A method for predicting icing of rotating blades of an aero-engine based on machine learning

Abstract

The application discloses an aero-engine rotating blade icing prediction method based on machine learning, relates to multiple technical fields of computational fluid dynamics, heat and mass transfer, multiphase flow simulation and artificial intelligence, and is used for quickly and accurately predicting icing forms of aero-engine fan or compressor rotating blade surfaces under different inflow conditions and blade type parameters. The method firstly constructs a parameterized model of the rotating blade, generates multiple sets of blade type and condition combinations in combination with orthogonal experimental design, and constructs a multi-source icing database through simulation and experiment; then, an icing prediction model is constructed based on a multilayer perception machine and a deconvolution neural network, and corresponding ice shape images can be output after inflow parameters and blade type parameters are input; finally, model optimization is realized through training and verification, and the recognition and fitting capacity for real ice shapes is improved. The application can effectively improve the efficiency and precision of icing prediction, and provides technical support for anti-icing design and safe operation of the aero-engine.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of aero-engine aerodynamic design and icing prediction, and relates to a compressor rotating blade icing prediction method, more particularly to an aero-engine rotating blade icing prediction method based on machine learning, which is used to quickly and accurately predict the icing shape and risk of aero-engine rotating blades under various flight conditions, and improve the safety and reliability of engine operation. BACKGROUND

[0002] During the flight of an aircraft, when passing through a cloud containing supercooled water droplets, the first few stages of the aero-engine blades are prone to icing. Once the aero-engine blades icing occurs, it will bring great hidden dangers to the safe operation of the aero-engine, seriously threaten the flight safety, and cause many safety accidents.

[0003] Firstly, the icing of the blades will change the original aerodynamic shape, leading to a decline in aerodynamic performance, seriously affecting the compressor pressurization capacity and operation stability. Secondly, due to the high-speed rotation of the engine, the icing parts are prone to fall off, and the fallen ice blocks are extremely 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, the icing of the inlet components will reduce the flow area, reduce the airflow into the engine, and cause the engine thrust to decrease, and in severe cases, it will also cause compressor surge. Finally, if the blade surface is iced, it will cause the rotor to unbalance, causing the power device to vibrate violently, and in severe cases, it can cause the engine rotor bearing to be damaged.

[0004] In view of the fact that the icing of the aero-engine blades has so many hazards, it is particularly important to accurately predict the icing of the blades. Knowing in advance whether the blades will ice, the degree, position and other information of the icing can help the crew to take effective anti-icing and de-icing measures in advance, avoid engine failures caused by icing, and ensure the safe flight of the aircraft. At the same time, accurate icing prediction is also helpful for optimizing the design of the engine, improving the anti-icing system of the engine, improving the operation reliability and stability of the engine in icing environment, and reducing the maintenance cost and flight delay caused by icing problems, which has important significance for the safety and economic benefits of the aviation industry. Therefore, the development of the prediction of the icing of the rotating blades of the aero-engine, and the formation of a fast and accurate prediction method, has important significance for the safety of the aero-engine.

[0005] Current researches on aircraft icing phenomenon mainly fall into three categories: experimental research, empirical estimation and numerical simulation. Experimental research explores the aircraft icing and its influence on flight performance by actual flight test or using icing wind tunnel to simulate icing weather conditions; empirical estimation relies on the experience charts and formulas obtained by experiments for calculation; numerical simulation solves the mathematical model after reasonable simplification by computer to analyze the icing process and its influence on aircraft performance.

[0006] However, the above three methods have certain defects. First, experimental research method can directly observe the icing phenomenon and measure the influence of icing on flight performance. However, this method requires large investment and long research period, and the experimental object has certain limitations. Second, the application range of empirical estimation is limited, and it cannot simulate the icing process, so this method is usually used for estimation in the preliminary stage of aircraft anti- / de-icing system design. Third, as an economical and efficient alternative, numerical simulation can effectively simulate the icing process under different flight conditions by reasonably simplifying the actual problem and constructing a mathematical model. However, the numerical simulation method still has certain defects:

[0007] First, the numerical simulation method is time-consuming. Although it shortens the long period compared with the experimental method, millions of calculation grids are still very time-consuming in solving the long-time icing process. At this point, the calculation time of the high-speed rotating structure such as an aero-engine is more prominent than the solution of the wing outflow, because the time characteristic scale of the aero-engine is very small (usually 10 -6 s), while the wing outflow is a static component, so the time characteristic scale is larger, and the time-consuming is less in solving the icing problem of the same time. Second, the precision of numerical simulation method itself also has certain defects, especially for the clear ice condition (water droplets do not completely freeze into ice after impact), because the water film movement cannot be accurately described, the precision has certain error compared with the experimental result. Finally, it is difficult to simulate icing in three-dimensional rotating environment. The aero-engine is a rotating component, and there are complex strong three-dimensional and strong unsteady flow characteristics inside. Under this condition, how to consider the influence of rotation on the traditional two-dimensional static component icing method and expand it to three-dimensional situation itself has great difficulty, so the existing simulation model makes some assumptions more or less.

[0008] In addition, in the prior art, Chinese invention CN116579383A and CN111680454A disclose icing prediction methods for wind turbine blades, but they mainly predict the icing quality or the single state of whether the wind turbine blade is iced, cannot provide key ice type distribution information, cannot solve the problem of transient flow field-icing coupling modeling of high-speed rotating blades of an aero-engine, and do not consider the influence of rotating effects such as centrifugal force and Coriolis force. The aero-engine stator blade ice crystal icing prediction system disclosed in CN117871592A still essentially establishes the correlation between temperature and icing thickness through ground simulation tests, relies on experimental means, and does not form a general prediction model that can comprehensively input multi-dimensional parameters and quickly output ice types.

[0009] In summary, the experimental research method is time-consuming and costly, and cannot be used as a regular prediction method for aero-engine blade ice types. The accuracy of the empirical estimation is low, and the application range is limited, so it cannot be used as a regular prediction method for icing. The numerical simulation method has time efficiency and accuracy, and has been widely used in daily icing research. However, the accuracy of numerical simulation still has a lot of room for improvement, and the time-consuming problem is also very prominent in aero-engine simulation. If a faster and more accurate icing prediction method can be developed and appropriately corrected based on experimental results, it will have extremely important value for improving the icing prediction ability of an aero-engine. SUMMARY

[0010] (I) Invention purposes

[0011] In view of the defects and deficiencies of the prior art in the icing prediction of aero-engine blades, such as long time consumption, low accuracy and the like, the present application aims to provide an aero-engine rotating blade icing prediction method based on machine learning, which comprises the following steps.

[0012] (II) Technical solutions

[0013] To achieve the purposes of the application and solve the technical problems, the application adopts the following technical solutions:

[0014] An aero-engine rotating blade icing prediction method based on machine learning is used to quickly and accurately predict the icing shape on the surface of a rotating blade of an aero-engine fan or compressor under different incoming flow conditions and blade structure parameters, and at least comprises the following steps:

[0015] S100. Leaf parameterized modeling and orthogonal design:

[0016] A parameterized geometric model of an aero-engine rotating blade is established based on blade design variables and blade design constants; the blade design variables are designed with multiple levels based on the orthogonal experimental method, and are combined with the blade design constants to generate multiple sets of blade profile schemes covering the variation range of typical geometric characteristics of the blade;

[0017] S200. Simulation example design and icing simulation database construction:

[0018] For each blade profile scheme, the incoming flow condition parameters are designed to have multiple levels within the preset range based on the orthogonal experimental method, and multiple sets of incoming flow condition parameter combinations covering the variation range of typical working conditions of the aero-engine rotating blade are generated, each incoming flow condition parameter combination corresponding to an icing simulation example; for each blade profile scheme and each incoming flow condition parameter combination, blade icing simulation calculation is performed to obtain corresponding blade surface ice shape data, and an icing simulation database is constructed accordingly;

[0019] S300. Experimental data collection and multi-source icing database construction:

[0020] Experimental data of aero-engine rotating blade icing are collected and processed uniformly with blade icing simulation data to construct a multi-source icing database that integrates experimental data and simulation data and has a unified data structure and parameter dimension, and the multi-source database is divided into a training set and a test set for the construction and training of a machine learning model and the evaluation of its generalization ability;

[0021] S400. Construction of icing prediction machine learning model:

[0022] An icing prediction machine learning model is established, which includes an input layer, a hidden layer and an output layer, the input layer is used to receive a multi-dimensional 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 the abstract features are restored to the spatial dimension of the ice shape image, and the output layer is used to generate a blade surface icing prediction image corresponding to the input parameters;

[0023] S500. Training and verification of icing prediction machine learning model:

[0024] The training set is used to train the icing prediction machine learning model, a weighted mean square error loss function is used to guide the model to preferentially fit the real ice shape features, and a deep learning framework is used for iterative training, and the generalization ability of the model is evaluated using the test set after training;

[0025] S600. Aero-engine rotating blade icing prediction:

[0026] The icing prediction model trained is used to predict icing of a rotating blade of an aero-engine.

