Fatigue reliability analysis method for turbine disk blades based on digital-physical fusion and active learning

Abstract

The gas turbine disc blade fatigue reliability analysis method of the application fuses combination of numbers and objects and active learning, comprising: S1: obtaining a probability distribution model of main characteristic parameters of a gas turbine; S2: establishing a geometric simulation model of a gas turbine disc blade, dividing a key failure area of the gas turbine disc blade, and establishing a gas turbine disc blade life prediction model of a number-object fusion multi-failure mode; S3: determining random variables according to the probability distribution model of the main characteristic parameters, calculating a function function by using the gas turbine disc blade life prediction model of the number-object fusion multi-failure mode and constructing a training sample set, gradually increasing the training sample set in the process of constructing a proxy model based on unsupervised learning and active learning ideas, and establishing a SOM-AL-Kriging proxy model for gas turbine disc blade reliability analysis; and S4: performing fatigue reliability analysis on the gas turbine disc blade. The application improves the efficiency and precision of the gas turbine disc blade fatigue reliability analysis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of gas turbines, and particularly relates to a wheel disc blade fatigue reliability analysis method combining data and physical fusion with active learning. BACKGROUND

[0002] A gas turbine is a core equipment of a clean power generation power system, has the advantages of high thermal efficiency, low pollution, good reliability and maintainability, can use various fuels, and has a small amount of cooling water usage. Developing a high-power and high-efficiency gas turbine can provide important technical support for clean and efficient use of energy and safe and stable operation of a power grid, and has an irreplaceable strategic position in the sustainable development of the energy and power industries.

[0003] A gas turbine has a decisive influence on the efficiency and safety of energy and power supply. Therefore, the reliability and safety are the top priority in the life cycle tasks of the design, operation, maintenance and the like of the gas turbine. In the operation process of the gas turbine, due to the complex operation environment and the large number of on-board devices and systems, the operation reliability of the gas turbine is affected by the mutual coupling of the complex situations such as the operation environment, performance degradation of each component, damage accumulation and the like, and failure of any key component or system may cause serious influence on the operation of the whole machine. The wheel disc blade is a core component for heat and power conversion of the gas turbine, and has a complex and harsh working environment. Once a more serious failure or fault occurs, the consequences are that the power supply is insufficient, the operation cost is increased, or even the machine is destroyed and people are killed. Therefore, it is of great significance for the gas turbine system to ensure that the reliability and safety of the wheel disc blade of the gas turbine are within a specified range.

[0004] Traditional wheel disc blade fatigue reliability analysis is usually based on a physical model or a data-driven method. On the one hand, the physical model cannot comprehensively estimate the actual working condition, cannot adjust the related model parameters in real time, and the information reflecting the actual change of the degradation state in the measured signal cannot be utilized in real time. On the other hand, the data-driven method excessively depends on the performance degradation information contained in the data, and the accuracy and authenticity of the method are limited by the amount of data. Moreover, the traditional fatigue reliability analysis method does not consider the differences between different failure modes and their influence on the overall reliability of the wheel disc blade, and a large number of simulation analysis programs need to be called in the calculation, which has the technical problems of limited accuracy, low efficiency and high calculation cost. Therefore, it is an urgent problem to establish a data and physical fusion reliability analysis method, comprehensively consider multiple site failures, and improve the efficiency and precision of the fatigue reliability analysis of the wheel disc blade of the gas turbine. SUMMARY

[0005] In view of the deficiencies of the existing gas turbine disc blade reliability analysis method, the application provides a disc blade fatigue reliability analysis method combining digital-physical fusion and active learning.The application method comprehensively considers the combination of multiple failure models of the gas turbine disc blade, establishes a digital-physical fusion life prediction model, combines unsupervised learning and active learning, automatically selects the most valuable sample points for sequence sampling, and uses importance sampling for reliability calculation, thereby greatly improving the efficiency and accuracy of the gas turbine disc blade fatigue reliability analysis, and having the advantages of self-adaptation, high efficiency, reliability and the like.

