PINN-based method and system for predicting erosion damage of aircraft engine compressor blade

Abstract

The invention belongs to the technical field of aircraft engine compressor blade detection, and discloses an aircraft engine compressor blade erosion damage prediction method and system based on a PINN. The method comprises the following steps: constructing an erosion damage rate equation and a flow field-particulate matter coupling equation aiming at the erosion damage judgment of the blade of the gas compressor; performing dual-drive loss function design, constructing a physical information neural network (PINN) model, and performing adaptive training on the constructed physical information neural network (PINN) model; the invention relates to gas compressor blade erosion damage prediction and engineering application. According to the method, the limitation of the prior art is overcome, a more accurate and reliable damage prediction model is provided, and powerful support is provided for design, operation and maintenance and safety evaluation of the aero-engine. According to the innovative method, the physical law and the data driving technology can be combined, and the damage prediction precision and efficiency of the aero-engine compressor blade under the complex working condition are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of aero-engine compressor blade detection, and particularly relates to a method and system for predicting erosion damage of aero-engine compressor blades based on PINN. Background Art

[0002] With the rapid development of aero-engine technology, compressor blades play a crucial role in aero-engines. The main function of compressor blades is to compress air for use in the combustion chamber. During engine operation, they are subjected to extremely high air flow velocities, temperature variations, and particulate impacts from the external environment. Due to these factors, compressor blades are prone to various damages including cavitation, corrosion, wear, and erosion. These damages not only affect the aerodynamic performance of the blades but also lead to a decrease in the structural strength of the blades, and in severe cases, may even cause failures in the entire engine.

[0003] Especially in sandstorms, dusty environments, or other airflows containing solid particles, the erosion damage problem of compressor blades is particularly prominent. Erosion damage is caused by the collision of particulate matter (such as sand, hail, rain, etc.) in high-speed airflows with the blade surface. It results in microscopic damage to the blade surface, which gradually forms larger breakages over time, thereby affecting the overall performance of the blades. Therefore, predicting and evaluating the erosion damage of compressor blades is crucial for engine design, maintenance, and safety assurance.

[0004] Traditional blade damage prediction methods mainly rely on experimental data, classical physical models, and finite element analysis methods. These methods generally require a large amount of computational resources and experimental data, and are difficult to provide accurate prediction results when facing complex working conditions (such as high-speed airflows, uneven distribution of particulate matter, etc.). For example, the finite element analysis method can simulate the stress and deformation of blades under specific working conditions, but for the prediction of erosion damage under the combined action of multiple factors, its accuracy and computational efficiency are often limited. With the increasing complexity of the operating environment of aero-engines, traditional methods face greater challenges in predicting blade erosion damage.

[0005] In recent years, technologies based on artificial intelligence and machine learning have gradually been introduced into the field of engineering prediction. In particular, the successful application of deep learning algorithms in fields such as image recognition and speech processing has stimulated their potential in engineering prediction. As a new type of deep learning technology, the Physics-Informed Neural Network (PINN) combines the prior knowledge of traditional physical models and data-driven machine learning methods, and can perform accurate damage prediction by combining with physical laws without completely relying on large-scale experimental data. The advantage of PINN is that it can force the output of the neural network to satisfy physical constraints during the training process, which makes it show high accuracy and generalization ability when dealing with engineering problems with complex physical backgrounds.

[0006] However, for the prediction of erosion damage of aero-engine compressor blades, the current research is still in the exploratory stage, and the relevant research results are relatively limited. The existing damage prediction methods based on machine learning usually cannot fully combine the complex flow field and particle characteristics of compressor blades under actual working conditions, nor take into account the precise constraints of physical laws. Due to the large differences in blade damage modes under different working conditions, it is difficult for the existing technologies to achieve efficient and accurate erosion damage prediction. Summary of the Invention

[0007] To overcome the problems existing in the related technologies, the disclosed embodiments of the present invention provide a method and system for predicting erosion damage of aero-engine compressor blades based on PINN. Specifically, it relates to a method based on the Physics-Informed Neural Network (PINN) for predicting the erosion damage of aero-engine compressor blades. This method combines the advantages of computational fluid dynamics (CFD) and machine learning, and is applicable to the damage assessment and prediction of aero-engine compressor blades under complex working conditions.

[0008] The technical solution is as follows: A method for predicting erosion damage of aero-engine compressor blades based on PINN includes:

[0009] S1, Physical mechanism and mathematical modeling. According to the dynamic process of the interaction between the kinetic energy of particles and the material surface, based on fluid mechanics and erosion dynamics, an erosion damage rate equation and a flow field-particle coupling equation are constructed for the discrimination of compressor blade erosion damage.

[0010] S2, Based on the constructed erosion damage rate equation and flow field-particle coupling equation, a dual-driven loss function is designed. The dual-driven loss function L integrates the data-driven term and the physical constraint term to ensure that the experimental data is fitted and the conservation law is obeyed simultaneously.

[0011] S3. Based on the constructed erosion damage rate equation, the fluid field-particle coupling equation, and the designed dual-drive loss function, construct a physics-informed neural network PINN model and perform adaptive training on the constructed physics-informed neural network PINN model;

[0012] S4. Compressor blade erosion damage prediction and engineering applications.

