Aircraft engine rotating blade icing prediction method based on machine learning

Abstract

The invention discloses an aeroengine rotating blade icing prediction method based on machine learning, and relates to multiple technical fields of computational fluid mechanics, heat and mass transfer, multiphase flow simulation, artificial intelligence and the like. The method is used for rapidly and accurately predicting the icing form of the surface of the rotating blade of the aero-engine fan or the gas compressor under different incoming flow working conditions and blade profile parameter conditions. The method comprises the steps that firstly, a parameterized model of a rotating blade is constructed, multiple blade profile and working condition combinations are generated in combination with orthogonal experimental design, and a multi-source icing database is constructed through simulation and experiments; then, an icing prediction model is constructed based on a multi-layer perceptron and a deconvolution neural network, and after incoming flow parameters and blade profile parameters are input, corresponding ice shape images can be output; and finally, model optimization is realized through training and verification, and the recognition and fitting capability of a real ice shape is improved. According to the method, the icing prediction efficiency and precision can be effectively improved, and technical support is provided for anti-icing design and safe operation of the aero-engine.

Description

Technical Field

[0001] The present invention belongs to the technical field of aero-engine aerodynamic design and icing prediction, and relates to a method for predicting icing on compressor rotating blades. More specifically, it relates to a method for predicting icing on aero-engine rotating blades based on machine learning. The method is used to quickly and accurately predict the icing morphology and risk of aero-engine rotating blades under various flight conditions, thereby improving the safety and reliability of engine operation. Background Art

[0002] When an aircraft flies through clouds containing supercooled water droplets, the blades of the first few stages of an aircraft engine are very susceptible to ice formation. Once ice forms on aircraft engine blades, it poses a serious risk to safe operation, seriously threatening flight safety and causing numerous accidents.

[0003] First, when ice forms on the blades, their original aerodynamic shape will change, resulting in a decrease in aerodynamic performance and seriously affecting the compressor's boost capacity and operational stability. Secondly, because the engine is rotating at high speed, the iced parts are easy to fall off. The ice that falls off is very likely to be sucked into the compressor and combustion chamber, causing damage to the compressor blades and even causing serious consequences such as combustion chamber flameout and engine shutdown. Thirdly, ice on the intake components will reduce the airflow area, reducing the airflow entering the engine, resulting in reduced engine thrust, and in severe cases, it will cause compressor surge. Finally, if ice forms on the blade surface, it will cause unbalanced rotor rotation, causing violent vibration of the power unit, and in severe cases, it can cause damage to the engine rotor bearings.

[0004] Given the numerous hazards associated with icing on aircraft engine blades, accurate prediction of blade icing is particularly important. Knowing in advance whether ice will form on the blades, as well as the extent and location of the icing, can help crews take effective anti-icing and de-icing measures in advance, avoiding engine failures caused by icing and ensuring safe flight. At the same time, accurate icing predictions can also help optimize engine design, improve the engine's anti-icing system, enhance the engine's operational reliability and stability in icing environments, and reduce maintenance costs and flight delays caused by icing problems, which is of great significance to improving the safety and economic benefits of the aviation industry. Therefore, developing research on the prediction of icing on aircraft engine rotating blades and developing a fast and accurate prediction method are of great significance to aircraft engine safety.

[0005] Current research on aircraft icing falls into three main approaches: experimental studies, empirical estimation, and numerical simulation. Experimental studies investigate aircraft icing and its impact on flight performance through actual flight tests or by simulating icing conditions in ice wind tunnels. Empirical estimation relies on empirical charts and formulas derived from experiments. Numerical simulations analyze the icing process and its impact on aircraft performance by solving simplified mathematical models using computers.

[0006] However, all three methods mentioned above have certain defects. First, the experimental research method can visually observe the icing phenomenon and directly measure the impact of icing on flight performance. However, this method requires a large investment and a long research cycle, and the experimental objects have certain limitations. Second, the application scope of empirical estimation is relatively limited, and it cannot simulate the icing process. Therefore, this method is usually used for estimation in the preliminary stage of aircraft anti-icing / de-icing system design. Third, numerical simulation, as an economical and efficient alternative, can effectively simulate the icing process under different flight conditions by reasonably simplifying practical problems and constructing mathematical models. However, the numerical simulation method still has certain defects: First, the numerical simulation method is time-consuming. Although it shortens the cycle compared to the experimental method, it is still very time-consuming to solve the long icing process with millions of computational grids. In this regard, the computation time of high-speed rotating structures such as aircraft engines is more prominent than that of solving wing outflow. This is because the time characteristic scale of aircraft engines is very small (usually 10 -6 s), while the outflow from the wing is a stationary component, so the time characteristic scale is larger, and it takes less time to solve the icing problem of the same time. Secondly, the accuracy of the numerical simulation method itself is also flawed, especially for the case of clear ice (water droplets that are not completely frozen after impact). Since it cannot accurately describe the movement of the water film, the accuracy has a certain error compared to the experimental results. Finally, the simulation of icing in a three-dimensional rotating environment is more difficult. The aircraft engine is a rotating component with complex, strong three-dimensional and strong unsteady flow characteristics inside. In this case, how to consider the influence of rotation in the traditional two-dimensional stationary component icing method and extend it to the three-dimensional case is itself quite difficult, so the existing simulation models have more or less made some assumptions.

[0007] Furthermore, in the prior art, Chinese inventions CN116579383A and CN111680454A disclose icing prediction methods for wind turbine blades. However, these methods primarily predict the ice quality or presence of ice on wind turbine blades, failing to provide critical ice type distribution information. Furthermore, they fail to address the transient flow field-icing coupled modeling problem for high-speed rotating aeroengine blades, and fail to consider the influence of rotational effects such as centrifugal and Coriolis forces. Furthermore, the ice crystal icing prediction system for aeroengine stator blades disclosed in CN117871592A essentially relies on ground simulation tests to establish a correlation between temperature and ice thickness. This reliance on experimental methods fails to form a universal prediction model that can integrate multi-dimensional input parameters and rapidly output ice types.

[0008] In summary, experimental research methods are time-consuming and costly, making them unsuitable for routine prediction of ice types on aeroengine blades. Empirical estimates are inaccurate, have limited applicability, and are therefore unsuitable for routine icing prediction. Numerical simulation methods offer both timeliness and accuracy, and are widely used in routine icing research. However, the accuracy of numerical simulations still leaves much room for improvement, and the time consumption is particularly prominent in aeroengine simulations. Developing faster and more accurate icing prediction methods, capable of making appropriate corrections based on experimental results, would be extremely valuable in improving aeroengine icing prediction capabilities. Summary of the Invention

[0009] (1) Purpose of the invention In response to the defects and shortcomings of the existing technology in predicting icing on aircraft engine blades, such as long time consumption and low accuracy, the present invention aims to provide an aircraft engine rotating blade icing prediction method based on machine learning. By constructing a multi-source database of engine blade experiments and simulations covering a variety of blade parameters and incoming flow conditions, the experimental data and simulation data are integrated during the training process, and the accuracy of the simulation data is improved through the experimental data. Machine learning training is performed based on a model of a convolutional neural network, and finally a reasonable mapping of incoming flow conditions, structural parameters and icing profiles is formed. Through this machine learning model, the blade ice profile under a given incoming flow condition can be obtained more conveniently, quickly and accurately, effectively improving the icing prediction capability of aircraft engine rotating blades.