[0027] (Three) Technical effects

[0028] Compared with the prior art, the icing prediction method of the rotating blade of the aero-engine based on machine learning has the following beneficial and remarkable technical effects:

[0029] (1) The traditional numerical simulation method needs to be calculated for a long time (such as solving the icing process of millions of calculation grids) for a complex three-dimensional rotating flow field, while the present application directly maps the incoming flow condition, the structure parameter and the icing profile through the machine learning model, and can complete the prediction in seconds or minutes, which is more than 90% shorter than the traditional numerical simulation, and meets the real-time early warning and rapid evaluation requirements of the aero-engine.

[0030] (2) The model training data includes more than ten thousand groups of autonomous simulation data, literature experimental data (50 groups) and autonomous experimental data (50 groups). And in each training data, the 100 fixed experimental data are fused, which fully corrects the precision and accuracy of the simulation data. Through the weighted loss function (simulation error weight 0.4, experimental error weight 0.6), the model is forced to preferentially fit the real experimental data, especially for complex icing shapes such as clear ice (traditional numerical simulation leads to error due to insufficient water film motion description), the prediction accuracy is significantly improved compared with single simulation model.

[0031] (3) The icing prediction can be completed without relying on high-cost ice wind tunnel experiments or large-scale numerical simulation, which reduces the hardware investment and time cost of aero-engine icing research, and is especially suitable for rapid iterative optimization in the early design stage. It can process multiple variable inputs such as incoming flow temperature, liquid water content and blade geometric parameters, and cover different flight conditions (such as low-altitude cloud and mist environment, high-altitude supercooled water droplet area) and engine structure design schemes, providing full-condition data support for ice prevention system optimization and blade aerodynamic design.

[0032] (4) The present application breaks through the bottleneck of traditional methods in efficiency and accuracy, and provides a new way of fast, accurate and economical icing prediction for aero-engine, which has remarkable engineering application value and economic benefit for improving flight safety, reducing maintenance cost and promoting engine design optimization. DETAILED DESCRIPTION

[0033] Figure 1 The flow chart of the icing prediction method of the rotating blade of the aero-engine based on machine learning is shown in the figure.

[0034] Figure 2 The schematic diagram of the blade type geometry is shown in the figure.

[0035] Figure 3Grid partitioning for fluid domain;

[0036] Figure 4 Ice shape prediction for different operating conditions, where (a)-(d) correspond to operating condition 1-4, respectively.

[0037] Figure 5 Ice shape prediction for different operating conditions, where (a)-(d) correspond to operating condition 1-4, respectively.

[0038] Figure 6 Ice shape prediction for different operating conditions, where (a)-(d) correspond to operating condition 1-4, respectively.

[0039] Figure 7 Ice shape prediction for different operating conditions, where (a)-(d) correspond to operating condition 1-4, respectively.

[0040] Figure 8 Ice shape prediction for different operating conditions, where (a)-(d) correspond to operating condition 1-4, respectively.

[0041] Figure 9 Ice shape prediction for different operating conditions, where (a)-(d) correspond to operating condition 1-4, respectively.

[0042] Figure 10 Ice shape prediction for different operating conditions, where (a)-(d) correspond to operating condition 1-4, respectively.

[0043] Figure 11 Ice shape prediction for different operating conditions, where (a)-(d) correspond to operating condition 1-4, respectively. DETAILED DESCRIPTION

[0044] The present application aims to provide a machine learning-based aero-engine rotating blade icing prediction method for quickly and accurately predicting the icing shape of the surface of the rotating blades of an aero-engine fan or compressor under different incoming flow conditions and blade parameter conditions. To make the purpose, technical solution and advantages of the present application clearer, the technical solution in the embodiments of the present application will be described in more detail below in combination with the drawings in the embodiments of the present application. According to the technical solution content of the present application, the implementation process will be introduced in detail in combination with the step decomposition and key parameter description in the technical solution:

[0045] As a specific example, as shown in Figure 1 The machine learning-based aero-engine rotating blade icing prediction method of the present application mainly includes the following steps when implemented:

[0046] S100. Blade parameterized model establishment and orthogonal design:

[0047] An aero-engine rotating blade parameterized geometric model is established based on blade design variables and blade design constants. The blade design variables are designed at multiple levels based on the orthogonal experimental method, and after being combined with the blade design constants, a plurality of blade profile schemes covering the variation range of typical geometric characteristics of the blades are generated.

[0048] In the embodiment of the present application, the blade parameterized modeling in step S100 at least includes the following sub-steps:

[0049] S111. Determine the design variables, define the blade design variables as the inlet geometric angle , the outlet geometric angle , the installation angle , the solidity σ , the relative maximum camber position x mc , the maximum turning angle h mc , the relative maximum thickness position x mth , the maximum thickness r mth . The above parameters are variables in the blade type design process, and different types of blade type distributions are generated by adjusting the numerical values of the above parameters. Each design variable is used to control the blade geometry and aerodynamic characteristics, and the value range should cover the typical blade geometric characteristics in the target application scenario.

[0050] S112. Determine several blade type design constants, including blade chord length c , mean camber inlet angle α 1, mean camber outlet 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 these blade constants remain unchanged during the design process. The same numerical value is used for different blade types to constrain the basic geometric dimensions and boundary shapes of the blade, and to maintain a stable structure reference in the parameterized expression. In combination with the above blade design variables and constants, a unique blade type geometry can be constructed to meet the design requirements for different geometries.

[0051] S113. Based on the set blade design variables and design constants, Bezier curve, B-spline curve or piecewise polynomial combination curve is used to construct the expression equation of the mean camber line and thickness distribution of the blade type, respectively. The mean camber line expression equation should ensure its continuity and smoothness in the full chord direction, and the thickness distribution expression equation should ensure the continuity and derivability of the suction surface and pressure surface shape. Specifically:

[0052] In the embodiment of the present application, cubic polynomial combination curve is used to construct the mean camber line of the blade type, and the expression is:

[0053]

[0054] Wherein, x and y are the spatial coordinates of the point on the mean camber line, a 10 ​a 23 are coefficients, and the specific values are:

[0055]

[0056] In the above coefficient expressions, x max and y max are the horizontal and vertical coordinate values of the maximum camber point, and the values are:

[0057]

[0058] In this embodiment, the thickness distribution of the blade profile is also constructed by using a cubic polynomial combined curve, and the expression is:

[0059]

[0060] wherein x is the horizontal coordinate of the point on the camber line, h is the thickness of the blade profile, a 30 ~ a 43 are coefficients, and the specific values are:

[0061]

[0062] In the above coefficient expressions, y 0 is the thickness value reference at the maximum thickness point, S is the length of the camber line.

[0063] S114. The blade profile geometry distribution is obtained by superimposing the blade profile camber line coordinates and the thickness distribution, the coordinate points of the blade leading edge, trailing edge and pressure surface / suction surface are generated, and the parameterized blade profile geometry is formed, as shown in Figure 2 .