[0006] To achieve the above object, the application realizes the technical scheme as follows:

[0007] The disc blade fatigue reliability analysis method combining digital-physical fusion and active learning comprises the following steps:

[0008] S1: Uncertainty characterization research is carried out for the gas turbine disc blade, and the probability distribution model of the main characteristic parameters of the gas turbine is obtained;

[0009] S2: A geometric simulation model of the gas turbine disc blade is established, the sensitivity analysis of the main characteristic parameters is carried out through finite element analysis, the key failure area of the gas turbine disc blade is divided, and a digital-physical fusion multiple failure mode life prediction model of the gas turbine disc blade is established;

[0010] S3: The random variables are determined according to the probability distribution model of the main characteristic parameters, the digital-physical fusion multiple failure mode life prediction model of the gas turbine disc blade is used for function function calculation and a training sample set D is constructed, the training sample set D is gradually increased in the process of proxy model construction based on unsupervised learning and active learning idea, and a SOM-AL-Kriging proxy model for the reliability analysis of the gas turbine disc blade is established;

[0011] S4: The SOM-AL-Kriging proxy model is used for fatigue reliability analysis of the gas turbine disc blade.

[0012] The further improvement of the application lies in that in step S1, under the condition of having sufficient experimental construction, the actual gas turbine disc blade sample is scanned to obtain the measurement data of the actual geometric shape, and then the main uncertainty parameters are extracted through principal component analysis to obtain the probability distribution model of the uncertainty parameters; in the case of lacking measured data, the uncertainty parameters and the probability distribution model are selected according to the past experience or engineering examples.

[0013] The further improvement of the application lies in that the uncertainty parameters include working environment parameters, structure parameters, material properties and load changes.

[0014] The further improvement of the present application is that the distribution of the uncertain parameter is normal distribution, lognormal distribution, Weibull distribution or uniform distribution.

[0015] The further improvement of the present application is that in step S2, according to the gas turbine disc blade entity, a geometric simulation model is established for fluid-structure coupling analysis, main characteristic parameters are taken as change input of the geometric simulation model for sensitivity analysis, overall disc blade aerodynamic load, temperature field and stress distribution are obtained; according to the temperature stress state of the overall disc blade, a stress value, a stress concentration coefficient and a region with a temperature greater than a specified value are selected as a dangerous region, the turbine disc is divided into different failure prone regions, and a theoretical life prediction model between structure parameters and service life is constructed.

[0016] The actual multi-source monitoring vibration data of the actual gas turbine disc blade under actual operation is obtained through a sensor, a denoising method based on wavelet transform or a denoising method based on independent variable analysis is used to denoise the obtained vibration signal of the gas turbine disc blade, the data collected in each time period is preprocessed, first, the training data is standardized under different working conditions through clustering, and each feature is normalized, then the value of the training set is established by setting labels and data enhancement through piecewise linear, and finally the training sample and the test sample are constructed through a sliding time window, a mapping relationship between the characteristic data and the life characteristic parameter of the gas turbine disc blade is established based on an Autoencoder-TCN neural network, and a data-driven life prediction model is obtained.

[0017] The theoretical life prediction model and the data-driven life prediction model are heterogeneously fused to obtain a complete numerical-physical fusion multi-failure mode gas turbine disc blade life prediction model.

[0018] The further improvement of the present application is that the theoretical life prediction model includes a low-cycle fatigue life prediction model, a high-cycle fatigue life prediction model, a fatigue fracture life prediction model, a high-temperature creep life prediction model and a high-low cycle composite fatigue life prediction model.

