A numerical and real fusion test method for reliability of aero-engine blade disk structure

Abstract

The present invention discloses a method for digital-real fusion testing of the reliability of an aero-engine blade disk structure, which is applied to the field of time-varying reliability assessment of an aero-engine. The method comprises: collecting response data of key parts of the blade disk structure under a full flight cycle; establishing a finite element virtual entity to obtain time-varying input variables and time-varying output responses under a full flight cycle of the aero-engine blade disk structure; establishing a loss function for embedding physical information, and establishing an optimal physical information embedding multiple response regression model based on a Bayesian neural network inference framework; extracting a large number of time-invariant data sets, fitting the limit state function using the established multi-response regression model, calculating the reliability of the time-varying system based on the Monte Carlo idea, and completing the reliability assessment of the time-varying system. The present invention converts a complex time-varying reliability assessment problem into a time-invariant response regression problem, effectively improving the calculation accuracy while simplifying the calculation task.

Description

Technical Field

[0001] The invention belongs to the field of time-varying reliability assessment of aircraft engines, and in particular relates to a numerical-real fusion testing method for reliability of an aircraft engine blade disk structure. Background Art

[0002] As a key component of aircraft engines, the blade disk structure is subjected to harsh loads and complex operating conditions (such as startup, idling, takeoff, climb and cruise), which directly affects the performance and safety of the engine. However, traditional reliability assessment methods usually rely on data acquisition of physical models, but often face many challenges in practical applications. The manufacturing and testing process of physical models is not only long-term, resource-intensive, and costly, but also particularly demanding in the field of aerospace, where the accuracy and real-time nature of test data are particularly demanding.

[0003] In recent years, the research on time-varying reliability assessment has gradually increased. The existing methods can be roughly divided into three categories: methods based on cross-rate, methods based on time discretization, and surrogate model methods. Due to the advantages of surrogate model methods in computational efficiency and accuracy, especially in dealing with implicit time-varying problems, they have become the research focus in the field of time-varying reliability assessment. However, for complex structures with multiple operating states such as aircraft engine blades, traditional surrogate model methods need to rely on a large amount of test data and historical data to describe the reliability changes of the system in different operating states. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a digital-physical fusion test method for the reliability of an aero-engine blade disk structure. By effectively fusing finite element simulation data with limited physical model test results, the limitation caused by the difficulty in obtaining physical model data can be overcome to a certain extent. Based on the idea of ​​extreme value conversion, the analysis process is decomposed into multiple time periods, and the corresponding extreme values ​​are extracted as output responses in each time period, so as to convert time-varying problems into time-invariant problems, thereby improving the reliability calculation accuracy and efficiency of complex structural systems.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] A method for numerical and real fusion testing of the reliability of an aircraft engine blade disk structure, comprising:

[0007] Step S1, based on the geometric dimensions and material properties of the actual blade disk structure, a finite element virtual entity is established to identify the parts of the virtual entity blade disk structure that are subjected to the maximum stress and strain in each stage of starting, idling, take-off, climbing and cruising;

[0008] Step S2, based on the identified key positions of the maximum stress and strain of the virtual entity blade disk structure in each operation stage, stress, strain and deformation sensors are installed at the corresponding positions of the actual blade disk structure, response data of the actual blade disk structure under full flight cycle conditions are collected, and the response data are preprocessed;

[0009] Step S3, using the preprocessed response data to calibrate the parameters of the finite element virtual entity to obtain an optimized finite element virtual entity;

[0010] Step S4, using optimized finite element virtual entity calculation to obtain time-varying input variables and time-varying output responses under full flight cycle conditions;

[0011] Step S5, discretizing the time-varying input variables into random input variables, and using the extreme value response mapping method to convert the time-varying output response into a time-invariant extreme value response, using the linkage sampling technology to simultaneously extract multidimensional random input variables and multiple time-invariant extreme value responses, and establishing a time-invariant sample set of multidimensional input variables and multiple output responses;

[0012] Step S6, constructing a digital-real fusion modeling framework based on the time-invariant sample data set to fit the nonlinear mapping between the multidimensional input variables and the limit state function;

[0013] Step S7, using physical information as a constraint embedding loss function to establish a Bayesian neural network model for physical information embedding, wherein the loss function for physical information embedding includes a data error term, a regularization term, and a physical loss term;

[0014] Step S8, by maximizing the posterior probability, finding the optimal hyperparameters of the Bayesian neural network and establishing the optimal Bayesian neural network model;

[0015] Step S9, based on the time-invariant data set, and using the optimal Bayesian neural network model to fit the limit state function, the reliability of the time-varying system is calculated based on the Monte Carlo simulation method to complete the time-varying reliability evaluation of the actual blade disk structure.