[0010] (2) Technical solution In order to achieve the purpose of the invention and solve the technical problems, the present invention adopts the following technical solutions: A machine learning-based method for predicting icing on rotating blades of an aircraft engine is used to quickly and accurately predict the icing morphology on the surface of rotating blades of an aircraft engine fan or compressor under different inflow conditions and blade structural parameters. The method comprises at least the following steps: S100. Blade Parametric Modeling and Orthogonal Design: Establish a parametric geometric model of an aero-engine rotating blade based on blade design variables and blade design constants. Perform a multi-level design of the blade design variables using the orthogonal experimental method, and combine them with the blade design constants to generate multiple sets of blade profiles covering the range of typical blade geometric characteristics. S200. Simulation case design and icing simulation database construction: For each blade profile, the orthogonal experimental method is used to perform multiple level selections of the incoming flow parameters within a preset range. Multiple sets of incoming flow parameter combinations covering the typical operating range of aircraft engine rotating blades are designed and generated. Each incoming flow parameter combination corresponds to an icing simulation case. Blade icing simulation calculations are performed for each blade profile and each incoming flow parameter combination to obtain the corresponding blade surface ice type data, and an icing simulation database is constructed based on this data. S300. Experimental data collection and multi-source icing database construction: Collect experimental data on icing on rotating blades of aircraft engines and perform unified data processing with blade icing simulation data. Construct a multi-source icing database that integrates experimental and simulation data and has a unified data structure and parameter dimensions. The multi-source database is then organized and divided into training and test sets, which are used for machine learning model construction and training and generalization capability evaluation, respectively. S400. Building a machine learning model for ice prediction: A machine learning model for icing prediction was established, consisting of an input layer, a hidden layer, and an output layer. The input layer receives a multidimensional feature vector containing blade profile parameters and incoming flow parameters. The hidden layer performs nonlinear mapping on the input features to extract abstract features and then converts the abstract features into the spatial dimensions of an ice shape image. The output layer generates an icing prediction image for the blade surface corresponding to the input parameters. S500. Icing Prediction Machine Learning Model Training and Validation: The ice prediction machine learning model was trained based on the constructed training set. A weighted mean square error loss function was used to guide the model to prioritize fitting the characteristics of real ice types. It was then iteratively trained using a deep learning framework. After training, the model's generalization ability was evaluated using a test set. S600. Aeroengine Rotating Blade Icing Prediction: The trained icing prediction model is used to predict icing on rotating blades of aircraft engines.

[0011] (3) Technical effects Compared with the prior art, the aircraft engine rotating blade icing prediction method based on machine learning of the present invention has the following beneficial and significant technical effects: (1) Traditional numerical simulation methods require long-term calculations for complex three-dimensional rotating flow fields (such as solving the icing process using millions of computational grids). However, the present invention uses a machine learning model to establish a direct mapping between the incoming flow conditions, structural parameters, and icing profiles. This allows predictions to be completed in seconds or minutes, which is more than 90% shorter than traditional numerical simulations and meets the needs of real-time warning and rapid assessment of aircraft engines.

[0012] (2) The model training data includes tens of thousands of sets of self-simulated data, 50 sets of experimental data from literature, and 50 sets of self-experimental data. Furthermore, these 100 sets of fixed experimental data are integrated into each set of training data, fully correcting the precision and accuracy of the simulation data. A weighted loss function (simulation error weight 0.4, experimental error weight 0.6) forces the model to prioritize fitting real experimental data. This is particularly true for complex ice formations such as clear ice (traditional numerical simulations suffer from errors due to insufficient description of water film motion), resulting in significantly improved prediction accuracy compared to a single simulation model.

[0013] (3) Icing prediction can be completed without relying on high-cost ice wind tunnel experiments or large-scale numerical simulations, reducing the hardware investment and time cost of aircraft engine icing research. It is particularly suitable for rapid iterative optimization in the early design stage. It can simultaneously process multiple variable inputs such as inflow temperature, liquid water content, and blade geometric parameters, covering different flight conditions (such as low-altitude cloud and fog environments and high-altitude supercooled water droplet areas) and engine structural design schemes, providing full-condition data support for anti-icing system optimization and blade aerodynamic design.

[0014] (4) This invention breaks through the bottleneck of traditional methods in efficiency and accuracy, and provides a new fast, accurate and economical way for aircraft engine icing prediction. It has significant engineering application value and economic benefits in improving flight safety, reducing maintenance costs and promoting engine design optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flow chart of the invented method for predicting icing on rotating blades of an aero-engine based on machine learning; Figure 2 Construct schematics for blade geometry; Figure 3 Schematic diagram of fluid domain mesh division; Figure 4 The velocity (a) and pressure (b) distribution cloud diagram of the blade flow field; Figure 5 is the droplet water content distribution cloud map; Figure 6 This is a schematic diagram of the method for selecting the icing control body; Figure 7 This is a schematic diagram of the internal structure of the ice control body; Figure 8Schematic diagram of ice profile distribution on the leading edge of the blade; Figure 9 Schematic diagram of neural network architecture based on MLP and DNN; Figure 10 Training loss monitoring graph for the model; Figure 11 Schematic diagram of ice shape feature prediction under different working conditions, where (a)~(d) correspond to working conditions 1 to 4 respectively. DETAILED DESCRIPTION

[0016] The present invention aims to provide a method for predicting icing on rotating blades of aircraft engines based on machine learning, which is used to quickly and accurately predict the icing morphology on the surfaces of rotating blades of aircraft engine fans or compressors under different inflow conditions and blade parameter conditions. To make the purpose, technical solutions, and advantages of the implementation of the present invention more clear, the technical solutions in the embodiments of the present invention will be described in more detail below in conjunction with the accompanying drawings in the embodiments of the present invention. Based on the content of the technical solution of the invention, combined with the decomposition of the steps in the technical solution and the description of key parameters, the implementation process is described in detail: As a specific example, Figure 1 As shown, the method for predicting icing of rotating blades of an aircraft engine based on machine learning of the present invention mainly includes the following steps when implemented: S100. Blade Parametric Modeling and Orthogonal Design: A parametric geometric model of aero-engine rotating blades is established based on blade design variables and blade design constants. Multi-level design of blade design variables is performed based on the orthogonal experimental method, and after combining with blade design constants, multiple sets of blade profile schemes covering the range of variation of typical blade geometric characteristics are generated.

[0017] In the embodiment of the present invention, the blade parametric modeling in step S100 includes at least the following sub-steps: S111. Determine the design variables and define the blade design variables as the inlet geometry angle , outlet geometry angle , installation angle , consistency σ , relative maximum curvature position x mc , maximum turning angle h mc , relative maximum thickness position x mth , maximum thickness r mth The above parameters are variables in the blade design process. By adjusting the values ​​of these parameters, different types of blade profiles can be generated. Each design variable is used to control the blade geometry and aerodynamic characteristics, and its value range should cover the typical blade geometry characteristics in the target application scenario.

[0018] S112. Determine several blade design constants, including blade chord length c , mid-arc inlet angle α 1. Mid-arc exit angle α 2. Leading edge radius r 1. Trailing edge radius r 2. The values ​​of each design constant are set based on the blade type and design requirements. These constants remain constant throughout the design process, maintaining the same value for different blade profiles. They constrain the basic geometric dimensions and boundary shape of the blade and maintain a stable structural benchmark in the parametric representation. Combining these blade design variables and constants allows for the construction of a unique blade geometry, enabling the design of blades with varying geometries.

[0019] S113. Based on the specified blade design variables and constants, construct expression equations for the blade profile's median camber and thickness distribution using Bezier curves, B-spline curves, or piecewise polynomial combination curves. The median camber expression equation should ensure continuous smoothness along the entire chord length, and the thickness distribution expression equation should ensure continuous differentiability with respect to the suction and pressure surface shapes. Specifically: In the embodiment of the present invention, a cubic polynomial combination curve is used to construct the blade profile camber line, and the expression is: in, x and y are the spatial coordinates of the points on the mid-arc, a 10 ~ a 23 is the coefficient, and its specific value is: In the above coefficient expressions, x max and y max are the horizontal and vertical coordinates of the maximum curvature point, respectively, and their values ​​are: In this embodiment, a cubic polynomial combination curve is also used to construct the blade thickness distribution, and its expression is: in x is the horizontal coordinate of the point on the mid-arc, h is the thickness of the blade, a 30 ~ a 43 is the coefficient, specifically:

[0020] In the above coefficient expressions, y 0 is the thickness value reference at the maximum thickness point, S is the length of the mid-arc.