[0064] In the embodiment of the application, the step S100 of carrying out batch design of the blade profile scheme based on the orthogonal design method at least includes the following sub-steps:

[0065] S121. Determining the level range of the design variable: based on the set blade design variable, combining the actual design requirements and application working conditions of the rotating blade of the aero-engine, the value interval of each design variable in the target blade design space is determined to ensure that the design space covers the typical blade geometry characteristic change range;

[0066] S122. Design variable level division and discretization processing: according to the value interval of the blade design variable, combining the sensitivity of the variable to the blade geometry shape and aerodynamic performance, the number of levels of the design variable is determined and the corresponding value interval is divided into multiple discrete level values at equal or unequal intervals.

[0067] S123. Constructing an orthogonal experiment design matrix: selecting an appropriate orthogonal table according to the number of design variables and the number of levels, and systematically combining the discrete level values of each design variable according to the arrangement rules of the orthogonal table to obtain an orthogonal design matrix with the maximum factor combination coverage under the condition of limited samples;

[0068] S124. Generating multiple sets of airfoil design schemes: combining each row of parameter combinations in the orthogonal matrix with the preset airfoil design constants, generating corresponding airfoil geometries one by one through the constructed airfoil parameterized model, and the obtained sample airfoils should have clear boundary contours, thickness distribution and mid-chord line continuity;

[0069] S125. Geometric verification and optimization of airfoil schemes: checking the geometric feasibility of each generated airfoil scheme, eliminating abnormal airfoils, and optimizing airfoils with slight geometric defects to finally form a standardized airfoil scheme library.

[0070] More specifically, in the airfoil orthogonal experiment design, the L1000 (10 8 ) orthogonal table is used to design ten levels for eight key variables (including inlet geometric angle , outlet geometric angle , installation angle , solidity σ , relative maximum camber position x mc , maximum turning angle h mc , relative maximum thickness position x mth , maximum thickness r mth ), and the other five variables (including blade chord length c , mid-chord line inlet angle α 1, mid-chord line outlet angle α 2, leading edge radius r 1, trailing edge radius r 2) are fixed as typical values to generate 1000 airfoil schemes, covering the commonly used geometric parameter range of the engine compressor.

[0071] S200. Design, calculation and construction of icing simulation database for icing simulation examples:

[0072] For each blade profile scheme, based on the orthogonal experiment method, the incoming flow condition parameters are multi-level valued in the preset range, a plurality of groups of incoming flow condition parameter combinations covering the typical working condition change range of the rotating blades of the aero-engine are designed and generated, and each incoming flow condition parameter combination corresponds to an icing simulation example; the blade icing simulation calculation is carried out for each blade profile scheme and each incoming flow condition parameter combination, the corresponding blade surface ice type data are obtained, and the icing simulation database is constructed accordingly.

[0073] In the embodiment of the application, when the batch design of the icing simulation example based on the orthogonal design method is carried out for each blade profile scheme in step S200, at least the following sub-steps are included:

[0074] S211. Determine the incoming flow condition parameters and their design range: according to the working environment of the rotating blades of the aero-engine, determine several incoming flow condition parameters that have a significant impact on the icing characteristics, and set the value interval of each parameter in combination with the icing history and / or test data;

[0075] S212. Incoming flow condition parameter level design and discretization: based on the value interval of each incoming flow condition parameter and its influence sensitivity on the icing characteristics, multi-level discretization processing is performed on each parameter, and an equidistant or non-equidistant division method is used to determine the level value;

[0076] S213. Construct the orthogonal design matrix of the incoming flow condition parameters: using the orthogonal experimental design method, according to the design variables and their level number, select an appropriate orthogonal table to generate an orthogonal design matrix of the working condition parameter combination with statistical representativeness and spatial uniformity;

[0077] S214. Generate the design scheme of the icing simulation example: combine each row of the working condition parameter combination in the orthogonal design matrix with the blade profile scheme to form a complete icing simulation example design, and each example includes a clear blade geometry definition and incoming flow boundary condition setting.

[0078] More specifically, in the embodiment, the working condition orthogonal experiment method design uses an L50(5 4 ) orthogonal table to perform five-level design on four working condition parameters (including incoming flow velocity v , incoming flow temperature T, liquid water content LWC, and droplet size MVD), and generates 50 groups of working condition schemes covering the commonly used working condition range of the engine compressor. In combination with 1000 groups of blade profiles, a total of 1000x50=50000 groups of icing simulation examples are generated.

[0079] Further, in the embodiment of the application, the simulation calculation based on the icing simulation example in step S200 and the construction of the icing simulation database at least include the following sub-steps:

[0080] S221. Fluid domain modeling and meshing:

[0081] For the selected icing simulation example, a fluid domain model containing the rotating blade passage of an aero-engine is constructed based on the corresponding blade profile scheme. The fluid domain is divided using structured or unstructured grid technology. Structured grid is preferred for dividing the fluid domain, as it not only improves computational efficiency but also ensures the accuracy of flow field simulation. In particular, in the key area of the blade wall, local grid refinement is performed to capture the complex flow phenomena and subtle pressure changes near the wall. In terms of grid quantity control, the total number of grids is strictly controlled between 100,000 and 200,000, which not only ensures the reasonable use of computing resources but also meets the general accuracy requirements of compressor cascade flow field simulation. Selecting one of the typical blade profiles as an example, the corresponding grid division structure diagram is drawn as shown in Figure 3 .

[0082] S222. Boundary condition setting of calculation model:

[0083] According to the inflow condition parameter combination corresponding to the icing simulation example, the boundary conditions of the calculation model are set accordingly, at least including setting the inflow velocity and 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 the periodic boundary condition according to the actual working state of the rotating blade to simulate the flow between blade rows.

[0084] S223. Air flow field calculation:

[0085] After completing the grid division and boundary setting, the Reynolds-averaged Navier-Stokes (RANS) equation is used to calculate the air flow in the rotating blade passage, obtaining the air flow field distribution including velocity field, pressure field and temperature field, which can reasonably control the calculation time and resources while meeting the calculation requirements. When calculating the flow field, an appropriate turbulence model is selected for turbulence characteristic calculation, and the Spalart-Allmaras model is preferred. The S-A model uses the Boussinesq assumption to establish the relationship between the eddy viscosity coefficient and the corresponding turbulence model transport variable, and the S-A model has good applicability for structured and unstructured grids. Figure 4 Taking the velocity and pressure parameters as examples, the flow field distribution around one of the blades is shown.

[0086] S224. Droplet motion and impact calculation:

[0087] Based on the solution results of the air flow field, the Euler method is used to solve the water droplet motion equation under the action of air. Due to the small content of liquid droplets in air, a one-way coupling method is generally used, that is, the air flow field is solved first, and then the liquid droplet motion under the action of the air flow field is solved, and the impact characteristics of the liquid droplets and the blade surface are calculated, including the impact position, impact speed, impact angle and impact momentum distribution, to provide the liquid droplet impact boundary conditions for subsequent icing calculation. Specifically, the liquid droplet motion and impact calculation can be further divided into the following sub-steps:

[0088] S2241. Establishing liquid droplet motion control equation: the liquid droplet motion control equation is established in the rotating coordinate system, considering the aerodynamic drag, gravity, centrifugal force and Coriolis force of the liquid droplets in the rotating environment, wherein the drag is calculated based on the relative speed difference between the liquid droplets and the airflow, and the centrifugal force and the Coriolis force are determined according to the engine speed and the liquid droplet motion state. The present application expands the liquid droplet motion equation in the rotating coordinate system, and considers the influence of centrifugal force and Coriolis force. The specific control equation is as follows:

[0089] 1) Continuity equation:

[0090] 2) Momentum equation:

[0091]

[0092] In the formula, is the water droplet density, is the water droplet relative speed, is the water droplet volume fraction, g is the gravity acceleration, t is the time, ω is the rotation angular velocity. F is the air drag on the water droplet, and , wherein is the relaxation time, d p is the average radius of the water droplets. f is a function of the Reynolds number Re r , which takes the following form:

[0093]

[0094] In the formula, v ar is the air relative motion speed, μ is the air viscosity coefficient.

