Turbine blade high-low cycle composite fatigue reliability analysis method based on neural network

Abstract

The invention discloses a neural network-based turbine blade high-low cycle composite fatigue reliability analysis method, which comprises the following steps of: obtaining a small amount of data samples of a turbine blade through finite element analysis, extracting the characteristics of the data samples by using a convolutional neural network, and inputting the extracted characteristics into a BP neural network optimized by an improved seagull optimization algorithm to obtain the high-low cycle composite fatigue reliability of the turbine blade. The method comprises the following steps: establishing a CNN-MSOA-BP neural network model of a turbine blade under a high-low cycle composite fatigue load, taking multi-source uncertainty such as a load parameter and a material parameter into consideration, sampling a large number of input samples through Monte Carlo, and obtaining a corresponding output response through the established CNN-MSOA-BP neural network model, thereby performing reliability analysis of the turbine blade. According to the method provided by the invention, the reliability of the turbine structure under complex loads such as high-low cycle compound fatigue can be analyzed, the method has relatively high precision and operation efficiency, and theoretical support is provided for reliability analysis and optimization of turbine blades.

Description

Technical Field

[0001] The present invention relates to a reliability analysis method for high - low cycle compound fatigue of turbine blades, and more particularly, to a reliability analysis method for high - low cycle compound fatigue of turbine blades based on a neural network. Background Art

[0002] Turbine blades are one of the indispensable components in aero - engines. They operate under harsh conditions of high temperature, high pressure, and high rotational speed, resulting in frequent failures. During service, turbine blades are mainly subjected to the combined action of centrifugal load and temperature load, and are also affected by factors such as aerodynamic vibration. Their main failure mode is high - low cycle compound fatigue (CCF). Due to the influence of multi - source uncertainty factors such as material properties and applied loads, the high - low cycle compound fatigue of turbine blades shows great dispersion, making it impossible for deterministic analysis methods to accurately predict the reliability of turbine blades. Therefore, in order to ensure the reliability and structural stability of aircraft engine turbine blades during operation, it is of great significance to carry out life prediction and reliability assessment under high - low cycle compound fatigue loads.

[0003] Although the integrated method of Monte Carlo simulation and finite - element model is very effective in analyzing the probabilistic reliability of structures, its calculation requires a large amount of time and computing resources, resulting in great limitations of this method in engineering applications. Aiming at the computational cost of analyzing a large number of samples by the finite - element model, surrogate models have lower costs and higher computational efficiency in evaluating the probabilistic reliability of complex structures; at the same time, convolutional neural networks have good multi - level feature extraction ability and generalization ability. Therefore, it is necessary to carry out research on a reliability analysis method for turbine blades under high - low cycle compound loads based on convolutional neural networks to improve the prediction accuracy and computational efficiency in probabilistic reliability analysis of complex structures. Summary of the Invention

[0004] The purpose of the present invention is to provide a reliability analysis method for high - low cycle compound fatigue of turbine blades based on a neural network, so as to solve the problem of high cost and long time - consuming of the commonly used integrated method of finite - element model and Monte Carlo simulation, and to consider the multi - source uncertainty of factors such as material characteristics and applied loads in the process of reliability analysis.

[0005] To achieve the purpose of the present invention, the technical solution adopted is: a reliability analysis method for high - low cycle compound fatigue of turbine blades based on a neural network, comprising the following steps:

[0006] Step 1, establish a finite - element model of the turbine blade, set the constraint conditions and apply loads, and perform deterministic finite - element analysis to obtain the dangerous part;

[0007] Step 2, select random variables and determine their distribution characteristics;

[0008] Step 3: Use the Latin hypercube sampling method to separately extract a small number of input samples for low-cycle fatigue and high-cycle fatigue. These input samples serve as the inputs to the finite element model.

[0009] Step 4: Substitute the extracted input samples into the finite element model for simulation to obtain output samples under low-cycle fatigue and high-cycle fatigue loads of the turbine blade.

[0010] Step 5: Divide the input samples and output samples into a training set and a test set in a ratio of 7:3.

[0011] Step 6: Use a convolutional neural network to extract the sample features of the training set and use them as the inputs to the BP neural network.

[0012] Step 7: Use the improved seagull optimization algorithm to optimize the parameters of the BP neural network and establish a CNN-MAOS-BP neural network model.

[0013] Step 8: Use the test set to verify the prediction accuracy and convergence speed of the CNN-MAOS-BP neural network model.

[0014] Step 9: According to the distribution characteristics of the random input variables, use the Monte Carlo method to extract a large number of 10,000 - 100,000 groups of samples, and obtain the output responses of low-cycle fatigue and high-cycle fatigue through the CNN-MAOS-BP neural network model.