[0013] In step S1, the erosion damage rate equation includes that the damage rate D is jointly determined by the particle collision characteristics, fluid parameters, and material properties. A modified erosion damage rate equation is introduced, and its expression is:

[0014]

[0015] In the formula, D is the damage rate, α and β are material erosion coefficients, ρ and v are the gas density and gas flow velocity, d is the particle diameter, θ is the particle incidence angle, and η is the collision efficiency.

[0016] In step S1, the fluid field-particle coupling equation includes describing the fluid field characteristics through the Navier-Stokes equation and introducing the particle momentum exchange term. The expression is:

[0017]

[0018] In the formula, is the change rate of the velocity field with time, u is the velocity, is the Nabla operator, is the pressure gradient, is the Laplace operator, F p is the force exerted by discrete particles on the fluid;

[0019] The force F p exerted by discrete particles on the fluid has the following expression:

[0020]

[0021] In the formula, is the acceleration of the particle, i is the particle index identifier, m p , v p are the mass and velocity of the particle respectively, and t is the running time.

[0022] In step S2, the expression of the conservation law is:

[0023] L = λ1L data + λ2l phys (3)

[0024] In the formula, L is the dual-drive loss function, λ1 and λ2 are the weights corresponding to the data loss and physical loss, Ldata For data loss, L phys is the physical loss;

[0025] The data-driven term is used to minimize the mean square error between the neural network prediction value and the experimental / CFD data as follows:

[0026]

[0027] where N is the number of experimental data samples, D PINN (x i ) is the predicted value of the pinn model, D exp (x i ) is the value measured experimentally, |·| 2 is the squared error; x i is the input parameter, including flow velocity, particulate matter properties, and geometric parameters.

[0028] In step S2, the physical constraint term is used to force the output of the physics-informed neural network to satisfy equation (5):

[0029]

[0030] where M is the number of samples, j is the sample index identifier, is the rate of change of the damage rate D with respect to time t; is the convection term, representing the advection change of the damage rate D under the action of the velocity field u; S is the erosion source term, ||·|| 2 is the Euclidean norm;

[0031] The expression for the erosion source term D is:

[0032]

[0033] where D is the damage rate, α, β are the material erosion coefficients, ρ, v are the gas density and gas flow velocity, d is the particulate matter diameter, θ is the particulate matter incidence angle, and η is the collision efficiency.

[0034] In step S3, the physics-informed neural network PINN model includes:

[0035] Input layer: 9-dimensional parameters, gas flow velocity v, pressure p, particulate matter velocity vp, particle size d, density ρp, incidence angle θ, surface roughness Ra, material hardness H, operating time t;

[0036] Hidden layer: 4 fully connected layers, with 13 nodes in each layer, and the Swish function is selected as the activation function to balance non-linearity and gradient stability;

[0037] Output layer: Erosion rate D and its spatial gradient for physical residual calculation.

[0038] In step S3, the adaptive training includes dynamic weight adjustment and enhanced sampling in key areas;

[0039] The dynamic weight adjustment includes: the initial weights, λ1 = 1.0, λ2 = 0.1, and every 1000 iterations, λ2 is increased by 10% to increase the physical constraint strength and avoid early overfitting;

[0040] The enhanced sampling in key areas includes: in high-damage areas such as the leading edge of the blade, the sampling distribution is controlled by the probability density function PDF:

[0041]

[0042] In the formula, p(x) is the probability density function, x edge is the coordinate of the leading edge of the blade, and ∝ represents being directly proportional.

[0043] In step S4, the prediction of the erosion damage of the compressor blade includes:

[0044] Input preprocessing, mapping the sensor data including flow velocity and particulate concentration into network input parameters;

[0045] Real-time inference, the PINN model outputs the damage rate D and calculates the cumulative damage by integration; the calculation steps are as follows:

[0046] Discretization processing: discretize the sensor data into D(t1), D(t2)…D(tn) according to the time step Δt;

[0047] Numerical integration: calculate using the trapezoidal rule or Simpson's rule:

[0048]

[0049] In the formula, D total is the total damage rate, D(t k ) is the damage rate at time t k k-1 k-1 ) is the damage rate at time t k-1 k-1

[0050] Another object of the present invention is to provide a prediction system for the erosion damage of aero-engine compressor blades based on PINN. This system implements the prediction method for the erosion damage of aero-engine compressor blades based on PINN. This system includes:

[0051] The physical mechanism and mathematical modeling module is used to construct an erosion damage rate equation and a flow field - particulate coupling equation for the discrimination of the erosion damage of the compressor blade based on the dynamic process of the interaction between the kinetic energy of the particulate matter and the material surface, based on fluid mechanics and erosion dynamics;

[0052] Dual - drive loss function design module, based on the constructed erosion damage rate equation and fluid - field particulate matter coupling equation, designs a dual - drive loss function. The dual - drive loss function L integrates a data - driven term and a physical constraint term to ensure fitting experimental data while complying with the conservation law;

[0053] Physical - informed neural network model construction and adaptive training module, based on the constructed erosion damage rate equation, fluid - field particulate matter coupling equation, and the designed dual - drive loss function, constructs a physical - informed neural network PINN model and conducts adaptive training on the constructed physical - informed neural network PINN model;

[0054] Erosion damage prediction module, used for erosion damage prediction and engineering applications of compressor blades.

