution and a significant deviation from the baseline value. This result verifies the ability of this invention to capture macroscopic performance degradation characteristics under real three-dimensional operating conditions.

[0127] Furthermore, to demonstrate the accuracy of this invention in identifying sensitive geometric features, the system statistically analyzed the changes in the marginal contribution weights of each deviation parameter under different dimensional models. (See attached figure.) Figure 7 and attached Figure 9 , Figure 7 and Figure 9 The horizontal axis represents the global absolute average value of SHAP, the direction of the horizontal axis represents the shift of the influencing weights, and the vertical axis represents the geometric feature parameters. Figure 9 The bar chart output by the traditional two-dimensional blade cascade model ignores the response characteristics of the three-dimensional pressure gradient field to changes in the shock wave structure, resulting in the weight of the thickening intensity remaining within the main numerical range, while the weight of the starting position of fouling thickening is at a low level. Figure 7 The bar chart of the three-dimensional blade model of this invention dynamically quantifies the influence intensity of spanwise boundary layer migration, ensuring that the initiation location and intensity of fouling thickening maintain comparable weights in the sensitivity ranking. This result verifies that the present invention has significant technical advantages in correcting sensitivity misjudgments and identifying highly sensitive areas that cause performance fluctuations.

Claims

1. A method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation, characterized in that, Includes the following steps: Parametric modeling is performed on the geometric deviation characteristics of the turbomachinery blade profile to extract geometric deviation parameters and sample to construct a geometric deviation parameter sample. Based on the geometric deviation parameter sample, the original blade profile is modeled with deviation to generate a three-dimensional blade simulation model. Aerodynamic simulation calculations are performed on the three-dimensional blade simulation model. The aerodynamic indices of the geometric deviation parameter samples under the same working conditions are extracted using the working condition control strategy, and a mapping database of geometric deviation and aerodynamic indices is constructed. A surrogate model is trained using the geometric deviation and aerodynamic index mapping database to establish a nonlinear mapping relationship between the geometric deviation parameters and the aerodynamic index; The surrogate model is used to perform aerodynamic performance inference on randomly generated geometric deviation parameter samples to complete uncertainty quantification assessment. In addition, interpretable machine learning methods are used to calculate the contribution of each geometric deviation parameter in the randomly generated geometric deviation parameter samples to identify geometric features that are highly sensitive to aerodynamic performance.

2. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 1, characterized in that, The specific steps of parametrically modeling the geometric deviation characteristics of the turbomachinery blade profile, extracting geometric deviation parameters, and sampling to construct a geometric deviation parameter sample include: Extract the geometric deviation features under the real operating environment as the geometric deviation features of the impeller blade profile, and perform mathematical parameterization modeling on the geometric deviation features of the impeller blade profile to construct a geometric deviation input parameter vector as the geometric deviation parameter; The range of values ​​for each parameter in the geometric deviation input parameter vector is determined, and a geometric deviation parameter sample set containing multiple sets of geometric deviation parameter samples is constructed by sampling within the range of values ​​of each parameter using a random sampling method.

3. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 2, characterized in that, The step of performing deviation modeling on the original blade design based on the geometric deviation parameter samples to generate a three-dimensional blade simulation model specifically includes: For each set of geometric deviation parameter samples in the geometric deviation parameter sample set, a deviation air profile is generated at the target air profile section of the original design air profile to produce a real air profile. The actual blade profile is superimposed with the original design blade profile at the other blade heights except the target blade height section along the blade radial direction to reconstruct a three-dimensional blade geometric model, and the three-dimensional blade geometric model is output as the three-dimensional blade simulation model.

4. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 3, characterized in that, The geometric deviation characteristics under the real operating environment include deviations in the real geometric shape caused by fouling, and the geometric deviation input parameter vector includes the starting position and the intensity of fouling thickening. During the stacking process along the blade radial direction, a smooth transition process is used in the vicinity of the target blade height section to form a smooth transition zone for geometric deviation, ensuring that the surface of the three-dimensional blade geometric model is smooth and continuous.

5. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 2, characterized in that, The process of performing aerodynamic simulation calculations on the three-dimensional blade simulation model, extracting aerodynamic indices of the geometric deviation parameter samples under the same operating conditions using a working condition control strategy, and constructing a mapping database between geometric deviations and aerodynamic indices specifically includes: At the design speed, the original design airfoil is scanned under all operating conditions to obtain the highest efficiency point, and the benchmark operating condition parameters corresponding to the highest efficiency point are extracted. For each set of geometric deviation parameter samples in the geometric deviation parameter sample set, CFD calculation is performed as the aerodynamic simulation calculation. While maintaining the same inlet boundary conditions as when the reference operating condition parameters were obtained, the outlet back pressure corresponding to each set of geometric deviation parameter samples is dynamically adjusted as the operating condition control strategy to force the actual aerodynamic parameters of each set of geometric deviation parameter samples to be equal to the reference operating condition parameters. After meeting the same working conditions, the aerodynamic index corresponding to each set of geometric deviation parameter samples is extracted to form an aerodynamic index output vector. The geometric deviation input parameter vectors corresponding to all the geometric deviation parameter samples are combined with the aerodynamic index output vectors to construct the geometric deviation and aerodynamic index mapping database.

6. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 5, characterized in that, The reference operating parameters include the reference value of the inlet relative Mach number and the reference value of the relative airflow angle; the aerodynamic parameters include isentropic efficiency and pressure ratio.

7. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 5, characterized in that, The step of training a surrogate model using the geometric deviation and aerodynamic index mapping database to establish a nonlinear mapping relationship between the geometric deviation parameters and aerodynamic indexes specifically includes: A fully connected neural network is selected as the surrogate model. The input layer of the fully connected neural network is set as the geometric deviation input parameter vector, and the output layer of the fully connected neural network is set as the aerodynamic index output vector. The geometric deviation and aerodynamic index mapping database is divided into a training set and a validation set, and the geometric deviation and aerodynamic index mapping database is used as a data source to input into the fully connected neural network for model training. During the model training process, the K-fold cross-validation method is used to optimize and validate the hyperparameters of the fully connected neural network, and the optimal network hyperparameters with the highest prediction accuracy on the validation set are selected to establish the data-driven nonlinear mapping relationship.

8. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 7, characterized in that, The process of using the surrogate model to infer aerodynamic performance from randomly generated geometric deviation parameter samples to complete uncertainty quantification specifically includes: Given a geometric deviation probability distribution, a randomly generated sample of geometric deviation parameters containing each of the geometric deviation parameters is generated using the Monte Carlo method. The randomly generated geometric deviation parameter samples are input into the fully connected neural network to obtain the aerodynamic performance prediction results corresponding to the randomly generated geometric deviation parameter samples; The aerodynamic performance prediction results of all the randomly generated geometric deviation parameter samples are statistically analyzed and the mean of the aerodynamic performance prediction results is calculated. The design point performance of the original airfoil is compared with the mean of the aerodynamic performance prediction results. The performance degradation between the original airfoil and the randomly generated geometric deviation parameter samples is calculated to complete the uncertainty quantification assessment.

9. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 8, characterized in that, The calculation of the contribution of each geometric deviation parameter in the randomly generated geometric deviation parameter sample using interpretable machine learning methods specifically includes: The SHAP interpretable machine learning method is selected as the interpretable machine learning method. The aerodynamic performance prediction results of the fully connected neural network are mathematically decomposed, and the marginal contribution of each geometric deviation parameter input into the fully connected neural network is calculated as the contribution. The SHAP interpretability machine learning method is used to decompose the aerodynamic performance prediction result corresponding to a single randomly generated geometric deviation parameter sample into the sum of the aerodynamic performance prediction result baseline value and the marginal contribution of each geometric deviation parameter to the aerodynamic performance prediction result.

10. The method for quantitatively evaluating the performance uncertainty of turbomachinery blade geometry deviation according to claim 9, characterized in that, The identification of geometric features highly sensitive to aerodynamic performance specifically includes: After calculating the marginal contribution of each of the geometric deviation parameters, the global weight of the marginal contribution of each of the geometric deviation parameters in all the randomly generated geometric deviation parameter samples is statistically evaluated, and a bar chart containing the marginal contribution values ​​of the features is output. The highly sensitive geometric features that cause performance fluctuations are identified as the geometric features that are highly sensitive to aerodynamic performance, and a sensitivity analysis report containing the geometric features that are highly sensitive to aerodynamic performance is generated.

Images