[0015] Step 10: Predict the high-low cycle composite fatigue life and damage of the turbine blade under multi-source uncertainty based on the output responses.

[0016] Step 11: Establish a performance function for the high-low cycle composite fatigue of the turbine blade, count the number of samples in the safe domain, and calculate the reliability of the turbine blade under high-low cycle composite loads.

[0017] Preferably, in Step 3, according to the uncertainty factors of the material properties and applied loads of the turbine blade during service, some parameters, such as density, elastic modulus, turbine blade rotation speed, thermal conductivity, and temperature, are taken as random variables.

[0018] Preferably, in Step 6, the features extracted by the convolutional neural network for the training set are the features of the turbine blade under complex loads and are used as the input parameters of the BP neural network.

[0019] Preferably, in Step 7, the random parameter B for balancing exploration and exploitation in the improved seagull optimization algorithm is:

[0020] B = 2 × A 2 × R1

[0021] where A is the introduced additional variable and R1 is the scaling factor.

[0022] Preferably, in step 7, the new position where the improved seagull optimization algorithm moves towards the optimal individual is:

[0023]

[0024] where, is the position to avoid collision, represents the position at the current iteration number, and C is a variable related to the seagull search behavior.

[0025] Preferably, in step 7, the position after the improved seagull optimization algorithm attacks the prey is:

[0026]

[0027] where, w is the inertia weight factor, which can be calculated by the formula and w0 is the weight of the initial position, taking w0 = 1.2; x, y, z are the coordinates when the seagull attacks the prey with a spiral motion behavior in the air; is the current best position.

[0028] Preferably, in step 8, if the accuracy of the established CNN-MAOS-BP neural network model does not meet the requirements, return to step 7; if the established CNN-MAOS-BP neural network model meets the accuracy requirements, go to step 9.

[0029] Preferably, in step 10, the linear damage law is extended to the high-low cycle composite fatigue load, that is:

[0030]

[0031] where, D(N) is the cumulative fatigue damage of N composite blocks, N L is the low-cycle fatigue life, N H is the high-cycle fatigue life, and n is the high-low cycle frequency ratio.

[0032] Preferably, in step 11, the performance function of the turbine blade under the high-low cycle composite fatigue load is:

[0033]

[0034] where, a is the damage intensity parameter, and in practice a takes 1; N is the number of times; z > 0 represents the safe domain, and z ≤ 0 represents the failure domain.

[0035] Preferably, in step 11, the reliability of the multi-source uncertainty of the turbine blade under the high-low cycle composite fatigue load can be obtained by counting the number of samples in the safe domain, that is:

[0036]

[0037] Among them, I r (z i ) is an exponential function, and z i is a function, M r is the number of samples within the safety domain, and M is the total number of samples.

[0038] The beneficial effects of the present invention are as follows:

[0039] The turbine blade reliability analysis method proposed by the present invention combines finite element analysis and neural network models. First, a relatively small amount of output sample data of low-cycle load and output sample data of high-cycle load are obtained through finite element analysis. The convolutional neural network is used to extract features from the obtained sample data, and then the sample data after feature extraction is input into the improved seagull optimization algorithm to optimize the parameters of the BP neural network, thereby establishing a CNN-MSOA-BP neural network model for reliability evaluation. At the same time, the influences of uncertainty factors such as high-cycle fatigue load, low-cycle fatigue load, and material properties on the reliability of turbine blades are considered. The turbine blade reliability analysis method proposed by the present invention has high modeling accuracy and simulation efficiency, saves the cost and time of experiments, and provides assistance for the reliability design of turbine blades. Description of the Drawings

[0040] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not constitute an improper limitation of the present invention. In the drawings:

[0041] Figure 1 is the structure based on the convolutional neural network - neural network;

[0042] Figure 2 is the high-low cycle composite fatigue load spectrum;

[0043] Figure 3 is the comparison between the low-cycle fatigue stress test set and the prediction result;

[0044] Figure 4 is the comparison between the low-cycle fatigue strain test set and the prediction result;

[0045] Figure 5 is the comparison between the high-cycle fatigue stress test set and the prediction result;

[0046] Figure 6 is the reliability change curve of turbine blades using different methods. Detailed Embodiments

[0047] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. The schematic embodiments and descriptions of the present invention are used to explain the present invention, but not to limit the present invention.

[0048] A method for analyzing the high-low cycle composite fatigue reliability of turbine blades based on a neural network provided by the present invention includes the following steps:

[0049] Step 1, establish a finite element model of the turbine blade, set the constraint conditions and apply loads, perform deterministic finite element analysis, and obtain the dangerous parts.