[0055] Furthermore, the PINN - based aero - engine compressor blade erosion damage prediction system is installed on a computer device. The computer device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, it realizes the functions in the above - mentioned PINN - based aero - engine compressor blade erosion damage prediction system.

[0056] Combining all the above - mentioned technical solutions, the beneficial effects of the present invention are as follows: The present invention proposes a method for predicting the erosion damage of aero - engine compressor blades based on PINN, overcomes the limitations of the prior art, provides a more accurate and reliable damage prediction model, and provides strong support for the design, operation and maintenance, and safety assessment of aero - engines. The present invention can combine the innovative method of physical laws and data - driven technology, improving the damage prediction accuracy and efficiency of aero - engine compressor blades under complex working conditions.

[0057] The present invention has good expected benefits and commercial value: Reduction in operation and maintenance costs: By improving the prediction accuracy (error < 5%), the unplanned shutdown is reduced by more than 60%. It is expected that the annual maintenance cost of a single engine will be reduced by about 1.2 million yuan (calculated based on the operation and maintenance data of CFM56 engines). Extension of the life cycle: The blade replacement cycle is extended from 4000 hours to 6500 hours. Calculated at a rental rate of 80,000 yuan per hour, the annual income increase per unit can reach 20 million yuan. Market share: The technology can cover 42,000 in - service civil aviation engines globally (ICAO data in 2023). Calculated at a 10% penetration rate, the potential market size exceeds 50 billion yuan per year.

[0058] The present invention is based on multi - physical - field coupling modeling: for the first time, unsteady aerodynamic load - particle erosion - material fatigue coupling modeling is carried out, breaking through the limitations of traditional single - discipline analysis. Small - sample transfer learning: a spatio - temporal feature decoupling transfer algorithm is proposed, and the prediction accuracy remains 92% even when the training data is less than 200 groups (existing technologies such as CNN - LSTM require more than 1000 groups of data).

[0059] The present invention has achieved breakthroughs in technical problems: Optimization of conflicting indicators: overcome the trade - off problem of "prediction accuracy - calculation efficiency", while maintaining the accuracy of CFD, shorten the calculation time from hours to minutes (6.2 hours → 8 minutes). Cross - operating - condition generalization ability: through the physical - information regularization layer, solve the failure problem of traditional data - driven models under variable rotational speed and variable angle of attack. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The drawings herein are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure;

[0061] Figure 1 is a flowchart of a method for predicting erosion damage of aero - engine compressor blades based on PINN provided by an embodiment of the present invention;

[0062] Figure 2 is provided by an embodiment of the present invention Figure 2 schematic diagram of the architecture of a physics - informed neural network (PINN) model;

[0063] Figure 3 is a schematic diagram of a system for predicting erosion damage of aero - engine compressor blades based on PINN provided by an embodiment of the present invention;

[0064] Figure 4 is a schematic diagram of the principle of a method for predicting erosion damage of aero - engine compressor blades based on PINN provided by an embodiment of the present invention;

[0065] In the figure: 1. Physical mechanism and mathematical modeling module; 2. Dual - drive loss function design module; 3. Physics - informed neural network model construction and adaptive training module; 4. Erosion damage prediction module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] To make the above - mentioned objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention is provided in conjunction with the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0067] The innovation of the present invention lies in: the present invention proposes a method and system for predicting the erosion damage of aero-engine compressor blades based on a physics-informed neural network (PINN). By integrating the unsteady flow field control equation and multi-source sensor data, the present invention constructs a spatio-temporal dual-driven physics-informed constrained transfer learning framework, realizing the multi-physics field coupling modeling of aerodynamic load - particle erosion - fatigue damage. This method achieves a cross-condition prediction accuracy of 95.3% under the condition of only 200 sets of samples, and the calculation time is reduced by 83% compared with the traditional CFD method. It can output the damage rate and cumulative damage amount in real time, and trigger maintenance instructions through a dynamic threshold warning mechanism. Experiments show that the prediction error of this method is less than 5% under the conditions of variable rotational speed (50% - 110%) and variable angle of attack (±15°), and the false alarm rate is reduced by 63% compared with the traditional LSTM model, significantly improving the reliability of blade life prediction and the economy of maintenance.

[0068] Example 1, the present invention provides a method based on a physics-informed neural network (PINN) for predicting the erosion damage of aero-engine compressor blades in high-speed airflow. The core idea of this method is to use the PINN framework to combine the physical laws of fluid mechanics and erosion damage mechanism during the neural network training process, accurately predicting the erosion damage behavior of the blades under complex working conditions. By combining the physics equation with the data-driven model, it avoids the dependence on a large amount of experimental data and computing resources in the traditional method, improving the prediction accuracy and efficiency.