[0095] S2242. Droplet phase initialization and boundary injection configuration: based on the air flow field solution results, a one-way coupling strategy is adopted, without affecting the air flow state, according to the liquid water content and droplet size distribution in the incoming flow condition parameters, the droplet release position and initial velocity condition are set at the inlet of the fluid domain;

[0096] S2243. Droplet trajectory numerical solution: Euler method is used to numerically integrate the droplet motion control equation, and the droplet velocity, position and force state are updated according to the flow field information at the current position in each time step until the droplet leaves the calculation domain or collides with the blade surface;

[0097] S2244. Droplet impact judgment and parameter calculation: establish the impact judgment criterion of droplet and blade surface, when the droplet trajectory intersects with the blade surface geometry, determine the occurrence of impact event, extract the key impact parameters of each impact point, at least including impact position, velocity vector, impact angle and impact momentum, and finally form the impact boundary condition field. Figure 5 The droplet water content cloud is taken as an example for display.

[0098] S225. Ice thermodynamic process calculation:

[0099] After the droplet motion calculation is completed, the droplet distribution on the blade surface is extracted, and the blade surface temperature distribution is extracted, and the calculation of the ice thermodynamic process on the blade surface is carried out. In the calculation process, the simulation of different types of ice (frost ice, clear ice and mixed ice) is enhanced by considering the water film movement, and the heat flux density distribution and phase change characteristics of each position on the blade surface are obtained. Specifically, the ice thermodynamic process calculation can be subdivided into the following sub-steps:

[0100] S2251. Ice control body construction: the first layer of grid near the wall of the grid used for solving the air flow field is used as the grid for ice numerical simulation, and an ice control body is established thereon, as shown in Figure 6 The ice control body is divided into three layers from outside to inside, which are air layer, water film layer and ice layer, as shown in Figure 7 The water film flow and ice layer growth in the control body are described by constructing and solving the control equation.

[0101] S2252. Water film flow control equation establishment:

[0102] In the control volume, continuity, momentum, and energy equations for water film flow are established. The continuity equation considers the mass source term of droplet impact and the mass loss caused by phase change. The momentum equation comprehensively considers the effects 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 and ice layers, as well as the energy input caused by droplet impact. Specifically, this invention considers the flow of unfrozen water film in the freezing thermodynamic model, thereby enhancing the simulation capability for both open and mixed ice. Furthermore, the freezing thermodynamic process in a rotating coordinate system is extended to consider the effects of centrifugal force and Coriolis force on the surface water film. The specific control equations are as follows:

[0103] 1) Continuity equation for water film flow:

[0104]

[0105]

[0106] In the formula v r The relative flow velocity of the water film. u and v For this speed at x , y The two components of direction (two perpendicular directions on the icy surface), H w For water film in z The height of the direction (normal to the icy surface), H i For ice thickness, m imp The droplet impact rate is [value]. ρ i The density of ice, ρ w The density of water, t For time.

[0107] 2) Momentum equation for water film flow:

[0108] Water film flow is primarily influenced by air shear force and pressure; in rotating environments, centrifugal force and Coriolis force must also be considered. Since the velocity of the water film is very small, it can be considered as steady-state flow. Water film flow also satisfies the incompressible Navier-Stokes equations, which can be written in the following form:

[0109]

[0110] In the formula, the right side represents the gravity, viscous stress, and air pressure acting on the water film. p This refers to the local gas pressure.

[0111] 3) Energy equation of water film flow

[0112] Since the thickness of water film is very small and the flow velocity is also very small, the heat transfer process in water film can be considered as a steady process in each icing time step, satisfying the energy equation:

[0113]

[0114] where u is the water film flow velocity, a w is the thermal diffusivity of water, T w is the water film temperature. The temperature on the water-ice interface is always the icing phase change temperature, and there is convective heat transfer and energy exchange caused by water droplet impact on the water-air interface.

[0115] S2253. Ice layer heat transfer and energy conservation calculation:

[0116] According to the same method, the energy equation in the ice layer can be obtained as: , T i is the ice layer temperature, considering the heat conduction process and temperature distribution of the ice layer, the one-dimensional steady-state heat transfer control equation uses an adiabatic boundary on the ice-wall interface, the temperature on the water-ice interface is always the icing phase change temperature, and there is convective heat transfer and energy exchange caused by water droplet impact on the water-air interface. Simulate the heat transfer path in the ice layer and the heat exchange process between the wall and the water film layer, and provide boundary conditions for the phase change calculation on the ice-water interface by solving the temperature field in the ice layer.

[0117] S2254. Phase change behavior and ice type judgment mechanism: Based on the water film flow state, temperature field distribution and impact flux characteristics, establish the judgment criteria for rime ice, clear ice and mixed ice, when the water film can flow on the surface and partially freeze, clear ice is formed, when the impact liquid droplets freeze instantaneously and there is no water film flow, rime ice is formed, when the two mechanisms coexist, mixed ice is formed, and the thermal physical parameters and phase interface boundary conditions are adjusted based on the phase change mechanism of different ice types to realize the adaptive simulation of ice type;

[0118] S2255. Heat flux distribution and phase change characteristics calculation: By solving the coupled thermodynamic equations of water film and ice layer in the control body, the heat flux distribution and phase change latent heat flux at each position on the blade surface are calculated, which provides the thermodynamic boundary conditions for the subsequent ice layer growth calculation.

[0119] S226. Icing growth thickness calculation:

[0120] Based on the calculation results of icing thermodynamic process, the ice layer growth rate is calculated by analyzing the difference in heat transfer capacity of the latent heat of freezing at the ice-water interface on both sides of the water film and ice layer, the ice layer 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, and the accuracy of the icing type judgment is ensured. Specifically, the ice layer growth thickness calculation can be divided into the following sub-steps:

[0121] S2261. Ice layer growth rate control equation: In the icing control body, an energy conservation model is established around the ice-water interface, the release and heat transfer process of the latent heat of freezing at the ice-water interface is analyzed, the temperature gradient difference in the water film and ice layer on both sides of the interface is considered, and the ice layer growth rate control equation of the ice layer thickness changing with time is established. Specifically, the latent heat released by the ice 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 the base. According to the Stefan condition, the ice layer growth rate in the control body mainly depends on the heat transfer capacity in the water film and ice layer, and the ice layer thickness growth rate can be expressed as follows:

[0122]

[0123] In the formula, L f is the latent heat of freezing phase change, and L f = 334400 J / kg, λ i is the thermal conductivity of ice, λ w is the thermal conductivity of water.

[0124] S2262. Comparison and minimum value selection of double mode results: The ice layer thickness distribution under the clear ice and frost ice freezing mode is calculated respectively, the numerical value of the two calculation results at each surface position is compared, and the smaller value of the two is selected as the final ice layer thickness at the position, which ensures the accuracy and rationality of the icing type judgment. That is, after the ice layer thickness of clear ice freezing is calculated according to the above equation, it needs to be compared with the frost ice freezing ice thickness. If it exceeds the frost ice freezing thickness, the final ice thickness is the frost ice thickness, and its value is .

[0125] S2263. Final determination of icing type and thickness distribution output: Based on the minimum value selection principle, the final icing type of each position on the blade surface is determined, 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 two are close, it is determined as mixed ice, and finally the complete ice layer thickness distribution data and icing type distribution are output. Figure 8 Taking a certain type of blade and working condition as an example, the icing ice type is displayed, and the icing mainly occurs at the leading edge of the blade.

[0126] S227. Ice layer type surface update and mesh self-adaptive adjustment:

[0127] Calculate a new ice layer type surface according to the ice layer growth thickness, adjust the blade geometry on the basis of the original blade geometry, and self-adaptively update the calculation mesh based on the updated blade geometry and in combination with the dynamic mesh technology, and ensure that the mesh quality meets the accuracy requirements of subsequent time step simulation calculation;

[0128] S228. Multi-time step icing evolution simulation:

[0129] Based on the updated blade geometry and mesh, repeat steps S223 to S227 while keeping the incoming flow condition parameters unchanged, to realize multi-time step dynamic evolution process simulation, until the preset icing time or ice layer thickness convergence condition is reached, and finally obtain complete icing evolution history data;

[0130] S229. Batch calculation of example and construction of database:

[0131] After completing the full flow simulation of one simulation example, iterate through all the preset icing simulation examples, repeat steps S221 to S228 for simulation calculation, and extract key icing characteristic parameters from each simulation example, to finally complete the construction of an icing simulation database covering multiple blade type schemes and working conditions.