[0050] Step 2, according to the material properties of the turbine blade and the random uncertainty of the loads applied during service, some parameters, such as density, elastic modulus, rotor speed, thermal conductivity, and temperature, etc., are taken as random input variables and are considered to follow a normal distribution.

[0051] Step 3, adopt the Latin hypercube sampling method, and respectively extract 100 groups of low-cycle fatigue and high-cycle fatigue samples as the input samples of the finite element model according to the distribution characteristics in Step 2.

[0052] Step 4, substitute the extracted input samples into the finite element model for analysis, and obtain the output samples of the turbine blade under low-cycle fatigue and high-cycle fatigue loads corresponding to the 100 groups of input samples.

[0053] Step 5, divide the obtained input samples and output samples into a training set and a test set, where 70 groups are used as the training set and 30 groups are used as the test set for subsequent verification of the accuracy of the CNN-MAOS-BP neural network model.

[0054] Step 6, since the convolutional neural network has good multi-level feature extraction ability and generalization ability, therefore, the convolutional neural network is used to extract the sample features of the training set. As Figure 1 shown, the extracted turbine blade sample features are connected to the input layer of the BP neural network through the convolutional neural network flattening layer, and then sequentially connected to the hidden layer of the BP neural network and the output layer of the BP neural network.

[0055] Step 7, in order to improve the training efficiency and prediction accuracy of the convolutional neural network - BP neural network model, an improved seagull optimization algorithm is used to optimize the parameters of the BP neural network, and a CNN-MAOS-BP neural network model is established, where CNN is the convolutional neural network and MAOS is the improved seagull optimization algorithm.

[0056] Here, the process of the improved seagull optimization algorithm includes a migration stage and a predation stage. The seagull group can make a spiral natural shape movement when attacking. First, initialize the positions of the seagull group. In the migration stage, the seagulls need to ensure that they do not collide with each other, and their new positions are:

[0057]

[0058] In the formula, represents the position to avoid being collided, represents the position at the current iteration number, A represents the migration behavior within the given search space, and A = f c -(t * (f c / Max iteration )); f C is the frequency for controlling variable A, t is the current iteration number, and Max itertion represents the maximum number of iterations.

[0059] Furthermore, the seagull will move closer to the position of the best individual, and there is

[0060]

[0061] In the formula, is the direction for the individual to search and iterate towards the best position, is the current best position.

[0062] The parameter B is a random parameter used to balance local search and global search, and B has:

[0063] B = 2 × A 2 × rd(3)

[0064] The seagull will move in the direction of the optimal individual, and the new position it reaches is:

[0065]

[0066] In the formula, is the direction for the individual to search and iterate towards the best.

[0067] To expand the search space of the seagull optimization algorithm, a scaling coefficient R1 is introduced during local search and global search, then the random parameter B for balancing local search and global search can be updated to:

[0068] B = 2 × A 2 × R1(5)

[0069] Among them, the scaling coefficient R1 is taken as 2.

[0070] Furthermore, to improve the global search ability of the seagull optimization algorithm, the position where the seagull moves towards the best individual can be updated to:

[0071]

[0072] Among them, c is a variable related to the seagull's search behavior, and c = 1 - A.

[0073] During the predation stage, when the seagull attacks the prey with a spiral motion in the air, there are:

[0074] x = rcos(θ)

[0075] y = rsin(θ) (7)

[0076] z = rθ

[0077] r = u×e θv

[0078] In the formula, r is the radius of the spiral, u and v are related constants controlling the spiral behavior, and k is a random number in the range [0, 2π].

[0079] The new position after attacking the prey is:

[0080]

[0081] Furthermore, to improve the iteration speed of the seagull optimization algorithm, increase the search step size in the early stage of the search, and reduce the search step size in the later stage of the search to improve the solution accuracy. A linearly decreasing inertia weight factor is introduced to obtain the new position of the seagull:

[0082]

[0083] Among them, w is the inertia weight factor, which can be calculated by the formula to obtain, w0 represents the weight of the initial position, and w0 = 1.2 is taken.

[0084] Step 8, use the test set samples to verify the prediction accuracy of the CNN - MAOS - BP neural network model. If the accuracy of the established CNN - MSOA - BP neural network model does not meet the requirements, return to Step 7. If it meets the accuracy requirements, then go to Step ⑨. As Figures 3 - 5 shown, they are respectively the comparison results of the low - cycle fatigue stress, strain and high - cycle fatigue stress predicted by the test set and the proposed model.