[0021] S114. The blade geometry distribution is obtained by superimposing the blade mid-camber coordinates and thickness distribution, and the coordinate points of the leading edge, trailing edge, and pressure / suction surface of the blade are generated to form the parameterized blade geometry, such as Figure 2 shown.

[0022] In the embodiment of the present invention, the batch design of blade profiles based on the orthogonal design method in step S100 includes at least the following sub-steps: S121. Determine the horizontal range of design variables: Based on the set blade design variables and incorporating the actual design requirements and application conditions of aircraft engine rotating blades, define the range of values ​​for each design variable within the target blade profile design space to ensure that the design space covers the typical range of blade geometric characteristics. S122. Design Variable Level Division and Discretization: Based on the value range of the blade design variable and its sensitivity to blade geometry and aerodynamic performance, the number of levels of the design variable is determined and the corresponding value range is divided into multiple discrete levels with equal or unequal spacing. S123. Construct an orthogonal experimental design matrix: Select an appropriate orthogonal array based on the number of design variables and levels. Systematically combine the discrete levels of each design variable according to the arrangement rules of the orthogonal array to obtain an orthogonal design matrix that maximizes factor combination coverage under finite sample conditions. S124. Generate multiple blade profile designs: Combine each row of the orthogonal matrix with the preset blade design constants. Generate corresponding blade geometry using the constructed blade parameterization model. The resulting blade profiles should have clear boundary contours, thickness distribution, and median camber continuity. S125. Blade profile geometric verification and optimization: Each generated blade profile is geometrically verified for feasibility, abnormal blade profiles are eliminated, and blade profiles with minor geometric defects are optimized, ultimately forming a standardized blade profile library.

[0023] More specifically, in this embodiment, when performing the orthogonal experimental design of blade profile, L1000 (10 8 ) Orthogonal table for 8 key variables (including inlet geometry angle , outlet geometry angle , installation angle , consistency σ , relative maximum curvature position x mc , maximum turning angle h mc, relative maximum thickness position x mth , maximum thickness r mth ) for a ten-level design, and the other five variables (including blade chord length c , mid-arc inlet angle α 1. Mid-arc exit angle α 2. Leading edge radius r 1. Trailing edge radius r 2) Fixed to typical values, 1000 sets of blade profiles were generated to cover the range of common geometric parameters of engine compressors.

[0024] S200. Icing simulation case design, calculation, and icing simulation database construction: For each blade profile scheme, the inflow operating condition parameters are taken at multiple levels within a preset range based on the orthogonal experimental method, and multiple groups of inflow operating condition parameter combinations covering the typical operating condition variation range of aircraft engine rotating blades are designed and generated. Each inflow operating condition parameter combination corresponds to an icing simulation case; blade icing simulation calculations are performed for each blade profile scheme and each of its inflow operating condition parameter combinations to obtain the corresponding blade surface ice type data, and an icing simulation database is constructed based on this.

[0025] In an embodiment of the present invention, when performing batch design of icing simulation examples for each blade profile scheme based on the orthogonal design method in step S200, at least the following sub-steps are included: S211. Determine inflow parameters and their design ranges: Based on the operating environment of aircraft engine rotor blades, determine several inflow parameters that significantly affect icing characteristics. Set the value ranges for each parameter based on icing history and / or test data. S212. Level Design and Discretization of Inflow Parameters: Based on the range of values ​​for each inflow parameter and its sensitivity to icing characteristics, each parameter is discretized at multiple levels and its level values ​​are determined using either equidistant or non-equidistant partitioning. S213. Construct an orthogonal design matrix for inflow parameters: Using the orthogonal experimental design method, select an appropriate orthogonal table based on the design variables and their number of levels to generate an orthogonal design matrix for statistically representative and spatially uniform combinations of operating parameters. S214. Generate icing simulation case designs: Combine each row of operating condition parameter combinations in the orthogonal design matrix with the blade profile design to form a complete icing simulation case design. Each case includes a clear blade geometry definition and inflow boundary condition settings.

[0026] More specifically, the orthogonal experimental design of the working condition in this embodiment adopts L50(5 4 ) Orthogonal table for 4 working parameters (including flow velocity vA five-level design was performed using the following parameters: inlet temperature T, liquid water content LWC, and droplet size MVD. This generated 50 operating scenarios, covering the common operating range of engine compressors. Combined with 1,000 blade geometry sets, this resulted in a total of 1,000 x 50 = 50,000 icing simulation cases.

[0027] Furthermore, in an embodiment of the present invention, performing simulation calculations based on the icing simulation example and constructing an icing simulation database in step S200 include at least the following sub-steps: S221. Fluid Domain Modeling and Meshing: For the selected icing simulation example, a fluid domain model containing the rotating blade channel of the aircraft engine is constructed based on its corresponding blade profile scheme, and the fluid domain is divided using structured or unstructured grid technology. Structured grids are preferably used to divide the fluid domain. Structured grids can not only improve computational efficiency, but also ensure the accuracy of flow field simulation. In particular, in the critical area of ​​the blade wall, local mesh encryption is performed to capture complex flow phenomena and subtle pressure changes near the wall. In terms of controlling the number of grids, the total number of grids is strictly controlled between 100,000 and 200,000, which can not only ensure the rational use of computing resources, but also meet the general accuracy requirements of compressor blade flow field simulation. One of the typical blade profiles is selected as an example, and a corresponding grid division structure diagram is drawn, as shown below. Figure 3 shown.

[0028] S222. Setting the boundary conditions of the calculation model: According to the combination of incoming flow parameters corresponding to the icing simulation example, the boundary conditions of the calculation model are set accordingly, including at least setting the incoming flow velocity and incoming flow temperature at the inlet of the fluid domain, setting the pressure boundary condition at the outlet, setting the no-slip wall boundary condition on the blade surface, and setting periodic boundary conditions according to the actual working state of the rotating blade to simulate the flow between blade rows.

[0029] S223. Air flow field calculation: After completing meshing and boundary settings, the Reynolds-averaged Navier-Stokes (RANS) equations are used to calculate the air flow in the rotating blade channel. The air flow field distribution, including velocity, pressure, and temperature fields, is obtained, ensuring reasonable control of computational time and resources while meeting computational requirements. An appropriate turbulence model is selected for turbulent characteristics during flow field calculations, with the Spalart-Allmaras model being the preferred choice. This SA model utilizes the Boussinesq assumption and establishes a relationship between the eddy viscosity coefficient and the corresponding turbulence model transport variables. The SA model is also well-suited for both structured and unstructured grids. Figure 4 The flow field distribution around one of the blades is shown, taking velocity and pressure parameters as an example.

[0030] S224. Droplet Motion and Impact Calculation: Based on the results of the air flow field solution, the Euler method is used to solve the water droplet motion equation under the action of air. Since the content of droplets in the air is very small, a one-way coupling method is generally used to solve the air flow field first, then solve the droplet motion under the action of the air flow field, and calculate the impact characteristics of the droplet with the blade surface, including the impact position, impact velocity, impact angle and impact momentum distribution, to provide the droplet impact boundary conditions for subsequent icing calculations. Specifically, the droplet motion and impact calculations can be further subdivided into the following sub-steps: S2241. Establishing the droplet motion control equation: The droplet motion control equation is established in a rotating coordinate system, taking into account the aerodynamic drag, gravity, centrifugal force, and Coriolis force acting on the droplet in a rotating environment. The drag force is calculated based on the relative velocity difference between the droplet and the airflow, while the centrifugal force and Coriolis force are determined based on the engine speed and the droplet motion state. This invention expands the droplet motion equation in a rotating coordinate system to account for the effects of centrifugal and Coriolis forces. The specific control equation is as follows: 1) Continuity equation: 2) Momentum equation: Where, is the water droplet density, is the relative velocity of the water droplet, is the water droplet volume fraction, g is the acceleration due to gravity, t For time, ω is the angular velocity of rotation. F is the drag force of the air on the water droplet and ,in For relaxation time, d p is the average radius of the water droplet. f is the Reynolds number Re r The function here takes the following form: Where, v ar is the relative velocity of air, μ is the air viscosity coefficient.