[0069] Specifically, as Figure 1 shown, the method for predicting the erosion damage of aero-engine compressor blades based on PINN provided by the embodiment of the present invention includes:

[0070] S1, physical mechanism and mathematical modeling. According to the dynamic process of the interaction between the kinetic energy of particulate matter and the material surface, based on fluid mechanics and erosion dynamics, for the discrimination of the erosion damage of compressor blades, an erosion damage rate equation and a flow field - particulate matter coupling equation are constructed;

[0071] Among them, the erosion damage rate equation includes: the damage rate D is jointly determined by the collision characteristics of particulate matter, fluid parameters, and material properties. A modified erosion damage rate equation is introduced, and the expression is:

[0072]

[0073] In the formula, D is the damage rate, α and β are material erosion coefficients (calibrated by experiments), ρ and v are the gas density and air flow velocity, d is the particulate matter diameter, θ is the incident angle of particulate matter (relative to the blade normal), and η is the collision efficiency (related to the surface roughness).

[0074] The flow field-particle coupling equation includes: describing the flow field characteristics through the Navier-Stokes equation and introducing the particle momentum exchange term, and the expression is:

[0075]

[0076] In the formula, is the rate of change of the velocity field with time, u is the velocity, is the Nabla operator, is the pressure gradient, is the Laplace operator, F p is the force exerted by discrete particles on the fluid;

[0077] The force F exerted by discrete particles on the fluid p has the following expression:

[0078]

[0079] In the formula, is the acceleration of the particle, i is the particle index identifier, m p , v p are the mass and velocity of the particulate matter respectively, and t is the running time.

[0080] S2. Based on the constructed erosion damage rate equation and the flow field-particle coupling equation, a dual-driven loss function is designed. The dual-driven loss function L integrates the data-driven term and the physical constraint term to ensure simultaneous fitting of experimental data and compliance with the conservation law;

[0081] The expression of the conservation law is:

[0082] L = λ1L data + λ2L phys (3)

[0083] In the formula, L is the dual-driven loss function, λ1 and λ2 are the weights corresponding to the data loss and the physical loss, L data is the data loss, L phys is the physical loss;

[0084] The data-driven term is used to minimize the mean square error between the neural network prediction value and the experimental / CFD data as:

[0085]

[0086] In the formula, N is the number of experimental data samples, D PINN (x i ) is the prediction value of the pinn model, D exp (x i ) is the value measured experimentally, |·| 2 is the squared error; xi is the input parameter, including flow rate, particulate matter properties, and geometric parameters.

[0087] The physical constraint term is used to force the output of the physics-informed neural network to satisfy Equation (5):

[0088]

[0089] where M is the number of samples, and j is the sample index identifier. is the rate of change of the damage rate D with respect to time t; is the convection term, representing the advection change of the damage rate D under the action of the velocity field u; S is the erosion source term, ||·|| 2 is the Euclidean norm;

[0090] The expression of the erosion source term D is:

[0091]

[0092] where D is the damage rate, α and β are material erosion coefficients, ρ and v are the gas density and gas flow velocity, d is the particulate matter diameter, θ is the particulate matter incident angle, and η is the collision efficiency.

[0093] S3. Based on the constructed erosion damage rate equation and the flow field - particulate matter coupling equation, as well as the designed dual - drive loss function, construct a physics - informed neural network PINN model, and perform adaptive training on the constructed physics - informed neural network PINN model;

[0094] Such as Figure 2 Schematic diagram of the architecture of the physics - informed neural network (PINN) model; The physics - informed neural network (PINN) model includes:

[0095] Input layer: 9 - dimensional parameters, including gas flow velocity v, pressure p, particulate matter velocity vp, particle size d, density ρp, incident angle θ, surface roughness Ra, material hardness H, and operating time t;

[0096] Hidden layer: 4 - layer fully connected, with 13 nodes in each layer, and the Swish function is selected as the activation function to balance nonlinearity and gradient stability;

[0097] Output layer: erosion rate D and its spatial gradient , which is used for physical residual calculation.

[0098] Adaptive training includes dynamic weight adjustment and enhanced sampling in key regions;

[0099] Dynamic weight adjustment includes: initial weights, λ1 = 1.0, λ2 = 0.1, and every 1000 iterations, λ2 is increased by 10% to increase the physical constraint strength and avoid early overfitting;

[0100] Enhanced sampling in key areas, including: in areas with high damage such as the leading edge of the blade, using the probability density function PDF to control the sampling distribution:

[0101]

[0102] In the formula, p(x) is the probability density function, and x edge is the coordinate of the leading edge of the blade, and ∝ represents being directly proportional.

[0103] S4. Prediction and engineering application of compressor blade erosion damage.

[0104] The prediction process includes:

[0105] Input preprocessing: Mapping sensor data (flow rate, particle concentration, etc.) to network input parameters.

[0106] Real-time inference. The PINN model outputs the damage rate D and integrates to calculate the cumulative damage amount. The calculation steps are as follows:

[0107] Discretization processing: Discretizing the sensor data into D(t1), D(t2)…D(tn) at time step Δt;

[0108] Numerical integration: Calculating using the trapezoidal rule or Simpson's rule:

[0109]

[0110] In the formula, D total is the total damage rate, D(t k ) is the damage rate at time t k k-1 k-1 ) is the damage rate at time t k-1 k-1

[0111] Real-time update: Updating the integral value every time a new data point is received to ensure timeliness.