[0085] Step 9, according to the distribution characteristics of the random input variables, use the Monte Carlo method to randomly extract 10,000 groups of samples, and obtain the corresponding 10,000 groups of low - cycle fatigue and high - cycle fatigue output responses through the CNN - MSOA - BP neural network model.

[0086] Step 10. Since the turbine blade is subjected to complex alternating loads during service, in order to further analyze its fatigue behavior and reliability, the actual load spectrum is simplified into a high-low cycle composite fatigue load spectrum under experimental conditions. Among them, the low-cycle fatigue generated by centrifugal load and temperature load under the main cycle of the aero-engine is represented by a trapezoidal wave, and the high-cycle fatigue load caused by aerodynamic load during the cruise stage is represented by a triangular wave. As Figure 2 shown.

[0087] After calculating the low-cycle fatigue stress and low-cycle fatigue strain according to the CNN-MSOA-BP neural network model, substitute them into the model with mean stress correction proposed by Morrow to calculate the corresponding low-cycle fatigue life:

[0088]

[0089] where, σ m is the mean stress, Δε is the strain range, N L is the low-cycle fatigue life, E is the elastic modulus, σ′ f is the fatigue strength coefficient, b is the fatigue strength index, ε' f is the fatigue ductility coefficient, and c is the fatigue ductility index.

[0090] After calculating the high-cycle fatigue stress according to the CNN-MSOA-BP neural network model, use the stress-life relationship to calculate the corresponding high-cycle fatigue life:

[0091] N H S β = C (11)

[0092] In the formula, N H is the high-cycle fatigue life, S is the high-cycle stress amplitude, and β and C are constants fitted through material fatigue data.

[0093] Furthermore, introduce the linear damage rule into the high-low cycle composite fatigue load, that is, calculate the low-cycle fatigue damage and high-cycle fatigue damage generated during each cycle respectively, and the total fatigue cumulative damage can be obtained;

[0094]

[0095] where, D(N) is the cumulative fatigue damage generated by N cycles, N H is the high-cycle fatigue life, and n is the high-low cycle frequency ratio.

[0096] Furthermore, substitute the low-cycle fatigue life and high-cycle fatigue life into Equation (14) to calculate the fatigue cumulative damage of the turbine blade under the high-low cycle composite load.

[0097] Step 11: Establish its performance function based on the fatigue cumulative damage under high-low cycle composite loads, which can be represented by the difference between the damage strength parameter and the fatigue cumulative damage, i.e.:

[0098] z = a - D (13)

[0099] where a is the damage strength parameter, a is taken as 1, z > 0 represents the safe domain, and z ≤ 0 represents the failure domain; z is the fatigue cumulative damage, and D is the cumulative fatigue damage generated by cycles.

[0100] Furthermore, substituting Equation (13) into the performance function, we get;

[0101]

[0102] According to the performance function, count the number of samples in the safe domain and calculate the reliability of the turbine blade under high-low cycle composite loads:

[0103]

[0104] where I r (z i ) is the exponential function, z i is the performance function, which is equal to 1 when the performance function is greater than or equal to 0, and otherwise equal to 0; M r is the number of samples in the safe domain, and M is the total number of samples.

[0105] Furthermore, the reliability can be calculated and the reliability curve can be plotted, as Figure 6 shown.

[0106] To verify the effectiveness of the proposed high-low cycle composite fatigue reliability analysis method for turbine blades based on neural networks, the reliability of the turbine blades predicted by the CNN-MAOS-BP neural network model in this method is compared with the reliabilities predicted by Monte Carlo simulation, convolutional neural network, and BP neural network models. The results show that the reliability calculated by the method proposed in this invention has higher accuracy compared with the Monte Carlo method, as Figure 6As shown. The method for analyzing the high-low cycle composite fatigue reliability of turbine blades based on neural network proposed by the present invention takes into account the uncertainties of material parameters and applied loads, extracts features from irregular data samples through a convolutional neural network, and then establishes a CNN-MAOS-BP neural network model with a BP neural network optimized by an improved seagull optimization algorithm, while considering the effects of high-cycle fatigue and low-cycle fatigue on the reliability of turbine blades. Therefore, the method for analyzing the high-low cycle composite fatigue reliability of turbine blades based on neural network proposed by the present invention has high reliability analysis accuracy and calculation efficiency. It should be noted that the method for analyzing the high-low cycle composite fatigue reliability of turbine blades based on neural network proposed by the present invention is not limited to the turbine blades of aero-engines under high-low cycle composite loads, and the reliability analysis of rotor components of aero-engines under complex loads is within the protection scope of the present invention.

