Turbine blade global mechanical property inversion method driven by data fusion

Abstract

The invention discloses a data fusion driven turbine blade global mechanical property inversion method, which belongs to the field of material science and technology and calculation, and constructs a layered correction data system: high-temperature nanoindentation experiment data is the highest fidelity reference, 3D finite element simulation is high fidelity, and 2D finite element simulation is low fidelity. The elastic-plastic training parameter range is set as follows: the hardening index is 0.1-0.3, the elastic modulus is 100-230 GPa, and the yield stress is 500-1400 MPa. A cascading correction strategy is adopted, a 3D simulation result is corrected through high-temperature experimental data, the precision of low-fidelity 2D data is improved through corrected high-fidelity 3D data, data fusion is achieved through residual feature learning, and the optimal ratio of 2D simulation data to 3D simulation data is determined under the condition of the minimum loss value. The mechanical properties of different parts of the turbine blade have differences, the method provides a new direction for characterization of the global mechanical properties of the turbine blade, and the accuracy and calculation efficiency of nanoindentation mechanical property inversion are improved through efficient data fusion and proportion optimization.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of material science and technology and calculation, and relates to a material attribute inversion analysis method for nanoindentation testing in mechanical property characterization of a turbine blade. BACKGROUND

[0002] As a core component of an aero-engine, the mechanical properties of a turbine blade directly affect the safety and reliability of the engine. Due to the microstructure differences in different areas of the blade, such as the root, tip, thin area and thick area, and the alloy fluidity, and the complex temperature gradient and stress field during the service of the blade, the mechanical properties of different parts are significantly different. Therefore, the overall mechanical property characterization of the blade is needed to evaluate its overall performance state. However, due to the size limitation and the difference between the test sample and the actual shape and size of the blade, the traditional tensile test cannot give the overall service state of the blade under different stress (the maximum stress of the blade is at the root, and decreases towards the tip) and temperature (the maximum temperature of the blade is in the middle of the blade body) conditions. In contrast, nanoindentation testing, as a key material mechanical property characterization method, has been widely used in metal material mechanical property evaluation. It can accurately obtain the mechanical response of the microstructure at the micro / nano scale, and then characterize and describe the overall mechanical properties of the blade. The core principle of nanoindentation inversion technology is to apply load to the material at the nano scale and accurately measure the indentation depth of the indenter, so as to indirectly invert the key mechanical properties of the material such as elastic modulus, yield strength and material hardening index according to the load-displacement curve. However, the traditional nanoindentation material attribute inversion method still faces many challenges in practical application: (1) Large difference in overall performance of turbine blade: Due to the complex geometry and service environment of the turbine blade, the mechanical properties vary significantly from the root to the tip and from the leading edge to the trailing edge. The material hardening index, elastic modulus and yield stress vary with position. The traditional experimental method is difficult to cover such a wide range of performance, and cannot realize real overall characterization.

[0003] (2) Difficulty in obtaining experimental data: High-temperature nanoindentation experiments can provide the highest fidelity material response data, but the experimental conditions are harsh and the cost is high. It is difficult to obtain a large amount of experimental data covering the overall performance range of the blade due to factors such as sample preparation and environmental control.

[0004] (3) High cost of high-fidelity simulation: 3D finite element model can provide high-precision indentation behavior simulation, but the calculation resource consumption is large and time-consuming. When a large amount of simulation data is needed to cover the overall performance range of the blade, this problem is particularly prominent, which greatly limits the generation of high-quality simulation data.

[0005] (4) Low-fidelity simulation accuracy is insufficient: 2D finite element model has high calculation efficiency, but oversimplifies the three-dimensional effect in the indentation process, and the simulation accuracy is often insufficient to meet the high requirements of turbine blade global performance characterization, which may introduce large errors.

[0006] (5) Lack of effective correction mechanism for multi-source data fusion: Existing methods cannot establish an effective correction relationship between experimental data, 3D simulation data and 2D simulation data, cannot fully play the correction role of high-fidelity experimental data, and the prediction accuracy is limited.