[0031] S2242. Droplet Phase Initialization and Boundary Injection Configuration: Based on the air flow field solution, a one-way coupling strategy is employed to set the droplet release location and initial velocity conditions at the fluid domain inlet, based on the liquid water content and droplet size distribution in the incoming flow parameters, without affecting the airflow state. S2243. Numerical Solution of Droplet Trajectory: The Euler method is used to numerically integrate and solve the equations governing droplet motion. At each time step, the droplet's velocity, position, and force state are updated based on the flow field information at the current position until the droplet leaves the computational domain or impacts the blade surface. S2244. Droplet Impact Judgment and Parameter Calculation: Establish criteria for the impact of a droplet on a blade surface. Determine the occurrence of an impact event when the droplet trajectory geometrically intersects the blade surface. Extract key impact parameters for each impact point, including at least the impact position, velocity vector, impact angle, and impact momentum, ultimately forming an impact boundary condition field. Figure 5 The droplet moisture content cloud map is used as an example for demonstration.

[0032] S225. Calculation of the Freezing Thermodynamic Process: After the droplet motion calculation is completed, the droplet distribution on the blade surface is extracted, and the blade surface temperature distribution is also extracted to calculate the thermodynamic process of icing on the blade surface. During the calculation process, the water film motion is considered to enhance the simulation of different types of ice (frost ice, clear ice, and mixed ice), and the heat flux density distribution and phase change characteristics at each location on the blade surface are obtained. Specifically, the calculation of the icing thermodynamic process can be broken down into the following sub-steps: S2251. Construction of icing control volume: The first layer of grid near the wall of the air flow field solution grid is used as the grid for icing numerical simulation, and the icing control volume is established based on it, such as Figure 6 The icing control body is divided into three layers, from the outside to the inside, namely the air layer, the water film layer, and the ice layer, as shown in Figure 7 By constructing and solving the governing equations, the water film flow and ice growth in the control volume are described.

[0033] S2252. Establishment of the water film flow control equations: The continuity equation, momentum equation and energy equation of the water film flow are established in the control body. The continuity equation takes into account the mass source term of the droplet impact and the mass loss caused by phase change. The momentum equation comprehensively considers the influence of air shear force, gravity, wall pressure gradient, centrifugal force and Coriolis force on the water film. The energy equation considers the heat exchange process between the water film and the air layer and ice layer, as well as the energy input caused by the droplet impact. Specifically, the present invention takes into account the flow of unfrozen water film in the ice thermodynamic model, thereby enhancing the simulation ability of clear ice and mixed ice. In addition, the ice thermodynamic process in the rotating coordinate system is expanded, taking into account the influence of centrifugal force and Coriolis force on the surface water film. The specific control equations are as follows: 1) Continuity equation of water film flow: In the formulav r is the relative flow velocity of the water film, u and v For this speed x , y The two components of the direction (two perpendicular directions on the ice surface), H w For water film z Height in the direction (normal to the ice surface), H i is the ice thickness, m imp is the droplet impact rate, ρ i is the density of ice, ρ w is the density of water, t For time.

[0034] 2) Momentum equation for water film flow: Water film flow is primarily influenced by air shear and pressure, and in a rotating environment, centrifugal and Coriolis forces must also be considered. Because the velocity of the water film is very low, it can be considered a steady-state flow. Water film flow also satisfies the incompressible NS equations, which can be expressed as follows: The right side of the formula represents the gravity, viscous stress and air pressure on the water film. p is the local gas pressure.

[0035] 3) Energy equation for water film flow Since the thickness of the water film is very small and the flow velocity is also very small, the heat transfer process in the water film can be regarded as a steady-state process in each freezing time step, satisfying the energy equation: In the formula u is the water film flow velocity, a w is the thermal diffusivity of water, T w is the water film temperature. The temperature at the water-ice interface is constant at the freezing phase transition temperature. Convective heat transfer and energy exchange due to water droplet impact occur at the water-air interface.

[0036] S2253. Calculation of Heat Transfer and Energy Conservation in Ice Layers: Using the same method, the energy equation in the ice layer can be obtained as follows: , T iThe one-dimensional steady-state heat transfer governing equations assumes an adiabatic boundary at the ice-wall interface, a constant freezing phase transition temperature at the water-ice interface, and energy exchange due to convective heat transfer and water droplet impacts at the water-air interface. The heat transfer path within the ice layer and its heat exchange with the wall and water film are simulated, and the temperature field within the ice layer is solved to provide boundary conditions for the phase transition calculation at the ice-water interface.

[0037] S2254. Phase Transition Behavior and Ice Type Identification Mechanism: Criteria for determining frost ice, clear ice, and mixed ice are established based on the water film flow state, temperature field distribution, and impact flux characteristics. Clear ice forms when the water film can flow on the surface and partially freezes. Frost ice forms when the impacting droplet freezes instantaneously without water film flow. Mixed ice forms when both mechanisms coexist. Based on the phase transition mechanisms of different ice types, thermophysical properties and phase interface boundary conditions are adjusted to achieve adaptive ice type simulation. S2255. Calculation of Heat Flux Density Distribution and Phase Change Characteristics: By solving the coupled thermodynamic equations governing the water film and ice layer within the blade, the heat flux density distribution and phase change latent heat flux at each location on the blade surface are calculated, providing thermodynamic boundary conditions for subsequent ice growth calculations.

[0038] S226. Ice Growth Thickness Calculation: Based on the calculation results of the freezing thermodynamic process, the ice growth rate is calculated by analyzing the difference in heat transfer capacity of the latent heat of freezing at the ice-water interface between the water film and the ice layer on both sides of the interface. The ice thickness distribution under the clear ice freezing mode and the frost ice freezing mode is solved respectively. The smaller value of the two is selected as the final ice thickness to ensure the accuracy of the ice type judgment. Specifically, the ice growth thickness calculation can be broken down into the following sub-steps: S2261. Establishment of the ice layer growth rate control equation: Within the ice control volume, an energy conservation model is established around the ice-water interface. By analyzing the release and heat transfer process of the latent heat of freezing at the ice-water interface, and considering the temperature gradient difference in the water film and ice layer on both sides of the interface, the ice layer growth rate control equation for the change of ice layer thickness over time is established. Specifically, the latent heat released by ice formation at the ice-water interface is first introduced into the water film and ice layer on both sides of the interface, and then further transferred to the airflow and substrate, respectively. According to the Stefan condition, the ice layer growth rate in the control volume mainly depends on the heat transfer capacity in the water film and ice layer. The growth rate of the ice layer thickness can be expressed as the following expression: In the formula L f is the latent heat of freezing phase change, L f =334400 J / kg, λ i is the thermal conductivity of the ice layer, λ w is the thermal conductivity of water.

[0039] S2262. Comparison of dual-mode results and minimum value selection: Calculate the ice thickness distribution under the clear ice and frost ice icing modes respectively. By comparing the numerical values ​​of the two calculation results at each surface location, select the smaller value of the two as the final ice thickness at that location to ensure the accuracy and rationality of the ice type judgment. That is, after calculating the ice thickness under clear ice icing according to the above equation, it is necessary to compare it with the ice thickness under frost ice icing. If it exceeds the frost ice icing thickness, the final ice thickness is set to the frost ice thickness, and its value is .

[0040] S2263. Final determination of ice type and thickness distribution output: The final ice type at each location on the blade surface is determined based on the minimum value selection principle. When the clear ice thickness is less than the frost ice thickness, it is determined as clear ice; otherwise, it is determined as frost ice. When the clear ice thickness and frost ice thickness are close, it is determined as mixed ice. Finally, the complete ice thickness distribution data and ice type distribution are output. Figure 8 Taking a certain blade shape and operating condition as an example, the icing pattern is demonstrated. Icing mainly occurs on the leading edge of the blade.