[0107] Those skilled in the art should understand that the above embodiments are only for clearly explaining the present invention, rather than limiting the scope of the present invention. For those skilled in the art, other changes or modifications can be made based on the above disclosure, and these changes or modifications are still within the scope of the present invention.

Claims

1. A method for analyzing the high - low cycle composite fatigue reliability of turbine blades based on neural network, characterized in that, It includes the following steps: Step 1: Establish a finite element model of the turbine blade, and perform deterministic finite element analysis based on the finite element model to obtain the dangerous parts; Step 2: Select random variables and determine their distribution characteristics; Step 3: Use the Latin hypercube sampling method to extract a small number of input samples for low-cycle fatigue and high-cycle fatigue respectively; Step 4: Substitute the extracted input samples into the finite element model for analysis to obtain the output samples of the turbine blade under low-cycle fatigue and high-cycle fatigue loads; Step 5: Divide the input samples and output samples into a training set and a test set according to a certain ratio; Step 6: Use a convolutional neural network to extract the sample features of the training set and use them as the input of the BP neural network; Step 7: Use the improved seagull optimization algorithm to optimize the parameters of the BP neural network and establish a CNN-MAOS-BP neural network model; Step 8: Verify the prediction accuracy and operation efficiency of the CNN-MAOS-BP neural network model; Step 9: According to the distribution characteristics of the random input variables, use the Monte Carlo method to extract a large number of samples and obtain the output responses of low-cycle fatigue and high-cycle fatigue through the CNN-MAOS-BP neural network model; Step 10: Predict the high-low cycle composite fatigue life and damage of the turbine blade under multi-source uncertainty according to the output response; Step 11: Establish a performance function for the high-low cycle composite fatigue of the turbine blade, count the number of samples in the safe domain, and calculate the reliability of the turbine blade under the high-low cycle composite load.

2. The method for analyzing the high-low cycle composite fatigue reliability of a turbine blade based on a neural network according to claim 1, wherein In Step 3, according to the uncertainty factors of the material properties and applied loads of the turbine blade during service, density, elastic modulus, turbine blade speed, thermal conductivity, and temperature are used as random variables.

3. The method for analyzing the high-cycle and low-cycle composite fatigue reliability of a turbine blade based on a neural network according to claim 1, wherein In Step 6, a convolutional neural network is used to extract the features of the training set and use them as the input parameters of the BP neural network.

4. The method for analyzing the high-cycle and low-cycle composite fatigue reliability of turbine blades based on a neural network according to claim 1, wherein In Step 7, the random parameter B for balancing exploration and exploitation in the improved seagull optimization algorithm is: B = 2 × A 2 × R1 Where A is the introduced additional variable and R1 is the scaling coefficient.

5. The method for analyzing the high-cycle and low-cycle composite fatigue reliability of a turbine blade based on a neural network according to claim 1, characterized in that, In step 7, the new position where the improved seagull optimization algorithm moves towards the optimal individual is as follows: Among them, Positions to avoid being collided with, Indicates the position at the current iteration number, and c is a variable related to the seagull search behavior.

6. The method for analyzing the high-cycle and low-cycle composite fatigue reliability of turbine blades based on a neural network according to claim 5, wherein In step 7, improve the position of the prey after the seagull optimization algorithm attacks the prey It is: Among them, w is the inertia weight factor, and x, y, and z are the coordinates when the seagull attacks the prey with a spiral motion behavior in the air. is the current best position.

7. The method for analyzing the high-cycle and low-cycle composite fatigue reliability of turbine blades based on a neural network according to claim 1, wherein In Step 10, the linear damage rule is extended to the high-low cycle composite fatigue load, that is: Among them, D(N) is the cumulative fatigue damage of N composite blocks, N L is the low-cycle fatigue life, N H is the high-cycle fatigue life, and n is the high-low cycle frequency ratio.

8. The method for analyzing the high-low cycle composite fatigue reliability of turbine blades based on a neural network according to claim 7, characterized in that, In Step 11, the performance function of the turbine blade under the high-low cycle composite fatigue load is: Where a is the damage intensity parameter and N is the number of cycles.

9. The method for analyzing the high-low cycle composite fatigue reliability of a turbine blade based on a neural network according to claim 1, characterized in that In Step 11, the reliability of the turbine blade under multi-source uncertainty in the high-low cycle composite fatigue load can be obtained by counting the number of samples in the safe domain, that is: Among them, I r (z i ) is an exponential function, z i is a functional function, M r is the number of samples within the security domain, and M is the total number of samples.

Images