[0132] S300. Construction of icing database from multiple sources of experiments and simulations

[0133] Collect experimental data of icing on rotating blades of an aero-engine, and perform unified data processing on the blade icing simulation data, to construct a multi-source icing database that fuses experimental data and simulation data and has a unified data structure and parameter dimension, and divide the multi-source database into a training set and a test set according to a set proportion, for model construction and training and generalization ability evaluation. Specifically, in this embodiment, the construction of the multi-source icing database and the division of the training set and the test set include the following sub-steps:

[0134] S301. Simulation data arrangement and experimental data collection:

[0135] Through the icing simulation method in step S200, 1000 blade types are sequentially simulated under 50 working conditions, i.e. 50000 simulation examples. The blade types and working conditions are selected by the orthogonal experiment method. Finally, an icing simulation database is constructed, and in the subsequent steps, the simulation data is divided into an 80% training set (40000 groups) and a 20% test set (10000 groups) according to an 8:2 ratio.

[0136] On the basis of the icing simulation database described above, further collection of blade icing experimental data, at least including several groups of typical icing experimental data collected from public literature and several groups of data measured through the rotating blade icing experiment carried out by the self-built ice wind tunnel platform, wherein:

[0137] Literature experimental data: A total of 50 groups of engine blade icing experimental data published in public literature were systematically collected, which came from multiple authoritative institutions, including but not limited to the wind tunnel test results conducted by well-known institutions such as NASA Lewis Research Center and Pratt & Whitney. These data comprehensively cover the ice type measurement values under different incoming flow conditions, specifically involving the icing conditions under various air flow velocities, temperatures, humidities and other environmental parameters, providing detailed and reliable experimental data support for training high-precision neural networks.

[0138] Independent experiment: In a small ice wind tunnel environment, independent experimental research on blade icing ice type was carried out. High-performance high-speed cameras (frame rate up to 1000 fps) and high-precision laser displacement sensors (measurement accuracy up to ±0.1 mm) were used to measure and record the ice type evolution during the blade icing process. Through a series of experiments, 50 groups of detailed experimental data were obtained, covering the icing thickness of the blade leading edge, the specific range of icing, etc. It should be noted that the compressor blade type used in the experiment fully covers the blade type design parameter range required in the icing simulation process, ensuring the wide applicability of the experimental results. At the same time, the simulated incoming flow conditions in the experiment fully cover various working condition parameters involved in the icing simulation, thereby providing a solid data support for subsequent machine learning training.

[0139] S302. Multi-source icing database construction: The blade icing simulation data, literature experimental data and independent experimental data are subjected to data cleaning, format conversion and normalization processing, the experimental data and simulation data are fused, and a multi-source icing database with unified data structure and parameter dimension is formed. Specifically, step S302 mainly includes the following steps when performing data normalization processing:

[0140] Data cleaning: Through careful review and selection, incomplete or abnormal data records are removed, thereby excluding noise data that may affect the training effect of the model.

[0141] Format conversion: The task is to ensure that data from different sources can be seamlessly connected. Given the diversity of data sources and the difference in formats, the data is unified to a standard format to eliminate problems that may be caused by inconsistent data formats.

[0142] Normalization: The main purpose is to deal with the dimension and value range differences of working condition parameters (such as temperature, speed, liquid water content, etc.), blade profile parameters (such as maximum camber, maximum thickness, installation angle, inlet / outlet geometric angle, etc.), and ice shape image pixel values. To avoid affecting the convergence due to large value differences during model training, the maximum and minimum normalization method is used. This method scales all input parameters to the range of 0 to 1, making the model training more stable. The incoming flow parameters and blade profile parameters are scaled to [1, 1] by standardization (ZScore), and the ice shape image pixel values are normalized to [0, 1] by MinMax, and the data format is unified (CSV file).

[0143] S303. Data set segmentation (training set and test set division): According to the principle of multi-dimensional uniformity of icing type, blade profile scheme and working condition parameters, the simulation data set is divided into 80% training set (40000 groups of data) and 20% test set (10000 groups of data) by stratified sampling strategy.

[0144] Specifically, the division process not only considers the distribution of different blade profile structures (such as geometric angle, maximum thickness position, installation angle, etc.) and incoming flow working condition parameters (such as speed, temperature, droplet size, etc.) in the sample, but also takes into account the coverage ratio of different icing types (such as frost, clear ice, mixed ice) in the sample, so as to ensure that each subset after division is representative in sample type and statistical characteristics. At the same time, part of the data can be reserved as a validation set for model parameter tuning and performance optimization. This validation set is also constructed according to the stratified sampling principle to ensure that the whole process data loop of training-validation-testing has comparability and systematicness.

[0145] S304. Training group segmentation: The training set is divided into multiple training groups according to the preset group size, each training group contains an equal amount of simulation data, and then literature experimental data and self-developed experimental data are added to each training group according to the equal proportion principle, so that each training group is composed of simulation data and experimental data. Specifically, in the above training set, 100 simulation data per group are divided into 400 training groups, and 50 groups of experimental data from literature and 50 groups of self-developed experimental data are added to each training group. That is, each training group contains 100 groups of simulation data + 100 groups of experimental data (mixed literature data and self-developed experimental data). This segmentation ratio helps to ensure that the model learns both simulation rules and real experimental characteristics during training, and verifies its generalization ability on the test set.

[0146] It should be noted that in the construction of the data support system for the icing prediction modeling of the rotating blades of the aero-engine, the present step breaks through the traditional simulation and experimental data separate use, data dimension non-uniformity, and insufficient physical reality of training samples, and has significant technical innovation, which is embodied in the following aspects: (1) a unified database is constructed by fusing multi-source icing data. The present application fuses simulation-literature experiment-self experiment three source data, ensures the physical accuracy, greatly improves the sample diversity and universality, and significantly improves the adaptability of the icing prediction model to complex flight conditions. (2) The training group is designed with a reasonable simulation-experiment ratio mixing strategy. This strategy not only improves the fitting ability of the model to the real ice type characteristics, but also effectively avoids the adverse effects of experimental data scarcity on the training process, which is a key link to ensure the generalization ability and prediction reliability of the model. (3) The data set division method of hierarchical sampling and multi-dimensional uniformity principle. This strategy ensures the statistical balance of training set and test set in blade type scheme, working condition parameters and icing type and other key variable dimensions, improves the representativeness of training data, and enhances the prediction ability of the model to boundary conditions and complex ice types.

[0147] S400. Icing prediction machine learning model construction:

[0148] An icing prediction machine learning model is established, which includes an input layer, a hidden layer and an output layer. The input layer is used to receive a multi-dimensional feature vector including blade type parameters and incoming flow condition parameters, the hidden layer is used to perform nonlinear mapping on the input features to extract abstract features, and the abstract features are restored to the spatial dimension of the ice shape image, and the output layer is used to generate a blade surface icing prediction image corresponding to the input parameters.

[0149] In the embodiment of the present application, the neural network of the machine learning model adopts the architecture combined with Multi-Layer Perceptron (MLP) and Deconvolutional Neural Network (DNN), which includes an input layer, a hidden layer and an output layer, as shown in Figure 9 .

[0150] Input layer: designed to receive incoming flow conditions and blade type parameters, input data is transmitted to the network in the form of a multi-dimensional numerical vector, which is the basis for model perception and representation. Specifically, in the present embodiment, the input layer receives a 15-dimensional parameter vector, including 8 blade type parameters (inlet geometric angle , outlet geometric angle , installation angle , solidity σ , relative maximum camber position x mc , maximum turning angle h mc, relative maximum thickness position x mth , maximum thickness r mth ), and 5 inflow parameters (inflow velocity v , inflow temperature T, liquid water content LWC, droplet size MVD, respectively).