[0019] The further improvement of the present application is that in step S3, the probability distribution model of the main characteristic parameters obtained in step S1 is taken as a random variable and a probability density function of the function function, and the random variable is converted to a standard normal space; a function function is established according to the numerical-physical fusion multi-failure mode gas turbine disc blade life prediction model obtained in step S2, and the function function is as follows:

[0020]

[0021] In the formula, g(x i ) is a function function, x iThe uncertainty parameters involved in the life prediction model for the i-th location, The design life of the i-th location, N i The actual life analyzed for the i-th location;

[0022] Subsequently, considering the failure of different locations of the gas turbine disc blade, a multi-failure mode joint failure probability model is established. Specifically, a series connection is used to build a reliability model for the gas turbine disc blade. When any location fails, the whole fails, and the failure probability model formula is as follows:

[0023] P f =P{∏g(x i )≤0}

[0024] In the formula, P f is the failure probability, and f(x i ) is the joint probability density function of all uncertainty parameters;

[0025] N s random variables are extracted to generate an initial sample by using a random sampling method. Then, the real function function sample set is obtained through fluid-structure coupling analysis-life analysis-function function calculation, and the training sample set D is constructed. Unsupervised classification and dimension reduction are performed on the input parameters by using the unsupervised learning self-organizing mapping network SOM. The normalized output of the SOM neural network is used as the input of the surrogate model, and a Kriging surrogate model of the input parameters to the function function is constructed. Then, the current evaluation point is obtained by using the first-order second-moment method for solving, and the reliability index β k and the failure probability P f k are calculated. Whether the current failure probability converges P f k is determined according to the difference between the failure probabilities before and after the two times. The formula is as follows:

[0026]

[0027] When the failure probability P f k does not converge, the sampling center close to the limit state curve g(x i ) = 0 is obtained by using a linear interpolation method, and the calculation formula is as follows:

[0028]

[0029] In the formula, is the sampling center point obtained by the k-th interpolation calculation, x *k is the evaluation point obtained by the k-th calculation, is the mean value of the sample points;

[0030] Monte Carlo method is adopted to take the SOM-AL-Kriging surrogate model as the center of sampling, 10000 groups of candidate samples are generated according to the probability density distribution, the function function of the candidate samples is predicted according to the above SOM-AL-Kriging surrogate model, and the learning function value of each candidate sample is calculated through the following formula:

[0031] In the formula, AL(x) is the learning function value,

[0032] is the average Euclidean distance of the point x and all current sample points, is the minimum Euclidean distance of the point x and all current sample points, f x (x) is the joint probability density function of the point x. The minimum value of the learning function in the candidate sample is selected as the best sample point, and the learning function ensures that the added sample point is close to the limit state curve, and also ensures that the added sample point is as sparse as possible and does not gather and falls in the high-probability area; the real function function of the best sample point and the checking point is calculated by using the S2 to establish the life prediction model of the gas turbine disc blade in the multi-failure mode of the data and material fusion, the best sample point and the checking point and the corresponding function function are added to the training sample set D, and the SOM-AL-Kriging surrogate model is re-constructed.

[0033] When the failure probability P f k When convergence is achieved, the SOM-AL-Kriging surrogate model for the reliability analysis of the gas turbine disc blade is output for subsequent reliability analysis.

[0034] Further improvement of the present application is that in step S4, the reliability analysis of the gas turbine disc blade is carried out according to the SOM-AL-Kriging surrogate model obtained in step S3, the checking point is obtained by solving the SOM-AL-Kriging surrogate model by using the FORM method, then the checking point is taken as the important sampling center and the important sampling function is constructed, and finally the failure probability and the reliability of the gas turbine disc blade are obtained by simulating the important sampling, and the formula is as follows:

[0035]

[0036] R=1-P f

[0037] In the formula, R is the reliability, p x (x) is the important sampling probability density function, I(·) is the indicator function, f x (x) is the joint probability density function of the point x.

[0038]

[0039] ​Compared with the prior art, the application has at least the following beneficial technical effects:

[0040] The number-material fusion combined with active learning is used in the fatigue reliability analysis method of the disc blade, the method has wide applicability for complex gas turbine disc blade models, greatly improves the efficiency and precision of the fatigue reliability analysis of the gas turbine disc blade, and has the advantages of self-adaptation, high efficiency, reliability and the like.