[0112] Decision support,

[0113] Parameter scanning: Generating sampling points in the blade geometric parameter space (such as leading edge angle θ, chord length c), and evaluating the damage amount under each parameter combination through the PINN model.

[0114] Optimization algorithm: Using the gradient descent method or genetic algorithm with minimizing damage as the objective function;

[0115] Result visualization: Generating a data graph to show the trade-off relationship between geometric parameters and damage.

[0116] Maintenance warning

[0117] Threshold setting: Define the dynamic threshold Dt(t) based on the fatigue limit of the material (such as the S-N curve) or historical failure data.

[0118] Early warning strategy:

[0119] Hard threshold: Trigger an emergency shutdown instruction when Dtotal > Dt.

[0120] Soft threshold: Initiate a preventive maintenance process (such as vibration detection, coating repair) when Dtotal ≥ 0.8Dt.

[0121] Adaptive adjustment: Dynamically correct Dt according to real-time working conditions (such as dust concentration, rotational speed) to enhance the early warning accuracy.

[0122] As can be seen from the above embodiments, the present invention provides a method for predicting erosion damage of aero-engine compressor blades based on a physics-informed neural network (PINN), aiming to overcome the limitations of traditional prediction methods in high-complexity working conditions in the prior art. Especially when facing factors such as dynamic changes, complex flow fields, and uneven particulate characteristics, the prediction accuracy and calculation efficiency often cannot meet the actual needs. Through this method, the present invention can not only improve the prediction accuracy and efficiency of blade erosion damage, but also provide a scientific basis for the design, operation and maintenance, fault diagnosis, and safety assessment of aero-engines.

[0123] The present invention improves the accuracy and reliability of erosion damage prediction. The present invention uses a physics-informed neural network (PINN) to combine traditional physical models with modern machine learning techniques, thereby achieving high-precision prediction of erosion damage of compressor blades in complex working environments. PINN can effectively integrate physical theories such as fluid mechanics and damage mechanics with blade damage data, enabling the model to perform high-precision damage prediction under unknown or partially known data conditions. This method effectively avoids the limitation of traditional machine learning methods that rely only on data training, ensuring that the prediction results are more scientific and accurate.

[0124] The present invention significantly reduces the computational cost and the need for experimental data. Traditional damage prediction methods require a large amount of experimental data and complex computational resources. Especially when simulating blade damage under different working conditions, the amount of calculation is huge and the cost is high. Through the PINN-based method, the present invention can, while ensuring the prediction accuracy, greatly reduce the dependence on experimental data and reduce the high experimental and computational costs. In addition, by adding physical constraints during the training process, PINN effectively improves the training efficiency of the model, reduces the demand for computational resources, and makes the prediction process more efficient and economical.

[0125] The present invention is applicable to a variety of complex working conditions and different damage modes. The method of the present invention is not only applicable to the blade damage prediction under conventional working environments, but also can effectively cope with the erosion damage under complex working conditions. For example, under working conditions with different air flow velocities, different environmental temperatures, and large variations in the type and concentration of particulate matter, traditional damage prediction methods often fail to adapt. PINN can adapt to the changing working condition environment through its flexible structure and physical information constraints, provide accurate prediction results, and provide effective support for the blade damage assessment under different operating states.

[0126] The present invention optimizes the design and material selection of aeroengine compressor blades. By accurately predicting the erosion damage of blades under complex working conditions, the present invention can provide important guidance for the design, material selection, surface treatment, etc. of aeroengine compressor blades. Based on the damage prediction results, designers can optimize the blade structure, select more suitable materials and surface coatings, enhance the erosion resistance of the blades, extend the service life of the blades, and improve the overall reliability and performance of the engine.

[0127] The present invention enhances the interpretability and adaptability of the model. The PINN model in the present invention not only has strong prediction ability, but also can improve the interpretability of the model through physical constraints. When predicting the erosion damage of blades, researchers can deeply analyze the mechanism of damage generation based on the relationship between the physical parameters provided by the model and the prediction results, providing a basis for subsequent blade optimization design and safety assessment. In addition, the PINN model can flexibly adapt to different types of erosion damage modes, such as microscopic damage, crack propagation, etc., providing accurate prediction support for various damage types under different working conditions.

[0128] The present invention provides decision-making support for the maintenance and fault diagnosis of aeroengines. The method of the present invention can provide decision-making support for the maintenance and fault diagnosis of aeroengines by real-time monitoring and predicting the erosion damage of blades. Through the accurate prediction of blade damage, operation and maintenance personnel can identify potential fault hazards in advance, implement targeted maintenance measures, reduce the probability of equipment failures, reduce operation risks, and improve the safety of aeroengines.

[0129] The present invention promotes the intelligent and digital transformation in the field of aeroengines. The proposal of the present invention not only has academic value and technical application value, but also can promote the intelligent and digital transformation in the field of aeroengines. By introducing advanced artificial intelligence technologies and combining with the physics-informed neural network (PINN) for damage prediction, the present invention can improve the intelligent level in aspects such as aeroengine design, manufacturing, operation and maintenance, and promote the technical development level of the whole industry.