[0007] (6) Challenges in model robustness and precision improvement: Traditional machine learning methods cannot effectively capture the deep correlation between multi-source heterogeneous data, lack of hierarchical correction mechanism, resulting in insufficient robustness of the model when facing actual test data.

[0008] In summary, the current nanoindentation material property inversion technology needs to build an innovative method that takes high-temperature experimental data as the highest fidelity benchmark and effectively fuses multi-fidelity data through a hierarchical correction mechanism. In terms of data fusion strategy, a multi-level correction system needs to be established, in which experimental data corrects 3D simulation, and corrected 3D prediction corrects 2D simulation. In terms of model architecture, a technical solution needs to be built to realize multi-fidelity data parallel learning and residual correction in a unified network framework, minimize the prediction difference between different fidelity data, and achieve high-fidelity prediction accuracy. SUMMARY

[0009] To solve the above technical problems, the present application provides a turbine blade global mechanical property inversion method driven by data fusion, which realizes multi-level precision improvement through building a hierarchical residual neural network to correct 3D simulation with experimental data, and correct 2D simulation with corrected 3D simulation.

[0010] The present application is implemented by the following technical solutions: (1) Turbine blade global multi-source data generation Obtain high-temperature nanoindentation experimental data, perform nanoindentation test under turbine blade service temperature conditions, and obtain high-fidelity P-h curve as benchmark data; Establish a finite element simulation model for nanoindentation test, set the elastic-plastic parameter range covering different parts of the turbine blade for global mechanical property characterization: the material hardening index value range is 0.1-0.3, the elastic modulus value range is 100-230 GPa, and the yield stress value range is 500-1400 MPa, to ensure coverage of global performance characteristics from blade root to blade tip, from leading edge to trailing edge; After setting input parameters and boundary conditions, simulating the nanoindentation process and performing calculation, a result curve of the nanoindentation, i.e. a P-h curve (load-displacement curve), is obtained, the P-h curve is processed to obtain required parameter data, and then 2D low-fidelity simulation data and 3D high-fidelity simulation data are obtained, wherein the low-fidelity (2D) simulation data is simulated by a 2D finite element model, and a multi-angle indenter is selected for simulation (50°, 60°, 70°, 80°), and the high-fidelity (3D) simulation data is simulated by a 3D finite element model; The 2D and 3D simulation data are ensured to have high independence, and effective complementation of data is realized; (2) Residual neural network architecture construction The residual network architecture is built, mainly involving how to build the residual module, and learning of two residual blocks will be used to measure the difference in prediction effect of different fidelity data, which is embedded into the neural network in the form of a skip connection. The residual network learns residual mapping F(x) = H(x) - x by introducing a skip connection, rather than directly learning target mapping H(x), which effectively solves the gradient vanishing problem in deep network training. The present application extends the residual learning concept to the multi-fidelity data fusion scene, and constructs a hierarchical correction residual network specially used for learning the systematic difference between different fidelity data; the core idea of the hierarchical correction residual network can be described by the following mathematical expression:

[0011] Among them, represents high-temperature experimental data, represents 3D simulation data, represents 2D simulation data, and are correction functions learned by the first layer and the second layer residual network respectively; the formula embodies the core concept of Low-fidelity + Residual Learning = High-fidelity, that is, the fidelity is improved by learning the residual between low-fidelity and high-fidelity prediction. The residual network R(x) establishes a nonlinear mapping relationship from input features to predicted residuals, so that the corrected low-fidelity prediction can approach the high-fidelity prediction accuracy, thereby obtaining a prediction result close to high-fidelity data without increasing the cost of obtaining high-fidelity data.

[0012] This design strategy embodies the essence of multi-fidelity modeling: using a small amount of high-fidelity data to guide systematic error correction of a large amount of low-fidelity data.