[0041] Further, the method obtains main uncertain variables of the gas turbine disc blade through a dimension reduction method, and uses a probability model to characterize the uncertain parameters, thereby reducing the complexity of reliability analysis and improving the efficiency of reliability analysis;

[0042] Further, the method divides the gas turbine disc blade into different dangerous regions according to the actual load state of the gas turbine disc blade, comprehensively considers the influence of different failure modes on reliability, can better simulate the actual working condition of the gas turbine turbine, and improves the precision of reliability analysis;

[0043] Further, the method proposes a data-driven life prediction model based on Autocoder-TCN, which has high prediction accuracy in a complex working environment;

[0044] Further, the method organically fuses the theoretical life model and the data-driven life prediction model, establishes a number-material fusion multi-failure mode gas turbine disc blade life prediction model for subsequent reliability analysis, fully utilizes the advantages of the two types of methods, and further improves the accuracy of reliability analysis;

[0045] Further, the method combines unsupervised learning with active learning, uses a SOM neural network to perform unsupervised dimension reduction on multiple uncertain parameters, proposes an efficient general learning function and a double-pointing strategy, uses the expansion points obtained by interpolation as sampling centers to generate a small number of candidate samples, and automatically selects the most valuable sample points for sequence sampling, thereby improving the precision of the reliability analysis model, and reducing the required computing resources without having to call original function functions a lot;

[0046] Further, the method uses importance sampling to perform reliability calculation, thereby improving the sampling efficiency of the reliability analysis and further improving the efficiency of the reliability analysis. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The figure is a flowchart of the number-material fusion combined with active learning in the fatigue reliability analysis method of the disc blade;

[0048] Figure 2 The figure is a schematic diagram of the gas turbine disc blade of the embodiment of the application;

[0049] Figure 3A flow chart of a life prediction model of a gas turbine disc blade of a number-physical fusion multi-failure mode is constructed for the present application.

[0050] Figure 4 An Autoencoder-TCN neural network structure diagram is constructed for the present application.

[0051] Figure 5 A flow chart of a reliability analysis SOM-AL-Kriging surrogate model construction is constructed for the present application. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical effect and technical scheme of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application; obviously, the described embodiments are part of the embodiments of the present application. Based on the disclosed embodiments of the present application, other embodiments obtained by those of ordinary skill in the art without creative labor should fall within the scope of protection of the present application.

[0053] Referring to Figure 1 The disc blade fatigue reliability analysis method provided by the present application combines active learning with number-physical fusion, and comprises the following steps:

[0054] S1: Uncertainty characterization research is carried out for a gas turbine disc blade, and a probability distribution model of main characteristic parameters of the gas turbine is obtained.

[0055] Specifically, taking the disc blade shown in Figure 2 as an example, under the condition of having sufficient experimental construction, the actual gas turbine disc blade sample is scanned to obtain measurement data of the actual geometric shape, and then the main uncertainty parameters are extracted by dimension reduction methods such as principal component analysis to obtain a probability distribution model of the uncertainty parameters; in the case of lacking measured data, the uncertainty parameters and the probability distribution model are selected according to previous experience or engineering examples, wherein the uncertainty parameters include working environment parameters, structure parameters, material properties, load changes, etc.; the distribution of the uncertainty parameters can be normal distribution, lognormal distribution, Weibull distribution, uniform distribution, etc.

[0056] S2: A geometric simulation model of the gas turbine disc blade is established, sensitivity analysis of the main characteristic parameters is carried out through finite element analysis, key failure areas of the gas turbine disc blade are divided, and a life prediction model of the gas turbine disc blade of a number-physical fusion multi-failure mode is established.

[0057] Specifically, referring to Figure 3The flow chart of the gas turbine disc blade life prediction model of the shown number of physical fusion multi-failure modes, according to the gas turbine disc blade entity, a geometric simulation model is established for fluid-structure coupling analysis, the main characteristic parameters are taken as the change input of the geometric simulation model for sensitivity analysis, the overall disc blade aerodynamic load, temperature field, stress distribution and the like are obtained; according to the temperature stress state of the overall disc blade, the stress value, the stress concentration coefficient, the region with a temperature greater than a specified value are selected as the dangerous region, the turbine disc is divided into different failure prone regions, such as the disc core prone to low cycle fatigue failure, the disc rim prone to fatigue-creep failure and the like, a theoretical life prediction model between the structural parameters and the service life is constructed; the theoretical life prediction model includes a low cycle fatigue life prediction model, a high cycle fatigue life prediction model, a fatigue fracture life prediction model, a high temperature creep life prediction model and a high-low cycle composite fatigue life prediction model and the like.