[0130] In summary, by innovatively combining the physics-informed neural network (PINN) technology, the present invention provides an efficient, accurate, and scalable method for predicting the erosion damage of aero-engine compressor blades, solving the limitations of traditional methods under complex working conditions, promoting the safety, reliability, economy, and intelligence development of aero-engines, and making contributions to the progress of aero-engine technology.

[0131] Example 2, as Figure 3 shown, the aero-engine compressor blade erosion damage prediction system based on PINN provided by the embodiment of the present invention includes:

[0132] The physical mechanism and mathematical modeling module 1 is used to construct an erosion damage rate equation and a flow field-particle coupling equation for the discrimination of compressor blade erosion damage based on fluid mechanics and erosion dynamics according to the dynamic process of the interaction between the kinetic energy of particles and the material surface.

[0133] The dual-driven loss function design module 2 designs a dual-driven loss function based on the constructed erosion damage rate equation and flow field-particle coupling equation. The dual-driven loss function K integrates the data-driven term and the physical constraint term to ensure that the experimental data is fitted simultaneously and the conservation law is obeyed.

[0134] The physics-informed neural network model construction and adaptive training module 3 constructs a physics-informed neural network PINN model based on the constructed erosion damage rate equation, flow field-particle coupling equation, and the designed dual-driven loss function, and performs adaptive training on the constructed physics-informed neural network PINN model.

[0135] The erosion damage prediction module 4 is used for the prediction of compressor blade erosion damage and engineering applications.

[0136] In the above embodiments, the descriptions of each embodiment have their own focuses. For the parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0137] The method for predicting the erosion damage of aero-engine compressor blades based on the physics-informed neural network (PINN) proposed by the present invention has achieved significant breakthroughs in terms of accuracy, efficiency, cost, and engineering applicability compared with traditional technologies. The specific effects are as follows:

[0138] (1) The prediction accuracy has been significantly improved.

[0139] Revolutionary breakthrough in computational efficiency.

[0140] Real-time prediction ability: The time-consuming for a single damage prediction is 12 ms, meeting the on-line monitoring requirements of aero-engines, and the speed is increased by 1.89 million times compared with the traditional CFD method (6.3 hours).

[0141] Resource saving: Only 200 sets of experimental data are required in the training phase (traditional DNN requires > 10,000 sets), and the experimental cost is reduced by 80%.

[0142] (2) Full life cycle management support.

[0143] Design optimization: Rapidly identify high-damage areas through parameter scanning, guide the design of the serrated structure at the leading edge of the blade, and reduce the maximum erosion rate by 30% - 50%.

[0144] Operation and maintenance decision-making: Real-time cumulative damage prediction can trigger early warnings, avoid unplanned shutdowns, and extend the service life of the blade by 20% - 40%.

[0145] Fault tracing: Combine gradient visualization technology to locate the key influencing factors of damage (such as the sensitivity of the particle incidence angle), and improve the efficiency of fault analysis.

[0146] (3) Technical compatibility and scalability.

[0147] Multi-scale adaptation: Support cross-scale prediction from microscopic single-crystal damage to macroscopic blade deformation.

[0148] Hardware lightweight: The model can be deployed to embedded devices, with memory occupancy < 50MB, meeting the requirements of airborne real-time monitoring.

[0149] (4) Economic benefits and industry value.

[0150] Operation and maintenance cost of a single aero-engine: It is expected to save $120,000 per year (reducing the number of unplanned repairs).

[0151] R & D cycle shortened: The anti-erosion design cycle of the new blade is compressed from 6 months to 2 weeks.

[0152] Through the deep integration of physical laws and deep learning, the present invention realizes the leap of aero-engine erosion damage prediction from "experience-driven" to "model-data dual-driven". Its characteristics of high precision, high efficiency and low cost have been verified in actual engineering, providing a subversive tool for the intelligent design, operation and maintenance, and airworthiness certification of aero-engines.

[0153] Experimental example: Predict the erosion damage distribution of the titanium alloy blade of the 3rd stage of the high-pressure compressor of a certain type of turbofan engine in a desert environment, and guide the optimization design of its anti-erosion coating. Such as Figure 4 The method for predicting the erosion damage of aero-engine compressor blades based on PINN specifically includes:

[0154] 1) Data collection.

[0155] Experimental data:

[0156] In the experiment, simulate the sand dust working condition, and use laser scanning to measure the damage depth of the blade surface (resolution 1μm).

[0157] A total of 120 groups of data were collected, covering 6 kinds of dust concentrations, 4 kinds of airflow velocities, and 3 kinds of incident angle ranges (20° - 80°).

[0158] CFD-assisted data:

[0159] 500 groups of flow field - particle trajectory data were generated through ANSYS Fluent as pre-training input.

[0160] 2) Model construction and training.

[0161] Input layer: 9-dimensional parameters including airflow velocity v, pressure p, particle velocity vp, particle size d, density ρp, incident angle θ, surface roughness Ra, material hardness H, and operation time t.

[0162] Hidden layer: 4 fully connected layers, with 13 nodes in each layer. The Swish function is selected as the activation function to balance nonlinearity and gradient stability

[0163] Output layer: Erosion rate D and its spatial gradient

[0164] Pre-training: Based on 500 groups of CFD data, with an initial learning rate of 1e-3, Adam optimizer, and batch size of 128.