[0151] The hidden layer is designed to consist of three fully connected layers and three deconvolution layers, wherein: the fully connected layer constitutes the main structure of the MLP network: the three fully connected layers contain 256, 128 and 64 neurons respectively, all using ReLU activation function, for abstract expression and nonlinear transformation of input features layer by layer, to realize nonlinear mapping of extracted parameter features. In order to improve the generalization performance of the model, batch normalization and dropout technology are added after each fully connected layer. Batch normalization technology adjusts the distribution of each small batch of data so that its mean is 0 and its variance is 1, to speed up the training process and enhance the stability of the model. The dropout technique randomly excludes the output of some neurons during training to avoid model overfitting. The deconvolution layer is a DNN module integrated on the main structure of the MLP network: the three deconvolution layers (convolution kernel size is 4x4, 4x4, 4x4 respectively, step is 2, 2, 1 respectively), each layer is connected with ReLU activation function, for restoring abstract features to the spatial dimension of ice shape image.

[0152] Output layer: designed to generate predicted ice shape image under the condition of corresponding input inflow working condition and blade profile parameter through convolution operation, at least containing the spatial distribution information of ice layer thickness on the blade surface. The output layer in this embodiment generates a 128x128 pixel single-channel grayscale image through a 1x1 convolution layer, and the pixel value corresponds to the ice layer thickness of each point on the blade surface (normalized to [0, 1]).

[0153] S500. Training and verification of icing prediction machine learning model:

[0154] Based on the constructed training set, the training of the icing prediction machine learning model is carried out, the weighted mean square error loss function is used to guide the model to preferentially fit the true ice shape feature, and the deep learning framework is used for iterative training. After training, the generalization ability of the model is evaluated by using the test set. Specifically:

[0155] In this embodiment, the model building and training are realized based on the PyTorch deep learning framework. The Adam optimizer is used, the initial learning rate is set to 1x10 -4 , and it is decayed by 0.5 times every 50 rounds. And the weighted mean square error WMSE = a x MSE sim + b x MSE exp is used as the loss function, wherein a is the simulation error MSE simweight of the simulation data, and β is the weight of the experimental error MSE exp weight of the simulation data, and β > α, by giving higher weight to the experimental error, the model is forced to prioritize fitting the real experimental data to correct the precision bias of the simulation data. For example, set the weight of MSE sim to 0.4, and the weight of MSE exp to 0.6 to force the model to prioritize fitting the real experimental data.

[0156] In addition, in order to train the neural network model, the back propagation algorithm and the gradient descent optimizer are used. By constantly iterating the model parameters, the loss function value between the predicted result and the real icing shape is minimized. In order to speed up the training speed and improve the model performance, learning rate decay, Dropout and early stopping strategies are also adopted.

[0157] The model training process is executed under GPU acceleration, which includes the following steps: 1) data loading: the preprocessed training data is loaded into memory in batches for model training. Each batch contains 100 sets of simulation data and 100 sets of experimental data combination. 2) forward propagation: input data is sent into the model, and after feature extraction by fully connected layer, ice shape image prediction value is generated by deconvolution layer. 3) loss calculation: according to the difference between the predicted ice shape image and the real ice shape image, the weighted mean square error of the predicted image and the real label (simulation / experimental ice shape data) is calculated. 4) back propagation: according to the loss value, the model weights and bias parameters are updated by gradient descent method to minimize the loss value. 5) model saving: at the end of each iteration period, the current optimal model weight is saved for subsequent evaluation and use. During the training process, visualization tools such as TensorBoard are used to monitor the training process, including the trend of loss value, the update of model weight, etc., so as to adjust the training strategy and hyperparameters in time.

[0158] Regarding the training monitoring and optimization of the model. Real-time monitoring: track the loss value trend of the training set and the validation set, neuron activation distribution, weight update amplitude, etc., to judge whether the model is overfitting or underfitting. Hyperparameter adjustment: if the validation set loss does not decrease within 20 rounds, trigger the early stopping mechanism; adjust the number of neurons in the fully connected layer, the step length of the deconvolution layer, the learning rate decay strategy and other parameters until the model converges.

[0159] Data augmentation: apply random noise (simulate measurement error) and rotation and translation transformation (expand ice shape diversity) to the simulation data to improve the model's generalization ability. Figure 10 The training loss monitoring chart is shown.

[0160] The training cycle of the model. The total number of iterations is set to 1000 rounds, and the single training time is about 12 hours (based on NVIDIA RTX3090 GPU), and finally the prediction accuracy of the simulation data MSE≤0.01, experimental data MSE≤0.02 is realized on the test set.

[0161] After training, the machine learning model for predicting icing of rotating blades of an aero-engine is obtained. Based on the model, the icing model under any combination of blade type parameters and working condition parameters can be predicted. Figure 11 The verification results of ice shapes under different working conditions are presented, in which the red curve represents the actual data and the green curve represents the predicted data. By comparing the shape of the predicted ice shape and the actual ice shape, it can be observed that the prediction ability of the model is relatively accurate, thereby verifying the excellent generalization ability and practical application value of the model in predicting icing.

[0162] S600. Prediction of icing of rotating blades of an aero-engine and result analysis:

[0163] The trained icing prediction model is used to predict the icing of rotating blades of an aero-engine. In the embodiment of the present application, the icing prediction of the blades according to the blade type and working condition parameters includes the following sub-steps:

[0164] Input parameter preprocessing. The blade type parameters (inlet geometric angle , outlet geometric angle , installation angle , solidity σ , relative maximum camber position x mc , maximum turning angle h mc , relative maximum thickness position x mth , maximum thickness r mth ) and incoming flow parameters (incoming flow velocity v , incoming flow temperature T, liquid water content LWC, droplet size MVD) of the working condition to be predicted are normalized and converted into a high-dimensional input vector.

[0165] Ice shape image generation. The input vector is imported into the trained icing prediction model, and a 128x128 ice shape image is output through the deconvolution layer, and the pixel value is mapped to the ice layer thickness.

[0166] Post-processing and visualization. Feature extraction: extract the ice type thickness distribution (front edge ratio, pressure surface / absorbing surface distribution) under different icing positions, calculate the maximum thickness, coverage area and other indicators. Risk assessment: based on the icing thickness and position, combined with the engine operation safety standard, output the risk level. Visualization: generate a two-dimensional ice shape profile or a three-dimensional ice type model, and superimpose the blade geometry to display the icing shape.

[0167] The objects of the present application are completely achieved by the above-described embodiments. It will be understood by those skilled in the art that the present application includes, but is not limited to, what is described in the above detailed description and the accompanying drawings. Although the present application has been described with respect to the presently preferred and other embodiments, it will be understood that the application is not limited to the disclosed embodiments, and modifications can be made by those skilled in the art without departing from the spirit and principles of the application.

Claims

1. A method for predicting icing on rotating blades of an aeroengine based on machine learning, the method comprising: At least comprising the following steps: S100. Establishing an aero-engine rotating blade parameterized geometry model based on blade design variables and blade design constants; designing multiple levels of blade design variables based on an orthogonal experimental method, and combining with blade design constants to generate multiple sets of blade profile schemes; S200. For each blade profile scheme, multiple levels of incoming flow condition parameters are valued based on an orthogonal experimental method, and multiple sets of incoming flow condition parameter combinations are designed to generate, each of which corresponds to an icing simulation example; blade icing simulation calculation is performed for each example to obtain corresponding blade surface ice shape data, and an icing simulation database is constructed accordingly; S300. Collecting experimental data of aero-engine rotating blade icing, and uniformly processing the experimental data with blade icing simulation data to construct a multi-source icing database integrating experimental data and simulation data, and dividing the multi-source database into a training set and a test set; S400. Establishing an icing prediction machine learning model including an input layer, a hidden layer and an output layer, the input layer being used to receive a multi-dimensional feature vector including blade profile parameters and incoming flow condition parameters, the hidden layer being used to perform nonlinear mapping on the input features to extract abstract features, and the output layer being used to generate a blade surface icing prediction image; S500. Training the icing prediction machine learning model based on the training set, using a weighted mean square error loss function to guide the model to preferentially fit the real ice shape features, and combining a deep learning framework to perform iterative training, and evaluating the generalization ability of the model using the test set after training is completed; S600. Using the trained icing prediction model to predict the icing of the rotating blade.