[0058] For example, for the disc core prone to low cycle fatigue failure, the Manson-Coffin life prediction model commonly used for low cycle fatigue life prediction is selected, a low cycle fatigue probability life prediction model is obtained through statistical analysis, the Morrow correction model is used to correct the model by considering the influence of the average stress, and the low cycle fatigue probability life prediction model can be obtained as follows:

[0059]

[0060] In the formula, Δε and σ m are the total strain amplitude and the average stress of the dangerous examination part respectively, σ′ f is the fatigue strength coefficient, ε′ f is the fatigue plasticity index, E is the elastic modulus of the material, and N is the low cycle fatigue life.

[0061] For the region prone to fatigue fracture, the Paris formula of crack generation and propagation is selected as the failure mechanism model, and the life prediction model is as follows:

[0062]

[0063]

[0064] In the formula, is the crack propagation rate, C and m are material constants, a c and a0 are the initial crack length and the critical crack length respectively, Δσ is the stress range of the part under the working condition, and C1 is the crack propagation parameter.

[0065] Subsequently, multi-source monitoring vibration data of the actual gas turbine disc blade under actual operation is obtained through a sensor, a denoising method based on wavelet transform or a denoising method based on independent variable analysis is used to perform denoising processing on the vibration signal of the gas turbine disc blade, the data collected in each time period is preprocessed, the training data is standardized under different working conditions through clustering, and each feature is normalized, then the value of the training set is established by setting labels and data enhancement through piecewise linearization, finally, the training sample and the test sample are constructed through a sliding time window, the mapping relationship between the feature data and the life representation parameter of the gas turbine disc blade is established based on the Autoencoder-TCN neural network, and a life prediction model based on data driving is obtained. The Autoencoder-TCN neural network structure is as shown in Figure 4 The Autoencoder layer can learn the knowledge hidden in the sample, so as to obtain deep features of the sample data in the high-level hidden layer. Then the features are input into a time convolution (TCN) network stacked by residual blocks, and the TCN network is mainly responsible for feature extraction of time dependence and final prediction. In the network training process, first, L sliding window samples with a size of N L ×N W are input into the Autoencoder network, and in the training process, the output features of the Autoencoder are made as similar as possible to the input features by continuously reducing the reconstruction error. After the training is completed, the low-dimensional features output in the decoding stage are extracted as the input of the TCN network. These low-dimensional features contain all the information of the original input features, can reduce the calculation amount and improve the generalization ability. In the TCN network, an adaptive learning rate is adopted, a higher learning rate is used to accelerate the acquisition of better parameters, and the learning rate is reduced as the number of iterations increases, so as to ensure the convergence of the model in the later period.

[0066] Finally, the theoretical life prediction model and the life prediction model based on data driving are heterogeneously fused to obtain a complete numerical-physical fusion multi-failure mode gas turbine disc blade life prediction model.

[0067] S3: According to the probability distribution model of the main feature parameters, a random variable is determined, a numerical-physical fusion multi-failure mode gas turbine disc blade life prediction model is used to perform function function calculation and construct a training sample set D, and based on the unsupervised learning and active learning idea, the training sample set D is gradually increased in the process of constructing the proxy model, and a SOM-AL-Kriging proxy model for reliability analysis of the gas turbine disc blade is established.

[0068] Specifically, referring to Figure 5The shown reliability analysis SOM-AL-Kriging proxy model construction process adopts the probability distribution model of the main characteristic parameters obtained in step S1 as the random variables and probability density functions of the function function, and converts the random variables to the standard normal space; the function function is established according to the number-material fusion multi-failure mode gas turbine disc blade life prediction model obtained in step S2, and the function function is as follows:

[0069]

[0070] In the formula, g(x i ) is a function function, x i is an uncertainty parameter involved in the life prediction model of the i th part, is the design life of the i th part, N i is the actual life obtained by analyzing the i th part.