[0016] The beneficial effects of the present invention are:

[0017] (1) Digital-real fusion test method: This method innovatively introduces the concept of digital-real fusion, organically combining finite element virtual entity simulation data and real experimental data to more accurately reflect the stress-strain behavior of the blade disk under complex flight conditions. Traditional finite element simulation is usually difficult to fully capture the complexity of actual flight conditions. By integrating measured data during the finite element model calibration process, the deviation between the simulation model and the actual situation can be effectively reduced, thereby improving the reliability and accuracy of the overall test. The digital-real fusion test method not only enhances the model's ability to simulate real flight conditions, but also improves the model's robustness and stability.

[0018] (2) Extreme response mapping method: By converting the time-varying dynamic response into a time-invariant extreme response, the analysis process of complex time-varying systems is simplified. Specifically, this method discretizes the dynamic random process into a series of random variables, divides the time domain into several sub-time domains, and extracts the extreme response of each time domain, thereby generating a time-invariant response sample set. This extreme value mapping method not only retains the key information in the time-varying system, but also significantly reduces the amount of model calculation.

[0019] (3) Bayesian neural network method with physical information embedding: By embedding physical information into the neural network as a constraint, the loss function includes not only the data error term and the regularization term, but also the physical loss term. This method introduces physical laws into the model training process to ensure that the prediction results meet the data accuracy requirements while maintaining the rationality and consistency of the physical process, effectively solving the problem of overfitting of data-driven models to data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The present invention is a flow chart of a method for fusion testing of the reliability of an aircraft engine blade disk structure using digital and real elements. DETAILED DESCRIPTION

[0021] The present invention will be further described below in conjunction with specific embodiments, but is not limited to the specific embodiments.

[0022] During the full flight phase of an aircraft engine, including starting, idling, takeoff, climbing and cruising, the loads such as gas temperature, rotation speed and inlet velocity borne by the turbine integral blade usually present complex time-varying characteristics. The following takes the time-varying reliability assessment of the integral blade of an aircraft engine during the full flight phase as an example to further illustrate the digital-real fusion test method for the reliability of the aircraft engine blade structure. The specific steps are as follows:

[0023] like Figure 1 As shown, a flow chart of a method for digital-real fusion testing of the reliability of an aircraft engine blade disk structure of the present invention includes the following steps:

[0024] Step S1, based on the geometric dimensions and material properties of the actual blade disk structure, a finite element virtual entity is established to identify the key parts of the blade disk that are subjected to the maximum stress and strain in each stage of starting, idling, take-off, climbing and cruising;

[0025] Step S2, based on the identified key positions of the maximum stress and strain of the virtual entity blade disk structure in each operation stage, stress, strain and deformation sensors are installed at the corresponding positions of the actual blade disk structure, response data of the actual blade disk structure under full flight cycle conditions are collected, and the response data are preprocessed;

[0026] Step S3, using the preprocessed response data to further calibrate the parameters of the finite element virtual entity (such as material elastic modulus, structural geometric dimensions, boundary conditions, etc.), to obtain an optimized finite element virtual entity, to ensure that the model can truly reflect the stress and strain distribution of the blade disk at different flight stages;

[0027] Step S4, considering the multiple uncertainties of load characteristics and material properties of the blade disk system under the full flight cycle, the inlet pressure p, thermal conductivity h, thermal expansion coefficient k, elastic modulus E, density ρ and other random variables are determined as time-invariant input variables by optimizing the finite element virtual entity; gas temperature T, speed , the air intake velocity v is a time-varying input variable; stress ,strain , deformation d m is the time-varying output response;

[0028] Step S5, using the extreme response mapping method, the entire time-varying input variable in the time domain [0s, 215s] is discretized into time-invariant random variables at 11 time points, and the time-varying output response (i.e., stress, strain, and deformation) of the turbine integral blade is calculated based on the fluid-thermal-structural coupling technology. Finite element analysis shows that the time-varying output response reaches a peak value in the time domain [165s-200s], which is regarded as a time-invariant extreme output response. 50 groups of input samples ( , T,v, p, h, k, E, ρ) and applied it to fluid-thermal-solid coupling analysis to obtain multi-output time-invariant extreme output responses (deformation, stress, strain), and establish a time-invariant sample set of multi-dimensional input variables and multiple output responses;