2. The machine learning based aeroengine rotating blade icing prediction method of claim 1, wherein, In step S100, the blade parameterized modeling at least includes the following sub-steps: S111. determining blade design variables including at least an inlet geometric angle , an outlet geometric angle , a setting angle , a solidity σ , a relative maximum camber position x mc , a maximum turning angle h mc , a relative maximum thickness position x mth , a maximum thickness r mth , each of the design variables being used to control the geometry and aerodynamic characteristics of the blade; S1 12. Determine blade design constants, including at least blade chord length c 1, camber line exit angle α 1, camber line 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 values in the blade parameterized modeling process; S113. Based on the set blade design variables and design constants, expression equations of the camber line and thickness distribution of the blade profile are constructed respectively, the expression equation of the camber line should ensure its continuity and smoothness in the full chord direction, and the expression equation of the thickness distribution should ensure the continuity and derivability of the pressure surface and suction surface profiles; S114. The blade profile geometry distribution is obtained by superimposing the camber line and thickness distribution, the coordinate points of the blade profile leading edge, trailing edge, pressure surface and suction surface are generated, and a blade profile parameterized geometry model is formed.

3. The machine learning based aeroengine rotating blade icing prediction method of claim 1 or 2, wherein, In step S100, when the orthogonal design method is used to design the batch of blade profile schemes, at least the following sub-steps are included: S121. Determining the level range of the design variables: based on the set blade design variables, combining the actual design requirements and application conditions of the aero-engine rotating blade, the value interval of each design variable in the target blade design space is determined to ensure that the design space covers the typical range of blade geometric feature changes; S122. Discretization processing of design variable levels: according to the value interval of the blade design variable, combining the sensitivity of the variable to the blade geometry shape and aerodynamic performance, the number of levels of the design variable is determined and the corresponding value interval is equally or unequally divided into multiple discrete level values; S123. Constructing orthogonal experiment design matrix: selecting an appropriate orthogonal table according to the number of design variables and the number of levels, systematically combining the discrete level values of each design variable according to the arrangement rules of the orthogonal table, and obtaining an orthogonal design matrix with the maximum factor combination coverage under the condition of limited samples; S124. Generating multiple sets of airfoil design schemes: combining each row of parameter combinations in the orthogonal design matrix with the preset airfoil design constants, generating corresponding airfoil geometries one by one through the constructed airfoil parameterized model, and the obtained airfoils should have clear boundary contours, thickness distribution and mid-arc line continuity; S125. Geometric verification and optimization of airfoil schemes: checking the geometric feasibility of each generated airfoil scheme, eliminating abnormal airfoils, and optimizing airfoils with slight geometric defects.

4. The machine learning based aeroengine rotating blade icing prediction method of claim 1, wherein, In step S200, for each airfoil scheme, the batch design of icing simulation examples based on the orthogonal design method at least includes the following sub-steps: S211. Determining the inflow condition parameters and their design range: according to the working environment of the rotating blade of the aero-engine, determining several inflow condition parameters that significantly affect the icing characteristics, and setting the value interval of each parameter combined with the icing history and / or test data; S212. Inflow condition parameter level design and discretization: based on the value interval of each inflow condition parameter and its influence sensitivity on icing characteristics, performing multi-level discretization processing on each parameter and determining its level value by using equidistant or non-equidistant division method; S213. Constructing an orthogonal design matrix of inflow condition parameters: using the orthogonal experimental design method, selecting an appropriate orthogonal table according to the design variables and their level numbers, and generating an orthogonal design matrix of parameter combinations with statistical representativeness and spatial uniformity; S214. Generating icing simulation example design schemes: combining each row of inflow condition parameter combinations in the orthogonal design matrix with the airfoil scheme to form complete icing simulation example designs, each example including clear blade geometry definition and inflow boundary condition setting.

5. The machine learning based aeroengine rotating blade icing prediction method of claim 1 or 4, wherein, In step S200, the construction of the icing simulation database at least includes the following sub-steps: S221. Fluid domain modeling and meshing: based on the airfoil scheme of the selected icing simulation example, constructing the blade passage fluid domain geometry, dividing the fluid domain using structured or unstructured mesh, and performing local mesh densification on the blade surface area to accurately capture the boundary layer flow characteristics; S222. Boundary condition setting: according to the inflow condition parameters corresponding to the icing simulation example, setting the boundary conditions of the calculation model accordingly, including setting the inflow velocity and 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 the periodic boundary condition according to the actual working state of the rotating blade to simulate the flow between blade rows; S223. Air flow field calculation: after completing meshing and boundary setting, the RANS equation is used to numerically solve the air flow in the rotating blade passage, and an appropriate turbulence model is selected for turbulence characteristic calculation, finally obtaining the air flow field results including velocity field, pressure field and temperature field; S224. Droplet motion and impingement calculation: based on the solution results of the air flow field, the droplet motion and impingement calculation in the rotating environment of the aero-engine is carried out using the Euler method and based on the one-way coupling strategy, the trajectory of the droplet in the flow field is tracked and the impingement characteristics of the droplet and the blade surface are calculated; S225. Icing thermodynamic process calculation: based on the calculation results of the droplet impingement and the temperature distribution of the blade surface, the icing thermodynamic model calculation of the blade surface is carried out, the simulation of different types of ice is enhanced by considering the water film movement, and the heat flux density distribution and phase change characteristics of each position of the blade surface are obtained; S226. Ice layer growth thickness calculation: based on the calculation results of the icing thermodynamic process, the ice layer growth rate is calculated by analyzing the heat transfer ability difference of the latent heat of ice at the ice-water interface in the water film and the ice layer on both sides of the interface, the ice layer thickness distribution in the clear ice icing mode and the frost ice icing mode is solved respectively, and the smaller value of the two is selected as the final icing thickness to ensure the accuracy of the icing type judgment; S227. Ice layer profile update and grid adaptive adjustment: a new ice layer profile is calculated according to the ice layer growth thickness, the blade geometry is adjusted based on the original blade geometry, and the calculation grid is adaptively updated based on the updated blade geometry to ensure that the grid quality meets the accuracy requirements of the subsequent time step simulation calculation; S228. Multi-time step icing evolution simulation: based on the updated blade geometry and grid, steps S223 to S227 are repeatedly executed while keeping the incoming flow condition parameters unchanged to realize the multi-time step icing dynamic evolution process simulation until the preset icing time or ice layer thickness convergence condition is reached, and finally the complete icing evolution history data is obtained; S229. Batch calculation of examples and database construction: after completing the whole process simulation of a simulation example, all preset icing simulation examples are traversed, steps S221 to S228 are repeatedly executed for simulation calculation, and key icing characteristic parameters are extracted from each simulation example to finally complete the construction of the icing simulation database covering various blade profile schemes and working conditions.

6. The machine learning based aeroengine rotating blade icing prediction method of claim 5, wherein, In step S224, the droplet motion and impingement calculation includes the following sub-steps: S2241. Establish droplet motion control equation: establish droplet motion control equation in rotating coordinate system, consider aerodynamic drag, gravity, centrifugal force and Coriolis force of droplet in rotating environment; S2242. Droplet phase initialization and boundary injection configuration: based on the solution results of the air flow field, the one-way coupling strategy is adopted, the liquid droplet release position and initial velocity condition are set at the inlet of the fluid domain according to the liquid water content and droplet particle size distribution in the incoming flow condition parameters without affecting the air flow state; S2243. Numerical solution of droplet trajectory: the Euler method is used to numerically integrate the droplet motion control equation, the velocity, position and force state of the droplet are updated according to the flow field information at the current position in each time step until the droplet leaves the calculation domain or impinges on the blade surface; S2244. Droplet impingement judgment and parameter calculation: Establish the judgment criterion of droplet impingement with the blade surface, determine the occurrence of impingement event when the droplet trajectory intersects with the blade surface geometry, extract the key impingement parameters of each impingement point, and finally form the impingement boundary condition field.