[0071] Then, considering the failure of different parts of the gas turbine disc blade, a joint failure probability model of multiple failure modes is established. Specifically, a series connection is used to construct the reliability model of the gas turbine disc blade. When any part fails, the whole fails, and the failure probability model formula is as follows:

[0072] P f =P{∏g(x i )≤0}

[0073] In the formula, P f is the failure probability, and f(x i ) is the joint probability density function of all uncertainty parameters.

[0074] N s random variables are extracted to generate an initial sample, and then the real function function sample set is obtained through fluid-structure coupling analysis-life analysis-function function calculation to construct the training sample set D, and the unsupervised learning self-organizing mapping network SOM is used for unsupervised classification and dimension reduction of the input parameters. The SOM neural network includes an input layer and a competition layer. The normalized output of the SOM neural network is used as the input of the proxy model, a Kriging proxy model of the input parameters to the function function is constructed, and then a first-order second-moment method (FORM) is used to solve to obtain the current checking point. The reliability index β k and the failure probability P f k are calculated. Whether the current failure probability converges P f k is judged according to the difference between the failure probabilities before and after the two times, and the formula is as follows:

[0075]

[0076] When the failure probability P f k When not converging, the sampling center close to the limit state curve (g(x i ) = 0) is obtained by linear interpolation method, and the calculation formula is as follows:

[0077]

[0078] In the formula, is the sampling center point obtained by the kth interpolation calculation, x *k is the checking point obtained by the kth calculation, is the mean value of the sample points.

[0079] The Monte Carlo (MC) method is used to generate 10000 groups of candidate samples with as the sampling center according to the probability density distribution, the function function of the candidate samples is predicted according to the above SOM-AL-Kriging surrogate model, and the learning function value of each candidate sample is calculated by the following formula:

[0080]

[0081] In the formula, AL(x) is the learning function value, is the average Euclidean distance of point x and all current sample points, d min,ζ (x) is the minimum Euclidean distance of point x and all current sample points, f x (x) is the joint probability density function of point x.

[0082] The minimum value of the learning function in the candidate samples is selected as the best sample point, which ensures that the added sample point is close to the limit state curve, and at the same time ensures that the added sample point is as sparse as possible and does not gather and falls in the high probability area. The real function function of the best sample point and the checking point is calculated by using the numerical-physical fusion multi-failure mode gas turbine disc blade life prediction model established by S2, the best sample point and the checking point and the corresponding function function are added to the training sample set D, and the SOM-AL-Kriging surrogate model is reconstructed.

[0083] When the failure probability P f k When converging, the SOM-AL-Kriging surrogate model for gas turbine disc blade reliability analysis is output for subsequent reliability analysis.

[0084] S4: fatigue reliability analysis of the gas turbine disc blade is performed by using the SOM-AL-Kriging surrogate model.

[0085] According to the SOM-AL-Kriging surrogate model obtained in step S3, the reliability analysis of the gas turbine disc blade is carried out, the FORM method is used to solve the checking point according to the SOM-AL-Kriging surrogate model, then the checking point is taken as the important sampling center and the important sampling function is constructed, and finally the failure probability and reliability of the gas turbine disc blade are obtained by the simulated important sampling, and the formula is as follows:

[0086]

[0087] R = 1 - P f

[0088] In the formula, R is the reliability, p x (x) is the important sampling probability density function, I(·) is the indicator function, f x (x) is the joint probability density function of the point x.