[0165] Fine-tuning: Load 120 groups of experimental data, dynamically adjust the loss weight (λ2 gradually increases from 0.1 to 1.0), and the learning rate decays to 1e-5.

[0166] Regional enhanced sampling: The sampling density of the leading edge of the blade (the first 20% area of the chord length) is increased by 3 times.

[0167] 3) Real-time prediction and verification.

[0168] Input parameters:

[0169] Airflow velocity: 420 m / s (Mach number 0.82);

[0170] Dust concentration: 6 g / m 3 , particle diameter d = 50 μm, density ρp = 2650 kg / m 3 ;

[0171] Blade material: Ti-6Al-4V, hardness H = 3.4 GPa, surface roughness Ra = 0.8 μm;

[0172] Output:

[0173] Predicted damage rate distribution (D) and its gradient field

[0174] Cumulative damage amount.

[0175] 4) Prediction results.

[0176] The accuracy comparison is shown in Table 1.

[0177] Table 1 Accuracy Comparison Table

[0178] Method Maximum relative error Average relative error Traditional CFD - empirical formula 12.3％ 8.7％ Pure data-driven DNN 18.5％ 15.2％ This embodiment (PINN) 4.9％ 3.2％

[0179] The calculation efficiency is shown in Table 2:

[0180] Table 2 Calculation Efficiency Table

[0181] Index Traditional CFD This embodiment (PINN) Single prediction time 4.5 hours 10 ms Total training time None 2.1 hours Experimental data requirement 500 sets of simulations 120 sets of experiments + 500 sets of simulations

[0182] Based on the PINN algorithm, the system can accurately predict the erosion damage of compressor blades under different working conditions. The specific prediction contents include:

[0183] The predicted damage time of each blade.

[0184] The depth and expansion trend of the damage area.

[0185] Emergency replacement warning and repair suggestions (for example: when the damage depth of a certain blade reaches 0.2 mm, prompt coating repair; when the damage depth exceeds 0.5 mm, suggest emergency replacement).

[0186] Through multiple experimental verifications, the accuracy of the prediction results has been fully verified.

[0187] 5) Engineering applications.

[0188] Design optimization: Based on the prediction results of PINN, the design of the blade leading edge has been optimized (such as by changing the incidence angle and surface roughness), and the erosion rate has been reduced by 37% in the leading edge area.

[0189] Maintenance decision: The real-time prediction results of the present invention can be used in the aircraft maintenance management system to help predict the blade life, avoid unplanned shutdowns, and extend the blade service life by 54%.

[0190] In summary, the present invention proposes a method for predicting the erosion damage of aero-engine compressor blades based on the physics-informed neural network (PINN), which successfully solves the limitations of traditional methods in terms of prediction accuracy, calculation efficiency, and adaptability. By embedding the fluid mechanics equation, the particulate motion model, and the material damage equation as physical constraints into the neural network, this method can accurately simulate the blade erosion damage process and demonstrate excellent prediction performance under different working conditions and environmental conditions.

[0191] The core innovation of the present invention lies in combining the physical information neural network with the complex physical model of erosion damage, accurately depicting the non-linear effects in gas-solid coupling, and ensuring that the prediction error is below 5%. In addition, through adaptive training and real-time data-driven, this method can dynamically adjust the prediction model to ensure high-precision and high-efficiency damage prediction under different working conditions. This dynamic adjustment ability significantly improves the robustness and adaptability of the model, meeting the damage prediction requirements of aero-engine blades in complex and changing environments.

[0192] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent replacement, and improvement made by those skilled in the art within the technical scope disclosed by the present invention, as long as they are made within the spirit and principle of the present invention, shall be covered by the protection scope of the present invention.

Claims

1. A PINN-based method for predicting erosion damage of aeroengine compressor blades, characterized in that: The method includes: S1, physical mechanism and mathematical modeling, based on the dynamic process of the interaction between the kinetic energy of particles and the material surface, based on fluid mechanics and erosion dynamics, for the identification of compressor blade erosion damage, the erosion damage rate equation and the flow field-particle coupling equation are constructed; S2, based on the constructed erosion damage rate equation and flow field-particle coupling equation, a dual-drive loss function is designed, wherein the dual-drive loss function L integrates the data-driven term and the physical constraint term to ensure that the experimental data is fitted and the conservation law is obeyed at the same time; S3, based on the constructed erosion damage rate equation and flow field-particle coupling equation, as well as the designed dual-drive loss function, a physical information neural network PINN model is constructed, and adaptive training of the constructed physical information neural network PINN model is performed; S4, Compressor blade erosion damage prediction and engineering application.

2. The PINN-based aeroengine compressor blade erosion damage prediction method according to claim 1, characterized in that: In step S1, the erosion damage rate equation includes: the damage rate D is determined by the particle collision characteristics, fluid parameters and material properties, and the modified erosion damage rate equation is introduced, and the expression is: Where D is the damage rate, α and β are the material erosion coefficients, ρ and v are the gas density and air velocity, d is the particle diameter, θ is the particle incident angle, and η is the collision efficiency.