[0041] S227. Ice Surface Update and Mesh Adaptation: Calculate the new ice layer profile based on the ice layer growth thickness, adjust the blade geometry based on the original blade geometry, and adaptively update the computational mesh based on the updated blade geometry in combination with dynamic mesh technology, ensuring that the mesh quality meets the accuracy requirements of subsequent time step simulation calculations; S228. Multi-time-step ice evolution simulation: Based on the updated blade geometry and mesh, steps S223 to S227 are repeatedly executed while keeping the incoming flow parameters unchanged, to achieve multi-time-step simulation of the dynamic evolution of icing until the preset icing time or ice thickness convergence conditions are reached, ultimately obtaining complete icing evolution history data; S229. Batch Calculation and Database Construction: After completing the full-process simulation of a simulation example, all preset icing simulation examples are traversed, and steps S221 to S228 are repeated to perform simulation calculations. Key icing characteristic parameters are extracted from each simulation example, and finally the construction of an icing simulation database covering multiple blade schemes and operating conditions is completed.

[0042] S300. Construction of an icing database based on multiple sources of experiments and simulations Experimental data on icing on rotating aircraft engine blades is collected and processed uniformly with blade icing simulation data. A multi-source icing database is constructed that integrates the experimental and simulation data and has a unified data structure and parameter dimensions. The multi-source database is then divided into training and test sets according to a set ratio, respectively, for use in building and training machine learning models and evaluating their generalization capabilities. Specifically, in this embodiment, the construction of the multi-source icing database and the division of the training and test sets include the following sub-steps: S301. Simulation data collation and experimental data collection: Using the icing simulation method in step S200, icing simulations were performed for 1,000 blade profiles under 50 operating conditions, totaling 50,000 simulation examples. Parameters for both blade profiles and operating conditions were selected using orthogonal experimental methods. This ultimately created an icing simulation database. In subsequent steps, the simulation data was divided into an 8:2 ratio, with 80% of the simulation data being a training set (40,000 sets) and 20% of the simulation data being a test set (10,000 sets).

[0043] Based on the above icing simulation database, we further collected blade icing experimental data, including at least several sets of typical icing experimental data collected from public literature and several sets of data measured by rotating blade icing experiments conducted through an independently built ice wind tunnel platform, including: Experimental Data from the Literature: We systematically collected 50 sets of publicly available experimental data on engine blade icing. These data were sourced from multiple authoritative institutions, including but not limited to wind tunnel test results conducted by renowned organizations such as NASA's Lewis Research Center and Pratt & Whitney. This data comprehensively covers ice type measurements under various incoming airflow conditions, specifically encompassing icing scenarios under various environmental parameters such as airflow velocity, temperature, and humidity. This data provides detailed and reliable experimental data support for training high-precision neural networks.

[0044] Independent experiments: In a small ice wind tunnel environment, independent experimental research on blade icing patterns was conducted. Using high-performance high-speed cameras (frame rates up to 1000fps) and high-precision laser displacement sensors (measurement accuracy of ±0.1mm), the evolution of ice patterns during the blade icing process was measured and recorded. Through a series of experiments, 50 sets of detailed experimental data were obtained, covering the ice thickness at the leading edge of the blade, the specific range of icing, etc. It should be noted that the compressor blade design used in the experiment fully covers the range of blade design parameters that need to be considered in the icing simulation process, ensuring the wide applicability of the experimental results. At the same time, the incoming flow conditions simulated in the experiment also fully cover the various operating parameters involved in icing simulation, thereby providing solid data support for subsequent machine learning training.

[0045] S302. Multi-source icing database construction: Data cleaning, format conversion, and normalization are performed on blade icing simulation data, literature experimental data, and independent experimental data. This data is then integrated with the experimental and simulation data to form a multi-source icing database with a unified data structure and parameter dimensions. Specifically, step S302, during data normalization, includes the following steps: Data cleaning: Through careful review and screening, incomplete or abnormal data records are eliminated, thereby eliminating noisy data that may affect the model training effect.

[0046] Format conversion: This ensures seamless integration of data from diverse sources. Given the diversity of data sources and formats, data is unified into a standard format to eliminate issues that may arise from inconsistent data formats.

[0047] Normalization: This primarily addresses differences in the dimensions and numerical ranges of operating parameters (such as temperature, velocity, and liquid water content), blade profile parameters (such as maximum camber, maximum thickness, mounting angle, inlet / outlet geometry), and ice image pixel values. To prevent large numerical differences from affecting convergence during model training, a maximum-minimum normalization method is employed. This method scales all input parameters to a range of 0 to 1, ensuring more stable model training. Inflow and blade profile parameters are scaled to [1, 1] using normalization (ZScore), while ice image pixel values ​​are normalized to [0, 1] using MinMax. The data format is then standardized (CSV file).

[0048] S303. Dataset Segmentation (Training and Test Sets): Based on the principle of multidimensional uniformity of icing type, blade profile, and operating parameters, a stratified sampling strategy was used to divide the simulation dataset into an 80% training set (40,000 data sets) and a 20% test set (10,000 data sets).

[0049] Specifically, the partitioning process not only considers the distribution of different blade structures (such as geometric angles, maximum thickness position, installation angle, etc.) and incoming flow parameters (such as velocity, temperature, droplet size, etc.) in the sample, but also takes into account the coverage ratio of different ice types (such as frost ice, clear ice, and mixed ice) in the sample, thereby ensuring that each subset is representative in terms of sample type and statistical characteristics. At the same time, some data can be reserved as a validation set for model parameter adjustment and performance optimization. This validation set is also constructed according to the principle of stratified sampling to ensure the comparability and systematicity of the data closed loop of the entire training-validation-testing process.

[0050] S304. Training Group Split: The training set is split into multiple training groups according to a preset group size, with each training group containing an equal amount of simulation data. Then, experimental data from the literature and self-conducted experimental data are added to each training group in equal proportions, ensuring that each training group is composed of both simulation and experimental data. Specifically, the training set is split into 400 training groups, each containing 100 simulation data sets. Each training group is then supplemented with 50 sets of experimental data from the literature and 50 sets of self-conducted experimental data. That is, each training group contains 100 sets of simulation data and 100 sets of experimental data (a mix of literature data and self-conducted experimental data). This split ratio helps ensure that the model learns both simulation patterns and real-world experimental characteristics during training, and its generalization ability is verified on the test set.

[0051] It should be noted that this step breaks through many bottlenecks in the construction of the data support system for the prediction modeling of icing on rotating blades of aircraft engines, such as the traditional separation of simulation and experimental data, inconsistent data dimensions, and insufficient physical authenticity of training samples. It has significant technical innovation, which is specifically reflected in the following aspects: (1) Fusion of multi-source icing data to build a unified database. By integrating the three-source data of simulation, literature experiment, and independent experiment, the present invention greatly improves the diversity and breadth of samples while ensuring physical accuracy, and significantly improves the adaptability of the icing prediction model to complex flight conditions. (2) The design of the simulation-experimental proportional mixing strategy within the training group is reasonable. This strategy not only improves the model's ability to fit the characteristics of real ice types, but also effectively avoids the adverse effects of the scarcity of experimental data on the training process. It is a key link in ensuring the generalization ability of the model and the credibility of predictions. (3) The data set division method based on stratified sampling and multidimensional uniformity principle. This strategy ensures statistical balance between the training set and the test set in terms of key variable dimensions such as blade profiles, operating parameters, and icing types, improves the representativeness of the training data, and enhances the model's ability to predict boundary conditions and complex ice types.

[0052] S400. Building a machine learning model for ice prediction: An icing prediction machine learning model consisting of an input layer, a hidden layer, and an output layer is established. The input layer is used to receive a multidimensional feature vector including blade profile parameters and incoming flow condition parameters. The hidden layer is used to perform nonlinear mapping on the input features to extract abstract features and restore the abstract features to the spatial dimensions of the ice shape image. The output layer is used to generate an icing prediction image of the blade surface corresponding to the input parameters.