[0013] The multi-fidelity finite element simulation data described above is input into the residual neural network for training, and first, the full connection neural network training of 2D and 3D data is respectively carried out to obtain the prediction accuracy. The fusion comparison and correction of different fidelity data are realized through residual feature learning, so as to improve the robustness and prediction accuracy of the model; (3) Data quantity matching The best matching of 2D and 3D finite element data is determined under the condition of the minimum loss value (LOSS) of the residual network; (4) Material property inversion Based on the trained residual neural network model, the corrected fusion data set is obtained, and the deep neural network is back-propagation trained under the Tensorflow framework using the data set; The mechanical properties of the material are obtained through neural network inversion, such as but not limited to elastic modulus, material hardening index, yield stress, etc.

[0014] Advantages of the present application (1) Realize the global performance characterization of turbine blades: by setting the range of elastic-plastic parameters covering the whole turbine blade, the mechanical performance difference of different parts of the blade can be effectively characterized, and comprehensive support is provided for the overall performance evaluation of the blade.

[0015] (2) Realize true high-fidelity prediction: taking high-temperature experimental data as the highest fidelity benchmark, through a hierarchical correction mechanism, the material property inversion accuracy is significantly improved. (3) Establish an innovative multi-level correction system: a hierarchical correction mechanism of experimental data→3D simulation→2D simulation is constructed, the advantages of different fidelity data are fully utilized, and the best balance between efficiency and accuracy is realized.

[0016] (4) Significantly reduce the demand for experimental data: through the residual learning mechanism, a small amount of high-temperature experimental data can effectively correct a large amount of simulation data, greatly reducing the experimental cost and time consumption.

[0017] (5) Improve the robustness and engineering applicability of the model: the hierarchical residual network architecture enhances the adaptability of the model to complex working conditions, provides reliable technical support for the global performance characterization of turbine blades, and has important engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 For the two-dimensional axisymmetric conical nanoindentation finite element model in the embodiment Figure 2 For the stress cloud map top view of the two-dimensional axisymmetric model in the embodiment Figure 3 For setting different indenter angles for the two-dimensional axisymmetric model in the embodiment Figure 4 The three-dimensional Berkovich indenter is a 1 / 6 symmetric finite element model in this embodiment. Figure 5 This is a top view of the stress cloud diagram of the 3D 1 / 6 model in the embodiment. Figure 6 The following is a description of the fitting results of training three material property parameter models using a 3D data DNN in the example. Figure 7 The LOSS curve (Average) of the 3D data DNN training parameter model in this embodiment is shown. Figure 8 The following is an example of how a two-dimensional data DNN trains a model to fit three material property parameters. Figure 9 The LOSS curve (Average) of the two-dimensional data DNN training parameter model in this embodiment. Figure 10 The confidence intervals for two-dimensional and three-dimensional data are displayed in the embodiment. Figure 11 This example illustrates the setup and connection of residual blocks in the residual network. Figure 12 In this example, the residual network training loss and validation loss are LOSS (including five-fold cross-validation). Figure 13 This example shows a comparison of residual distributions after the residual network has been trained. Figure 14 σ after residual network correction in the example y Predictive effect Figure 15 The ratio of 2D and 3D data is shown in the example where the residual network LOSS value is minimized. Detailed Implementation

[0019] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be further described below in conjunction with specific embodiments and accompanying drawings. However, the embodiments described are only used to illustrate the principles and technical effects of the present invention and are not intended to limit the scope of protection of the present invention.

[0020] The finite element simulation methods (including the construction and solution of two-dimensional axisymmetric models and three-dimensional solid models), high-temperature nanoindentation experimental methods, neural network modeling methods (including the construction of multi-layer residual network structures, model training and evaluation), and nanoindentation-related parameter extraction methods involved in the embodiments of this invention are all conventional techniques well known to those skilled in the art, unless otherwise stated. The simulation software (ABAQUS), deep learning framework (TensorFlow), high-temperature nanoindentation experimental equipment, and computing resources used in the embodiments are all publicly available unless otherwise specified.