[0089] The fatigue reliability analysis method of the disc blade provided by the application combines the principal component analysis method with active learning, the main uncertain variables of the gas turbine disc blade are obtained, and the uncertain parameters are characterized by using a probability model; according to different failure regions of the disc blade, the turbine disc is divided into different failure-prone regions according to the overall disc blade temperature stress state; the advantages of the theoretical model and the data-driven model are comprehensively considered to construct a complete numerical and physical fusion multi-failure mode life prediction model of the gas turbine disc blade; when the multi-mode reliability is calculated, an initial sample set is constructed by a small number of sampling points, the initial sample set is subjected to unsupervised learning and dimension reduction by using the SOM neural network, the output of the SOM neural network is taken as the input of the Kriging surrogate model, the key sample points required for calculation are automatically supplemented in the construction of the surrogate model according to the active learning double-point adding strategy, and a high-precision surrogate model can be obtained after the active learning is completed, and then the reliability of the gas turbine disc blade is obtained by importance sampling. The method provides a high-precision reliability analysis method of the gas turbine disc blade, comprehensively considers the influence of different failure modes, organically combines the theoretical life model and the data-driven life prediction model, simultaneously uses the surrogate model based on unsupervised learning and active learning, greatly reduces the calculation cost required for reliability analysis, and is high in precision and low in cost.

[0090] Although the application has been described in detail in the foregoing description with general principles and specific embodiments, some modifications or improvements can be made on the basis of the application, which is obvious to those skilled in the art. Therefore, these modifications or improvements made on the basis of not deviating from the spirit of the application, all belong to the scope of protection claimed by the application.

Claims

1. A fatigue reliability analysis method for rotary disk blades that combines data and physical systems with active learning, characterized in that: Includes the following steps: S1: Conduct uncertainty characterization research on gas turbine disk blades to obtain probability distribution models of the main characteristic parameters of gas turbines; S2: Establish a geometric simulation model of the gas turbine disk blades, conduct sensitivity analysis on the main characteristic parameters through finite element analysis, delineate the key failure areas of the gas turbine disk blades, and establish a gas turbine disk blade life prediction model based on multiple failure modes of data and physical components; based on the gas turbine disk blade entity, establish a geometric simulation model for fluid-structure interaction analysis, use the main characteristic parameters as inputs to the geometric simulation model for sensitivity analysis, and obtain the overall disk blade aerodynamic load, temperature field, and stress distribution; based on the overall disk blade temperature and stress state, select areas with stress values, stress concentration factors, and temperatures exceeding specified values ​​as dangerous areas, divide the turbine disk into different easily failed areas, and construct a theoretical life prediction model between structural parameters and service life; Multi-source monitoring vibration data of actual gas turbine disk blades under actual operation is obtained through sensors. The obtained vibration signals of gas turbine disk blades are denoised using a wavelet transform-based denoising method or an independent variable analysis-based denoising method. The data collected in each time period are preprocessed. First, the training data under different operating conditions is standardized by clustering, and each feature is normalized. Then, the values ​​of the training set are established by setting labels in a piecewise linear manner and data augmentation. Finally, training samples and test samples are constructed by sliding time windows. The mapping relationship between feature data and life characterization parameters of gas turbine disk blades is established based on the Autoencoder-TCN neural network, and a data-driven life prediction model is obtained. The theoretical life prediction model and the data-driven life prediction model are heterogeneously fused to obtain a complete data-physical fusion multi-failure mode life prediction model for gas turbine disk blades. S3: Based on the probability distribution model of the main characteristic parameters, random variables are determined. A multi-failure mode fusion model of gas turbine disk blades is used to calculate the function and construct a training sample set. Based on the ideas of unsupervised learning and active learning, the training sample set is gradually increased during the agent model construction process. A SOM-AL-Kriging surrogate model was established for reliability analysis of gas turbine disk blades; S4: The SOM-AL-Kriging surrogate model is used to perform fatigue reliability analysis on the gas turbine disk blades. The FORM method is used to solve the model based on the SOM-AL-Kriging surrogate model to obtain verification points. These verification points are then used as important sampling centers to construct an important sampling function. Finally, the failure probability and reliability of the gas turbine disk blades are obtained through simulated important sampling. The formulas are as follows: In the formula, For reliability, Let I(·) be the importance sampling probability density function, and let I(·) be the indicator function. For point The joint probability density function.