[0053] In the embodiment of the present invention, the neural network of the machine learning model adopts an architecture that combines a multi-layer perceptron (MLP) and a deconvolutional neural network (DNN), which includes three parts: an input layer, a hidden layer, and an output layer. Figure 9 As shown. Among them: Input layer: It is designed to receive the incoming flow conditions and blade parameters. The input data is transmitted to the network in the form of a multi-dimensional numerical vector, which serves as the basis for model perception and representation. Specifically, in this embodiment, the input layer receives a 15-dimensional parameter vector, including 8 blade parameters (inlet geometry angle, , outlet geometry angle , installation angle , consistency σ , relative maximum curvature position x mc , maximum turning angle h mc , relative maximum thickness position x mth , maximum thickness r mth ), and 5 incoming flow parameters (incoming flow velocity v , incoming flow temperature T, liquid water content LWC, droplet size MVD).

[0054] The hidden layers are designed to consist of three fully connected layers and three deconvolutional layers. The fully connected layers form the main structure of the MLP network: the three fully connected layers contain 256, 128, and 64 neurons, respectively, and all use the Reluctant Unit (ReLU) activation function, which abstracts and transforms the input features layer by layer, achieving nonlinear mapping of extracted parameter features. To improve the model's generalization performance, batch normalization and dropout techniques are added after each fully connected layer. Batch normalization adjusts the distribution of each mini-batch of data to have a mean of 0 and a variance of 1, accelerating training and enhancing model stability. Dropout randomly excludes the outputs of some neurons during training to prevent overfitting. The deconvolutional layers are DNN modules integrated into the main MLP network structure: three deconvolutional layers (with kernel sizes of 4×4, 4×4, and 4×4, with strides of 2, 2, and 1, respectively). Each layer is followed by a Reluctant Unit (ReLU) activation function to restore the abstract features to the spatial dimensions of the ice image.

[0055] The output layer is designed to generate a predicted ice shape image based on the input flow conditions and blade parameters through convolution operations. This image contains at least the spatial distribution of ice thickness on the blade surface. In this embodiment, the output layer generates a 128×128 pixel single-channel grayscale image using a 1×1 convolutional layer. The pixel values ​​correspond to the ice thickness at each point on the blade surface (normalized to [0, 1]).

[0056] S500. Icing Prediction Machine Learning Model Training and Validation:

[0057] Based on the constructed training set, the ice prediction machine learning model is trained. The weighted mean square error loss function is used to guide the model to prioritize fitting the real ice type characteristics. It is then iterated in conjunction with the deep learning framework. After training, the generalization ability of the model is evaluated using the test set. Specifically: In this embodiment, the model is built and trained based on the PyTorch deep learning framework. The Adam optimizer is used, and the initial learning rate is set to 1×10 -4 , decayed by 0.5 times every 50 rounds. And the weighted mean square error WMSE=α×MSE is used sim +β×MSE exp As the loss function, where α is the simulation error MSE sim The weight of β is the experimental error MSE exp The weight of β>α is used to give higher weight to the experimental error, forcing the model to fit the real experimental data first to correct the accuracy deviation of the simulation data. sim The weight is set to 0.4, MSE exp The weight of is set to 0.6 to force the model to fit the real experimental data first.

[0058] Furthermore, a backpropagation algorithm and gradient descent optimizer were used to train the neural network model. Model parameters were iteratively updated to minimize the loss function between the predicted results and the actual ice morphology. To accelerate training and improve model performance, strategies such as learning rate decay, dropout, and early stopping were also employed.

[0059] The model training process is GPU-accelerated and includes the following steps: 1) Data loading: The preprocessed training data is loaded into memory in batches for model training. Each batch contains a combination of 100 sets of simulated data and 100 sets of experimental data. 2) Forward propagation: The input data is fed into the model. After feature extraction by the fully connected layer, the model generates ice shape image predictions through the deconvolution layer. 3) Loss calculation: Based on the difference between the predicted ice shape image and the true ice shape image, the weighted mean squared error (MSE) between the predicted image and the true label (simulated / experimental ice shape data) is calculated. 4) Backward propagation: Based on the loss value, gradient descent is used to update the model weights and bias parameters to minimize the loss value. 5) Model saving: After each iteration, the optimal model weights are saved for subsequent evaluation and use. During training, visualization tools such as TensorBoard are used to monitor the training process, including the changing trend of the loss value and the update of the model weights, so that the training strategy and hyperparameters can be adjusted in a timely manner.

[0060] Regarding model training monitoring and tuning. Real-time monitoring: Tracking metrics such as loss trends, neuron activation distribution, and weight update amplitude across the training and validation sets to determine whether the model is overfitting or underfitting. Hyperparameter adjustment: If the validation set loss does not decrease within 20 epochs, trigger the early stopping mechanism; adjust parameters such as the number of neurons in the fully connected layer, the step size of the deconvolution layer, and the learning rate decay strategy until the model converges.

[0061] Data augmentation: Random noise (to simulate measurement errors) and rotation and translation transformations (to expand the diversity of ice shapes) are added to the simulation data to improve the generalization ability of the model. Figure 10 Shows a training loss monitoring plot.

[0062] Regarding the model training cycle, the total number of iterations was set to 1000 rounds, and a single training session took approximately 12 hours (based on an NVIDIA RTX3090 GPU). Ultimately, prediction accuracy on the test set was achieved with an MSE ≤ 0.01 for simulation data and an MSE ≤ 0.02 for experimental data.

[0063] After training, a machine learning model for predicting icing on rotating aircraft engine blades was obtained. Based on this model, icing patterns can be predicted for any combination of blade profile parameters and operating parameters. Figure 11 The validation results of ice shapes under different operating conditions are presented, with the red curve representing actual data and the green curve representing predicted data. By comparing the predicted ice shapes with the actual ice shapes, we can observe that the model's predictive ability is relatively accurate, thus verifying its excellent generalization ability and practical application value in predicting icing.

[0064] S600. Aeroengine Rotating Blade Icing Prediction and Analysis: Using the trained icing prediction model, icing prediction is performed on rotating blades of an aircraft engine. In an embodiment of the present invention, icing prediction of blades based on blade profile and operating parameters includes the following sub-steps: Input parameter preprocessing. The blade parameters (inlet geometry angle , outlet geometry angle , installation angle , consistency σ , relative maximum curvature position x mc , maximum turning angle h mc , relative maximum thickness position x mth , maximum thickness r mth ) and the incoming flow parameters (incoming flow velocity v , incoming flow temperature T, liquid water content LWC, droplet size MVD) are normalized and converted into a high-dimensional input vector.

[0065] Ice shape image generation: The input vector is fed into the trained ice prediction model, and the deconvolution layer outputs a 128×128 ice shape image, where pixel values ​​are mapped to ice thickness.

[0066] Post-processing and visualization. Feature extraction: Extracts ice thickness distribution (leading edge percentage, pressure / suction side distribution) at different icing locations, calculating metrics such as maximum thickness and coverage area. Risk assessment: Outputs a risk level based on ice thickness and location, combined with engine operating safety standards. Visualization: Generates a 2D ice profile or 3D ice model, overlaying blade geometry to visualize icing patterns.

[0067] The above embodiments fully and effectively achieve the objectives of the present invention. Those skilled in the art will appreciate that the present invention includes, but is not limited to, the contents described in the accompanying drawings and the above specific embodiments. Although the present invention has been described with reference to the embodiments currently considered to be the most practical and preferred, it should be understood that the present invention is not limited to the disclosed embodiments, and any modifications that do not deviate from the functional and structural principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A method for predicting icing on rotating blades of an aircraft engine based on machine learning, characterized in that: At least the following steps are included: S100. Establish a parametric geometric model of an aeroengine rotating blade based on blade design variables and blade design constants; perform multi-level design of the blade design variables using an orthogonal experimental method, and generate multiple sets of blade profiles by combining them with the blade design constants; S200. For each blade profile, perform multiple level selections on the incoming flow parameters using an orthogonal experimental method. Multiple combinations of incoming flow parameters are designed and generated, with each corresponding combination corresponding to an icing simulation case. Blade icing simulation calculations are performed for each case to obtain corresponding blade surface ice pattern data, and an icing simulation database is constructed based on this data. S300. Collect experimental data on aircraft engine rotating blade icing and perform unified data processing with blade icing simulation data to construct a multi-source icing database that integrates experimental and simulation data. The multi-source database is then organized and divided into training and test sets. S400. Establish an icing prediction machine learning model comprising an input layer, a hidden layer, and an output layer. The input layer receives a multidimensional feature vector including blade profile parameters and incoming flow parameters. The hidden layer performs nonlinear mapping on the input features to extract abstract features and converts these abstract features into the spatial dimensions of an ice shape image. The output layer generates an icing prediction image for the blade surface. S500. Train an icing prediction machine learning model based on the training set. Use a weighted mean squared error loss function to guide the model to prioritize fitting real ice type characteristics. Iterative training is performed using a deep learning framework. After training, the model's generalization ability is evaluated using a test set. S600. Utilize the trained icing prediction model to perform icing prediction on rotating blades.

2. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 1, characterized in that: In step S100, the blade parametric modeling includes at least the following sub-steps: S111. Determine blade design variables, including at least the inlet geometry angle. , outlet geometry angle , installation angle , consistency σ , relative maximum curvature position x mc , maximum turning angle h mc , relative maximum thickness position x mth , maximum thickness r mth , each design variable is used to control the geometric shape and aerodynamic characteristics of the blade; S1 12. Determine blade design constants, including at least blade chord length c , mid-arc inlet angle α 1. Mid-arc exit angle α 2. Leading edge radius r 1. Trailing edge radius r 2. The values ​​of each design constant are set according to the blade type and design requirements and remain fixed during the blade parametric modeling process; S113. Based on the specified blade design variables and constants, construct equations for the blade profile's median camber and thickness distribution. The median camber equation should ensure smooth continuity across the entire chord length, while the thickness distribution equation should ensure continuous differentiability with respect to the suction and pressure surface shapes. S114. The blade profile geometry is obtained by superimposing the blade profile camber line and thickness distribution. The coordinate points of the blade's leading edge, trailing edge, pressure side, and suction side are generated to form a parameterized blade profile geometry model.

3. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 1 or 2, characterized in that: In step S100, the batch design of blade profiles based on the orthogonal design method includes at least the following sub-steps: S121. Determine the horizontal range of design variables: Based on the set blade design variables and incorporating the actual design requirements and application conditions of aircraft engine rotating blades, define the range of values ​​for each design variable within the target blade profile design space to ensure that the design space covers the typical range of blade geometric characteristics. S122. Design Variable Level Division and Discretization: Based on the value range of the blade design variable and its sensitivity to blade geometry and aerodynamic performance, the number of levels of the design variable is determined and the corresponding value range is divided into multiple discrete levels with equal or unequal spacing. S123. Construct an orthogonal experimental design matrix: Select an appropriate orthogonal array based on the number of design variables and levels. Systematically combine the discrete levels of each design variable according to the arrangement rules of the orthogonal array to obtain an orthogonal design matrix that maximizes factor combination coverage under finite sample conditions. S124. Generate multiple blade profile designs: Combine each row of parameter combinations in the orthogonal design matrix with pre-set blade design constants. Generate corresponding blade geometry configurations using the constructed blade parameterized model. The resulting blade profiles should have clear boundary contours, thickness distribution, and median camber continuity. S125. Blade profile geometry verification and optimization: Each generated blade profile is geometrically verified for feasibility, abnormal blade profiles are eliminated, and blade profiles with minor geometric defects are optimized.

4. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 1, characterized in that: In step S200, for each blade profile, carrying out batch design of icing simulation examples based on the orthogonal design method includes at least the following sub-steps: S211. Determine inflow parameters and their design ranges: Based on the operating environment of aircraft engine rotor blades, determine several inflow parameters that significantly affect icing characteristics. Set the value ranges for each parameter based on icing history and / or test data. S212. Level Design and Discretization of Inflow Parameters: Based on the range of values ​​for each inflow parameter and its sensitivity to icing characteristics, each parameter is discretized at multiple levels and its level values ​​are determined using either equidistant or non-equidistant partitioning. S213. Construct an orthogonal design matrix for inflow parameters: Using the orthogonal experimental design method, select an appropriate orthogonal table based on the design variables and their number of levels to generate an orthogonal design matrix for statistically representative and spatially uniform combinations of operating parameters. S214. Generate icing simulation case designs: Combine each row of operating condition parameter combinations in the orthogonal design matrix with the blade profile design to form a complete icing simulation case design. Each case includes a clear blade geometry definition and inflow boundary condition settings.

5. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 1 or 4, characterized in that: In step S200, the icing simulation calculation and the construction of the icing simulation database include at least the following sub-steps: S221. Fluid Domain Modeling and Meshing: For selected icing simulation cases, construct the blade passage fluid domain geometry based on the blade profile. Use structured or unstructured meshing to mesh the fluid domain, and perform local mesh refinement on the blade surface to accurately capture boundary layer flow characteristics. S222. Boundary Condition Setting: Based on the flow parameters corresponding to the icing simulation example, the boundary conditions of the computational model are set accordingly. This includes setting the flow velocity and temperature at the fluid domain inlet, a pressure boundary condition at the outlet, and a no-slip wall boundary condition on the blade surface. Furthermore, periodic boundary conditions are set based on the actual operating conditions of the rotating blades to simulate the flow between blade rows. S223. Air Flow Field Calculation: After completing meshing and boundary setting, the RANS equations are used to numerically solve the air flow within the rotating blade channel. An appropriate turbulence model is selected to calculate the turbulence characteristics, ultimately obtaining air flow field results including velocity, pressure, and temperature fields. S224. Droplet Motion and Impact Calculation: Based on the air flow field solution, this study uses the Euler method and a one-way coupling strategy to calculate droplet motion and impact in a rotating aeroengine environment. This method tracks the motion of droplets in the flow field and calculates the impact characteristics of droplets on the blade surface. S225. Calculation of Icing Thermodynamic Processes: Based on droplet impact calculations and blade surface temperature distribution, a thermodynamic model of icing on blade surfaces was developed. By considering water film motion, the simulation of different ice types was enhanced, and the heat flux distribution and phase change characteristics at various locations on the blade surface were determined. S226. Ice Growth Thickness Calculation: Based on the results of the freezing thermodynamics process, the ice growth rate is calculated by analyzing the difference in heat transfer capacity between the water film and the ice layer on both sides of the ice-water interface. The ice thickness distribution is calculated for both the clear ice and frost ice modes, and the smaller value is selected as the final ice thickness to ensure the accuracy of ice type determination. S227. Ice layer profile update and mesh adaptive adjustment: Calculates a new ice layer profile based on ice layer thickness growth, adjusts blade geometry based on the original, and adaptively updates the computational mesh based on the updated blade geometry to ensure mesh quality meets the accuracy requirements of subsequent time-step simulations. S228. Multi-time-step icing evolution simulation: Based on the updated blade geometry and mesh, steps S223 to S227 are repeated while maintaining the inflow parameters unchanged, simulating the dynamic icing evolution over multiple time steps until the preset icing time or ice thickness convergence conditions are reached, ultimately obtaining complete icing evolution data. S229. Batch Case Calculation and Database Construction: After completing the full simulation process for a simulation case, iterate through all pre-set icing simulation cases, repeating steps S221 to S228 to perform simulation calculations. Key icing characteristic parameters are extracted from each simulation case, ultimately completing the construction of an icing simulation database covering multiple blade configurations and operating conditions.

6. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 5, characterized in that: In step S224, the droplet motion and impact calculation includes the following sub-steps: S2241. Establishing the Governing Equations for Droplet Motion: Establish the governing equations for droplet motion in a rotating coordinate system, taking into account aerodynamic drag, gravity, centrifugal force, and Coriolis force acting on the droplet in a rotating environment. S2242. Droplet Phase Initialization and Boundary Injection Configuration: Based on the air flow field solution, a one-way coupling strategy is employed to set the droplet release location and initial velocity conditions at the fluid domain inlet, based on the liquid water content and droplet size distribution in the incoming flow parameters, without affecting the airflow state. S2243. Numerical Solution of Droplet Trajectory: The Euler method is used to numerically integrate and solve the equations governing droplet motion. At each time step, the droplet's velocity, position, and force state are updated based on the flow field information at the current position until the droplet leaves the computational domain or impacts the blade surface. S2244. Droplet Impact Judgment and Parameter Calculation: Establish criteria for droplet impact with blade surfaces, determine the occurrence of an impact event when the droplet trajectory intersects the blade surface geometry, extract key impact parameters at each impact point, and ultimately form an impact boundary condition field.

7. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 6, characterized in that: In step S225, the calculation of the icing thermodynamic process includes the following sub-steps: S2251. Icing Control Volume Construction: The icing control volume is constructed using the first layer of the flow field solution grid near the wall. From the outside in, it is divided into an air layer, a water film layer, and an ice layer. The air layer maintains heat exchange with the main flow field, the water film layer carries the droplet impact and water film flow process, and the ice layer records the ice accumulation process. S2252. Establishing the Governing Equations for Water Film Flow: The continuity, momentum, and energy equations for water film flow are established within the control volume. The continuity equation accounts for the mass source term due to droplet impact and mass loss due to phase change. The momentum equation comprehensively considers the effects of air shear, gravity, wall pressure gradient, centrifugal force, and Coriolis force on the water film. The energy equation considers the heat exchange between the water film and the air and ice layers, as well as the energy input caused by droplet impact. S2253. Calculation of Heat Transfer and Energy Conservation in Ice Layers: A one-dimensional steady-state heat transfer equation is established in the ice layer, taking into account the heat conduction and temperature distribution of the ice layer. Adiabatic boundary conditions are applied to the ice-wall interface, and the ice-water interface temperature is kept constant at the freezing phase transition temperature. The heat transfer path within the ice layer and the heat exchange between the ice layer and the wall and water film are simulated. The temperature field within the ice layer is solved to provide boundary conditions for the phase transition calculation at the ice-water interface. S2254. Phase Transition Behavior and Ice Type Identification Mechanism: Criteria for determining frost ice, clear ice, and mixed ice are established based on the water film flow state, temperature field distribution, and impact flux characteristics. Clear ice forms when the water film can flow on the surface and partially freezes. Frost ice forms when the impacting droplet freezes instantaneously without water film flow. Mixed ice forms when both mechanisms coexist. Based on the phase transition mechanisms of different ice types, thermophysical properties and phase interface boundary conditions are adjusted to achieve adaptive ice type simulation. S2255. Calculation of heat flux density distribution and phase change characteristics: By solving the coupled thermodynamic equations governing the water film and ice layer within the blade, the heat flux density distribution and phase change latent heat flux at each location on the blade surface are calculated, providing thermodynamic boundary conditions for subsequent ice growth calculations.

8. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 7, characterized in that: In step S226, ice layer growth thickness calculation includes the following sub-steps: S2261. Establishing the governing equation for ice growth rate: An energy conservation model is established around the ice-water interface. By analyzing the release and heat transfer of latent heat of freezing at the ice-water interface and considering the temperature gradient differences in the water film and ice layer on both sides of the interface, an equation governing ice growth rate is established for the time-dependent change of ice thickness. S2262. Dual-mode results comparison and minimum value selection: Calculate the ice thickness distribution for both the clear ice and frost ice modes. Compare the two calculated results at each surface location and select the smaller value as the final ice thickness at that location. S2263. Final determination of ice type and thickness distribution output: The final ice type at each location on the blade surface is determined based on the minimum value selection principle. When the clear ice thickness is less than the frost ice thickness, it is determined as clear ice; otherwise, it is determined as frost ice. When the clear ice thickness and frost ice thickness are close, it is determined as mixed ice. Finally, the complete ice thickness distribution data and ice type distribution are output.

9. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 1, characterized in that: In step S300, the construction of the multi-source icing database and the division of the training set and the test set include at least the following sub-steps: S301. Experimental Data Collection: Based on the icing simulation database, further collect blade icing experimental data, including at least several sets of typical icing experimental data collected from public literature and several sets of data measured from rotating blade icing experiments conducted using a self-built ice wind tunnel platform; S302. Multi-source icing database construction: Data cleaning, format conversion, and normalization are performed on blade icing simulation data, experimental data from literature, and independent experimental data. This data is then integrated with the experimental and simulation data to form a multi-source icing database with a unified data structure and parameter dimensions. S303. Training and test set division: Based on the principle of multidimensional uniformity of icing type, blade profile, and operating parameters, a stratified sampling strategy is used to divide the icing simulation database into training and test sets at a predetermined ratio. A portion of the data is reserved as a validation set for model parameter tuning and performance optimization. S304. Training Group Split: Split the training set into multiple training groups according to the preset group size, with each training group containing an equal amount of simulation data. Then, add literature experimental data and independent experimental data to each training group in equal proportions, so that each training group consists of both simulation data and experimental data.

10. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 1, characterized in that: In step S400, the ice prediction machine learning model adopts an MLP architecture and is designed in combination with a DNN network. It consists of an input layer, a hidden layer, and an output layer, wherein: The input layer is designed to receive the incoming flow conditions and blade parameters. The input data is transmitted to the network in the form of a multi-dimensional numerical vector, which serves as the basis for model perception and representation. The hidden layer is designed to consist of three fully connected layers and three deconvolution layers. The fully connected layers constitute the main structure of the MLP network, which is used to perform layer-by-layer abstract expression and nonlinear transformation of input features. Each fully connected layer contains a certain number of neurons and uses the ReLU activation function for nonlinear transformation. The deconvolution layer is a DNN module integrated into the main structure of the MLP network. It is used to gradually decode the abstract features of the previous layer into spatially distributed icing prediction results, realizing the extraction and spatial reconstruction of icing pattern features. Each deconvolution layer is also followed by the ReLU activation function for nonlinear processing. The output layer is designed to generate a predicted ice shape image corresponding to the input flow conditions and blade parameters through convolution operations, which at least contains the spatial distribution information of the ice thickness on the blade surface.

11. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 1, characterized in that: In step S500, the icing prediction machine learning model is trained based on the training set. Each batch of training data is set to a fixed number of simulation samples and experimental samples, and the weighted mean square error WMSE = α × MSE is used. sim +β×MSE exp As the loss function, where α is the simulation error MSE sim The weight of β is the experimental error MSE exp The weight of , and β>α; Use a deep learning framework for iterative model training, and combine dynamic learning rate adjustment, batch normalization, dropout, and early stopping mechanisms to improve training convergence efficiency and stability. The training process includes an iterative cycle of data loading, forward propagation, loss calculation, backpropagation, and model parameter updates. Real-time monitoring of the training process, including loss trend, model weight updates, and neuron activation distribution, prevents overfitting through early stopping. When the validation set loss does not decrease within a preset period, early stopping is triggered and the optimal model weights are saved. Hyperparameters are tuned, including the number of neurons in the fully connected layer, the step size of the deconvolution layer, and the learning rate decay strategy, until the model converges. After training, the test set was used to evaluate the model's ability to predict icing on non-training data. The model's generalization ability and prediction accuracy were verified by quantitatively comparing the error levels between the predicted ice type and the real ice type in terms of thickness distribution, area share and image similarity indicators.

12. The method for predicting icing of rotating blades of an aircraft engine based on machine learning according to claim 1, characterized in that: In step S600, when performing icing prediction on rotating blades of an aircraft engine using the trained icing prediction model, the method includes at least the following sub-steps: S601 prepares input parameters, including blade parameters, flow parameters and engine operating parameters, as the input basis of the icing prediction model; S602. Preprocess the input parameters, organize them according to the unified format and dimensions used when training the model, and use the same normalization method as in the training phase; S603. The preprocessed input parameters are input to the icing prediction model, feature extraction and nonlinear mapping are performed through the hidden layer, and the corresponding blade ice shape image is generated by the output layer; S604. Post-processing the output ice shape image to extract ice thickness distribution and ice location, and quantify the degree of ice formation; S605. Present the location and shape of ice on the blade surface, and output information on the degree of ice formation and risk level.

Images