[0029] Step S6, constructing a digital-real fusion modeling framework based on the time-invariant sample data set to fit the nonlinear mapping between the multidimensional input variables and the limit state function;

[0030] Step S7, considering that the output response (deformation, stress, strain) and the input workload (i.e., rotation speed, gas temperature) always maintain a negative correlation, the negative correlation physical law is embedded in the loss function, and a Bayesian neural network model for physical information embedding is established, wherein the loss function for physical information embedding includes a data error term, a regularization term, and a physical loss term;

[0031] Step S8, by maximizing the posterior probability, finding the optimal hyperparameters of the Bayesian neural network and establishing the optimal Bayesian neural network model;

[0032] Step S9, based on the time-invariant data set, and using the optimal Bayesian neural network model to fit the limit state function, the reliability of the time-varying system is calculated based on the Monte Carlo simulation method to complete the time-varying reliability evaluation of the actual blade disk structure.

[0033] Step S10, after conducting more physical tests of the blade disk structure, the new data is incorporated into the training data to update the proxy model; when the new data shows that the finite element model deviates greatly from the actual situation, the new data is incorporated into the finite element model to update the finite element model, thereby realizing real-time update of the time-varying reliability assessment.

[0034] The step S2 includes the following sub-steps:

[0035] Step S21, select a suitable sensor according to the required measurement type and accuracy. Common sensors include: strain sensor: measures the strain change of the material when it is subjected to force; stress sensor: directly measures the stress state inside the material; deformation sensor: monitors the deformation of the material.

[0036] Step S22, install the sensor at the determined monitoring point to ensure that the sensor is in good contact with the surface of the blade disk. If necessary, use a special adhesive or a fixing fixture to fix the sensor. Ensure that the installation direction of the sensor is consistent with the measurement target to avoid external interference affecting data collection.

[0037] Step S23, setting up a suitable data acquisition system to collect data from multiple sensors in real time and ensure the accuracy and completeness of the data.

[0038] The step S5 includes:

[0039] Step S51, according to the extended optimal linear estimation method, the Gaussian random process P(t) is decomposed into a combination of a mean term, a standard deviation term and an autocorrelation function term, assuming that the Gaussian random process P(t) is represented by a linear combination of N orthogonal characteristic functions and independent standard normal random variables:

[0040] (1)

[0041] Among them, λ k represents the kth eigenvalue of the autocorrelation coefficient; represents the kth characteristic function corresponding to the kth eigenvalue; η k represents a standard normal random variable; represents the mean of P(t).

[0042] Step S52: based on the time-varying characteristics, decompose the entire time domain [0, T] into multiple subdomains [0, T] = ∑ [T i , T i+1 ], and get the i-th subdomain [T i , T i+1 ] to output the extreme values ​​of the response:

[0043] (2)

[0044] in, represents the sub-time domain [T i , T i+1 ] the mth extreme value output response; Represents a subdomain [T i , T i+1 ] is the minimum value of the mth limit state function in .

[0045] In step S7, the loss function of embedding physical information as a constraint is expressed as:

[0046] (3)

[0047] Among them, L d Represents data loss, which is the error between the predicted value and the true value; L p represents the regularization term, which is constructed through the prior distribution of the parameters according to the Bayesian framework; L phy represents the physical loss term; λ represents the weight; Represents a hyperparameter. The loss function that embeds physical information during training makes the model output not only conform to the data distribution but also follow the physical constraints, and regularizes the model parameters through the prior distribution, thereby preventing overfitting of the training data and improving the generalization performance of the model.

[0048] The step S8 comprises:

[0049] Step S81, optimize the hyperparameters by maximizing the posterior probability, and the posterior probability of the parameters is expressed by the Bayesian formula:

[0050] (4)

[0051] Where P(D|θ) represents the likelihood function, which reflects the fitness of data D under a given hyperparameter θ; P(θ) represents the prior distribution, which describes the prior information of the hyperparameter θ.

[0052] Step S82, combining the loss function of the model and the posterior probability of the hyperparameters, constructing the objective function, and minimizing the objective function to obtain the optimal hyperparameters:

[0053] (5)

[0054] in, It represents the negative logarithm of the posterior probability, reflecting the uncertainty of the hyperparameter θ combined with the data D.