7. The machine learning based aeroengine rotating blade icing prediction method of claim 6, wherein, In step S225, the icing thermodynamic process calculation includes the following sub-steps: S2251. Icing control body construction: The first layer of near-wall grid of the flow field solving grid is used to construct the icing control body, which is divided into air layer, water film layer and ice layer from outside to inside, the air layer maintains heat exchange with the main flow field, the water film layer bears droplet impingement and water film flow process, and the ice layer records the icing accumulation process; S2252. Water film flow control equation establishment: The continuity equation, momentum equation and energy equation of water film flow are established in the control body, the continuity equation considers the mass source term of droplet impingement and the mass loss caused by phase change, the momentum equation considers the influence of air shear force, gravity, wall pressure gradient, centrifugal force and Coriolis force, and the energy equation considers the heat exchange process between water film and air layer and ice layer and the energy input caused by droplet impingement; S2253. Ice layer heat transfer and energy conservation calculation: One-dimensional steady-state heat transfer control equation is established in the ice layer, considering the heat conduction process and temperature distribution of the ice layer, the ice-wall interface adopts adiabatic boundary condition, and the ice-water interface temperature is constant at the icing phase change temperature, simulating the heat transfer path inside the ice layer and the heat exchange process between the wall and the water film layer, and providing boundary conditions for the phase change calculation at the ice-water interface by solving the temperature field inside the ice layer; S2254. Phase change behavior and ice type judgment mechanism: Based on the water film flow state, temperature field distribution and impingement flux characteristics, establish the judgment criteria of rime ice, clear ice and mixed ice, when the water film can flow on the surface and partially freeze, clear ice is formed, when the impinging droplet freezes instantly and there is no water film flow, rime ice is formed, and when the two mechanisms coexist, mixed ice is formed, and the thermal physical parameters and phase interface boundary conditions are adjusted based on the phase change mechanism of different ice types to realize the adaptive simulation of ice type; S2255. Heat flux distribution and phase change characteristic calculation: By solving the coupled thermodynamic equations of water film and ice layer in the control body, the heat flux distribution and phase change latent heat flux of each position on the blade surface are calculated, and the thermodynamic boundary conditions are provided for the subsequent ice layer growth calculation.

8. The machine learning based aeroengine rotating blade icing prediction method of claim 7, wherein, In step S226, the ice layer thickness calculation includes the following sub-steps: S2261. Ice layer growth rate control equation establishment: An energy conservation model is established around the ice-water interface, the release of icing latent heat and heat transfer process at the ice-water interface are analyzed, the temperature gradient difference in the water film and ice layer on both sides of the interface is considered, and the ice layer growth rate control equation of ice layer thickness change with time is established; S2262. Double mode result comparison and minimum value selection: The ice layer thickness distribution under the clear ice icing mode and the rime ice icing mode is calculated respectively, the numerical value of the two calculation results at each surface position is compared, and the smaller value of the two is selected as the final ice layer thickness at that position; S2263. Final ice type determination and thickness distribution output: the final ice type of each position on the blade surface is determined based on the minimum value selection principle. When the rime ice thickness is less than the glaze ice thickness, it is determined as rime ice. When the two are close, it is determined as mixed ice. Finally, the complete ice thickness distribution data and ice type distribution are output.

9. The machine learning based aeroengine rotating blade icing prediction method of claim 1, wherein, In step S300, the construction of the multi-source icing database and the division of the training set and the test set at least include the following sub-steps: S301. Experimental data collection: on the basis of the icing simulation database, further collect blade icing experimental data, at least including several groups of typical icing experimental data collected from public literature and several groups of data measured by rotating blade icing experiments carried out through the self-built icing wind tunnel platform; S302. Construction of multi-source icing database: data cleaning, format conversion and normalization processing are performed on the blade icing simulation data, literature experimental data and self-built experimental data, experimental data and simulation data are fused and a multi-source icing database with unified data structure and parameter dimension is formed; S303. Division of training set and test set: according to the multi-dimensional uniformity principle of icing type, blade type scheme and working condition parameters, the layered sampling strategy is adopted to divide the icing simulation database into training set and test set according to the set proportion, and part of the data is reserved as the validation set for model parameter adjustment and performance optimization; S304. Training group segmentation: the training set is segmented into multiple training groups according to the preset group size, each training group contains equal amount of simulation data, and then literature experimental data and self-built experimental data are added in each training group according to the equal proportion principle, so that each training group is composed of simulation data and experimental data.

10. The machine learning based aeroengine rotating blade icing prediction method of claim 1, wherein, In step S400, the icing prediction machine learning model adopts MLP architecture and is designed in combination with DNN network, which is composed of input layer, hidden layer and output layer, wherein: The input layer is designed to receive incoming flow conditions and blade type parameters, and the input data is transmitted to the network in the form of a multi-dimensional numerical vector as the basis for model perception and representation; The hidden layer is designed to be composed of three fully connected layers and three deconvolution layers, wherein the fully connected layer constitutes the main structure of the MLP network, which is used to abstractly express and nonlinearly transform the input features layer by layer, each fully connected layer contains a certain number of neurons and adopts ReLU activation function for nonlinear transformation; the deconvolution layer is a DNN module integrated on the main structure of the MLP network, which is used to decode the abstract features of the previous layer step by step into spatially distributed icing prediction results, realize the extraction and spatial reconstruction of the icing pattern features, and each deconvolution layer is also nonlinearly processed by ReLU activation function after it; The output layer is designed to generate predicted ice shape images under the condition of corresponding input incoming flow conditions and blade type parameters through convolution operation, which at least includes the spatial distribution information of the blade surface ice thickness.

11. The machine learning based aeroengine rotating blade icing prediction method of claim 1, wherein, In step S500, training of the icing prediction machine learning model is carried out based on the training set, each batch of training data is set to a fixed number of simulation samples and experimental samples combined, and a weighted mean square error WMSE = a x MSE + b x MSE is used sim exp as a loss function, where a is the weight of the simulation error MSE sim , b is the weight of the experimental error MSE exp , and b > a;​ The model is iteratively trained using a deep learning framework, combined with learning rate dynamic adjustment, batch normalization, Dropout, and early stopping mechanism to improve training convergence efficiency and stability. The training process includes an iterative loop of data loading, forward propagation, loss calculation, back propagation, and model parameter updating; Real-time monitoring of the training process, including loss value trend, model weight update, and neuron activation distribution. The early stopping mechanism prevents overfitting, and the optimal model weight is saved when the validation set loss does not decrease within the preset period. Hyperparameters are optimized, including adjusting the number of fully connected layer neurons, deconvolution layer step size, and learning rate decay strategy, until the model converges; After training, the test set is used to evaluate the model's ice prediction ability on non-participating training data. The prediction error level of ice type, thickness distribution, area ratio, and image similarity is quantitatively compared with the true ice type to verify the model's generalization ability and prediction accuracy.

12. The machine learning based aeroengine rotating blade icing prediction method of claim 1, wherein, In step S600, the trained ice prediction model is used to predict the icing of the rotating blades of an aero-engine, including at least the following sub-steps: S601. Prepare input parameters, including blade type parameters, incoming flow condition parameters, and engine operating parameters, as the input basis for the ice prediction model; S602. Preprocess the input parameters according to the unified format and dimension during training, and use the same normalization method as in the training phase; S603. Input the preprocessed input parameters into the ice prediction model, perform feature extraction and nonlinear mapping through the hidden layer, and generate the corresponding blade ice shape image from the output layer; S604. Post-process the output ice shape image, extract the ice layer thickness distribution and icing location, and quantify the icing degree; S605. Present the icing location and shape on the blade surface, and output the icing degree and risk level information.

Images