2. The method for fatigue reliability analysis of wheel blades based on data-physical fusion and active learning according to claim 1, characterized in that, In step S1, under the condition of sufficient experimental setup, the actual gas turbine disk blade sample is scanned to obtain the measurement data of the actual geometry. Then, the main uncertainty parameters are extracted by principal component analysis to obtain the probability distribution model of the uncertainty parameters. In the absence of measured data, the uncertainty parameters and probability distribution model are selected based on past experience or engineering examples.

3. The method for fatigue reliability analysis of wheel blades based on data-physical fusion and active learning according to claim 2, characterized in that, Uncertain parameters include operating environment parameters, structural parameters, material properties, and load variations.

4. The method for fatigue reliability analysis of wheel blades based on data-physical fusion and active learning according to claim 2, characterized in that, The distribution of the uncertain parameter is normal, log-normal, Weibull, or uniform.

5. The method for fatigue reliability analysis of wheel blades based on data-physical fusion and active learning according to claim 1, characterized in that, The theoretical life prediction models include low-cycle fatigue life prediction models, high-cycle fatigue life prediction models, fatigue fracture life prediction models, high-temperature creep life prediction models, and high-low cycle combined fatigue life prediction models.

6. The method for fatigue reliability analysis of wheel blades based on data-physical fusion and active learning according to claim 1, characterized in that, In step S3, the probability distribution model of the main characteristic parameters obtained in step S1 is used as the random variable and probability density function of the function, and the random variable is transformed into the standard normal space; the function is established based on the data-physical fusion multi-failure mode gas turbine disk blade life prediction model obtained in step S2, and the function is as follows: In the formula, For functional purposes, For the first i Uncertainty parameters involved in the life prediction model for individual parts For the first i The design life of each part For the first i Actual lifespan obtained from analysis of individual components; Subsequently, considering the failure of different parts of the gas turbine disk blade, a multi-failure mode joint failure probability model is established. Specifically, a cascaded reliability model of the gas turbine disk blade is constructed. When any part fails, it is considered that the entire system has failed. The failure probability model formula is as follows: In the formula, This represents the probability of failure. Let be the joint probability density function for all uncertain parameters; Using random sampling method An initial sample is generated using random variables. Subsequently, through fluid-structure interaction analysis, lifetime analysis, and performance function calculation, the true performance function sample set is obtained, and a training sample set is constructed. Furthermore, an unsupervised learning self-organizing map network (SOM) is employed to perform unsupervised classification and dimensionality reduction of the input parameters. The SOM neural network includes an input layer and a competition layer. The normalized output of the SOM neural network is used as the input to the surrogate model to construct a Kriging surrogate model that maps the input parameters to the function. Subsequently, the first second-order moment method is used to obtain the current verification point and calculate its reliability index. and failure probability The difference between the failure probabilities of two consecutive failures is used to determine whether the current failure probability has converged. The formula is as follows: When failure probability When convergence fails, a curve close to the limit state can be obtained using linear interpolation. The sampling center is calculated using the following formula: In the formula, For the first k The sampling center points are obtained by interpolation. This is the verification point obtained from the k-th calculation. The mean of the sample points; Using the Monte Carlo method Using the sampling center, 10,000 candidate samples are generated based on the probability density distribution. The function function of the candidate samples is predicted according to the SOM-AL-Kriging surrogate model described above, and the learning function value of each candidate sample is calculated using the following formula: In the formula, To learn function values, For point The average Euclidean distance to all current sample points. For point The minimum Euclidean distance to all current sample points. For point The joint probability density function; The minimum value of the learning function among the candidate samples is selected as the optimal sample point. This learning function ensures that the added sample points are close to the limit state curve, while also ensuring that the added sample points are as sparse and non-clustered as possible and all fall within the high-probability region. The true function of the optimal sample point and the verification point is calculated using the data-physical fusion multi-failure mode prediction model of the gas turbine disk blade established by S2. The optimal sample point, the verification point, and the corresponding function are then added to the training sample set. Reconstruct the SOM-AL-Kriging proxy model; When failure probability Upon convergence, the SOM-AL-Kriging surrogate model for reliability analysis of gas turbine disk blades is output for subsequent reliability analysis.

Images