3. The PINN-based aeroengine compressor blade erosion damage prediction method according to claim 2, characterized in that: In step S1, the flow field-particle coupling equation includes: describing the flow field characteristics through the Navier-Stokes equation and introducing the particle momentum exchange term, the expression is: In the formula, is the rate of change of velocity field with time, u is the velocity, is the Nabla operator, is the pressure gradient, is the Laplace operator, F p is the force exerted by discrete particles on the fluid; The force F exerted by the discrete particles on the fluid p The expression is: In the formula, is the acceleration of the particle, i is the particle index, m p ,v p are the particle mass and velocity respectively, and t is the running time.

4. The PINN-based aeroengine compressor blade erosion damage prediction method according to claim 3, characterized in that: In step S2, the expression of the conservation law is: L=λ1L data +λ2L phys (3) Where L is the dual-drive loss function, λ1,λ2 are the weights corresponding to data loss and physical loss, and L data is the data loss, L phys For physical damage; The data-driven term is used to minimize the mean square error between the neural network prediction and the experimental / CFD data: Where N is the number of experimental data samples, D PINN (x i ) is the predicted value of the pinn model, D exp (x i ) is the experimentally measured value, |·| 2 is the square error; x i are input parameters, including flow velocity, particle properties, and geometric parameters.

5. The PINN-based aeroengine compressor blade erosion damage prediction method according to claim 4, characterized in that: In step S2, the physical constraint term is used to force the physical information neural network output to satisfy equation (5): In the formula, M is the number of samples, j is the sample index, is the rate of change of damage rate D with time t; is the convection term, which represents the advection change of the damage rate D under the action of the velocity field u; S is the erosion source term, ||·|| 2 is the Euclidean norm; The expression of the erosion source term D is: Where D is the damage rate, α and β are the material erosion coefficients, ρ and v are the gas density and air velocity, d is the particle diameter, θ is the particle incident angle, and η is the collision efficiency.

6. The PINN-based aeroengine compressor blade erosion damage prediction method according to claim 1, characterized in that: In step S3, the physical information neural network PINN model includes: Input layer: 9-dimensional parameters: air flow velocity v, pressure p, particle velocity vp, particle size d, density ρp, incident angle θ, surface roughness Ra, material hardness H, and running time t; Hidden layer: 4 layers of full connection, 13 nodes in each layer, the activation function uses the Swish function to balance nonlinearity and gradient stability; Output layer: erosion rate D and its spatial gradient Used for physical residual calculations.

7. The PINN-based aeroengine compressor blade erosion damage prediction method according to claim 1, characterized in that: In step S3, adaptive training includes dynamic weight adjustment and enhanced sampling of key areas; Dynamic weight adjustment, including: initial weight, λ1 = 1.0, λ2 = 0.1, every 1000 iterations, increase λ2 by 10% to increase the physical constraint strength and avoid early overfitting; Strengthen sampling in key areas, including: In high damage areas such as the leading edge of the blade, the probability density function PDF is used to control the sampling distribution: Where p(x) is the probability density function, x edge is the coordinate of the leading edge of the blade, and ∝ indicates that they are in direct proportion.

8. The PINN-based aeroengine compressor blade erosion damage prediction method according to claim 1, characterized in that: In step S4, compressor blade erosion damage prediction includes: Input preprocessing, mapping sensor data including flow rate and particle concentration into network input parameters; In real-time reasoning, the PINN model outputs the damage rate D and integrates to calculate the cumulative damage. The calculation steps are as follows: Discretization processing: discretize the sensor data into D(t1), D(t2)…D(tn) according to the time step Δt; Numerical integration: Calculated using the trapezoidal rule or Simpson's rule: Where D total is the total damage rate, D(t k ) is t k The damage rate at time, D(t k-1 ) is t k-1 The damage rate at the moment, Δt is the step size, and k is the discretized index.

9. A PINN-based aeroengine compressor blade erosion damage prediction system, characterized in that: Implementing the PINN-based aeroengine compressor blade erosion damage prediction method as described in any one of claims 1 to 8, the system comprises: Physical mechanism and mathematical modeling module (1), which is used to identify compressor blade erosion damage based on the dynamic process of the interaction between particle kinetic energy and material surface, fluid mechanics and erosion dynamics, and construct the erosion damage rate equation and flow field-particle coupling equation; A dual-drive loss function design module (2) designs a dual-drive loss function based on the constructed erosion damage rate equation and the flow field-particle coupling equation. The dual-drive loss function L integrates the data-driven term and the physical constraint term to ensure that the experimental data is fitted and the conservation law is obeyed at the same time. Physical information neural network model construction and adaptive training module (3), based on the constructed erosion damage rate equation and flow field-particle coupling equation, and the designed dual-drive loss function, construct a physical information neural network PINN model, and adaptively train the constructed physical information neural network PINN model; The erosion damage prediction module (4) is used for compressor blade erosion damage prediction and engineering application.

10. The PINN-based aeroengine compressor blade erosion damage prediction system according to claim 9, characterized in that: The PINN-based aero-engine compressor blade erosion damage prediction system is mounted on a computer device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, the functions of the above-mentioned PINN-based aero-engine compressor blade erosion damage prediction system are implemented.

Images