[0055] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for testing the reliability of an aircraft engine blade disk structure by integrating digital and real elements, characterized in that: The following steps are involved: Step S1, based on the geometric dimensions and material properties of the actual blade disk structure, a finite element virtual entity is established to identify the parts of the virtual entity blade disk structure that are subjected to the maximum stress and strain in each stage of starting, idling, take-off, climbing and cruising; Step S2, based on the identified key positions of the maximum stress and strain of the virtual entity blade disk structure in each operation stage, stress, strain and deformation sensors are installed at the corresponding positions of the actual blade disk structure, response data of the actual blade disk structure under full flight cycle conditions are collected, and the response data are preprocessed; Step S3, using the preprocessed response data to calibrate the parameters of the finite element virtual entity to obtain an optimized finite element virtual entity; Step S4, using optimized finite element virtual entity calculation to obtain time-varying input variables and time-varying output responses under full flight cycle conditions; Step S5, discretizing the time-varying input variables into random input variables, and using the extreme value response mapping method to convert the time-varying output response into a time-invariant extreme value response, using the linkage sampling technology to simultaneously extract multidimensional random input variables and multiple time-invariant extreme value responses, and establishing a time-invariant sample set of multidimensional input variables and multiple output responses; Step S6, constructing a digital-real fusion modeling framework based on the time-invariant sample data set to fit the nonlinear mapping between the multidimensional input variables and the limit state function; Step S7, using physical information as a constraint embedding loss function to establish a Bayesian neural network model for physical information embedding, wherein the loss function for physical information embedding includes a data error term, a regularization term, and a physical loss term; Step S8, by maximizing the posterior probability, finding the optimal hyperparameters of the Bayesian neural network and establishing the optimal Bayesian neural network model; Step S9, based on the time-invariant data set, and using the optimal Bayesian neural network model to fit the limit state function, the reliability of the time-varying system is calculated based on the Monte Carlo simulation method to complete the time-varying reliability evaluation of the actual blade disk structure.

2. The method for testing the reliability of an aero-engine blade disk structure by combining digital and real elements according to claim 1, characterized in that: In the step S2, the strain sensor is used to measure the strain change of the material when it is subjected to stress; the stress sensor is used to directly measure the stress state inside the material; and the deformation sensor is used to monitor the deformation of the material.

3. The method for testing the reliability of an aero-engine blade disk structure by combining digital and real elements according to claim 1, characterized in that: The step S5 comprises: Step S51, according to the extended optimal linear estimation method, the Gaussian random process P(t) is decomposed into a combination of a mean term, a standard deviation term and an autocorrelation function term, assuming that the Gaussian random process P(t) is represented by a linear combination of N orthogonal characteristic functions and independent standard normal random variables: (1) Among them, λ k represents the kth eigenvalue of the autocorrelation coefficient; represents the kth eigenfunction corresponding to the kth eigenvalue; represents a standard normal random variable; represents the mean value of P(t), Step S52: based on the time-varying characteristics, decompose the entire time domain [0, T] into multiple subdomains [0, T] = ∑[T i , T i+1 ], and get the i-th subdomain [T i , T i+1 The extreme values ​​of the output response in ] are: (2) in, Represents a subdomain [T i , T i+1 ] the mth extreme value output response; Represents a subdomain [T i , T i+1 ] is the minimum value of the mth limit state function in .

4. The method for testing the reliability of an aero-engine blade disk structure by combining digital and real elements according to claim 1, characterized in that: In step S7, the loss function of embedding physical information as a constraint is expressed as: (3) Among them, L d Represents data loss, which is the error between the predicted value and the true value; L p represents the regularization term, which is constructed through the prior distribution of parameters according to the Bayesian framework; L phy represents the physical loss term; λ represents the weight; Represents a hyperparameter.

5. The method for testing the reliability of an aero-engine blade disk structure by combining digital and real elements according to claim 1, characterized in that: The step S8 comprises: Step S81, optimize the hyperparameters by maximizing the posterior probability, and the posterior probability is expressed by the Bayesian formula: (4) Among them, P(D|θ) represents the likelihood function, which reflects the fitness of data D under a given hyperparameter θ; P(θ) represents the prior distribution, which describes the prior information of the hyperparameter θ. Step S82, combining the loss function of the model and the posterior probability of the hyperparameters, constructing the objective function, and minimizing the objective function to obtain the optimal hyperparameters: (5) in, It represents the negative logarithm of the posterior probability, reflecting the uncertainty of the combination of the prior information of the hyperparameter θ and the data D.

Images