[0021] This embodiment aims to verify the effectiveness and superiority of the nanoindentation inversion material property analysis method based on "high-temperature nanoindentation-multi-fidelity simulation data fusion residual neural network" proposed by the present application in the global mechanical property characterization of turbine blades. The specific steps are as follows: This example aims to predict the global performance of three key material properties of turbine blades, namely material hardening index n, elastic modulus E, and yield stress σy. To achieve global performance characterization of turbine blades, according to the actual mechanical property variation characteristics of different parts of turbine blades, the material property parameter range covering the whole blade is set: the material hardening index range is 0.1-0.3, the elastic modulus range is 100-230 GPa, and the yield stress range is 500-1400 MPa. Systematic integer values are taken within the value range and combined scientifically to generate material property combinations used in finite element simulation nanoindentation, ensuring that the global performance characteristics of turbine blades from the blade root to the blade tip and from the leading edge to the trailing edge are covered.

[0022] I. Global multi-fidelity experiment-simulation data generation (1) Obtain high-temperature nanoindentation experimental data High-temperature nanoindentation experiments are performed to obtain the mechanical response data of turbine blade materials under actual service temperature conditions. The experiment uses a Berkovich indenter to perform high-temperature nanoindentation testing on high-temperature alloy samples, including 30 indentation experiments at 600°C and 30 indentation experiments at 800°C.

[0023] Through the load control mode, the true load-displacement curve is obtained, and the relevant mechanical response characteristic parameters (C, dP / dh, and Wp / Wt, where C, dP / dh, and Wp / Wt are the loading curvature, initial unloading slope, and ratio of residual plastic work to total work, respectively) are extracted. Experimental data have the highest fidelity and can truly reflect the actual mechanical behavior of materials in a high-temperature environment, but the cost of obtaining data is high and the amount of data is limited. Process the experimental data, and obtain the high-temperature experimental data vector set

n, E, σy, C, dP / dh, Wp / Wt

[0024] (2) Establish a finite element simulation model to obtain a finite element dataset According to the experimental process and related settings of nanoindentation testing, a nanoindentation finite element simulation model is established in the finite element software ABAQUS, and the geometry of the indenter and the material being tested is selected. In the simulation data obtained in this embodiment, the indenter angle conditions of 50°, 60°, 70°, and 80° are included.

[0025] Set input parameters and boundary conditions. The material behavior is defined using an elastic-plastic constitutive relationship, with Young's modulus E and Poisson's ratio v defined in the elastic stage, and yield stress σy The power-law hardening curve is determined by the hardening exponent n. The indentation load-displacement curve for each set of data is extracted through finite element simulation. The curves are then processed to obtain relevant mechanical response characteristic parameters (C, dP / dh, Wp / Wt, where C, dP / dh, and Wp / Wt are the loading curvature, initial unloading slope, and the ratios of the calculated residual plastic work and total work, respectively). This forms a system containing material constitutive parameters (E, σ...). y A complete dataset of (n) and response feature parameters.

[0026] (3) Generate 2D low-fidelity simulation data The nanoindentation process was simulated using a 2D finite element model, as shown in the attached figure. Figure 1 The diagram shows a two-dimensional axisymmetric conical nanoindentation finite element model. The 2D model employs a plane strain assumption, simplifying the three-dimensional effects during the indentation process to reduce computational cost. While the two-dimensional axisymmetric conical indenter simulation model is computationally efficient, its significant simplification allows it to only approximately characterize the actual indentation deformation behavior of the material. The resulting stress distribution exhibits a concentric annular pattern, as shown in the attached diagram. Figure 2 The stress cloud diagram, shown in the top view of a two-dimensional axisymmetric model swept 60° around the Y-axis, differs somewhat from the actual three-dimensional indentation. This can be achieved by changing the indenter's half-apex angle (see attached diagram). Figure 3 By setting different indentation angles for the two-dimensional axisymmetric model, the sensitivity of the load-displacement curve to changes in material parameters is enhanced, thereby improving the dataset's ability to infer material performance parameters. A total of 367 sets of two-dimensional simulation data were obtained in this stage, resulting in 367 sets of data vectors [n, E, σ]. y ,C,dP / dh,Wp / Wt】。

[0027] (4) Generate 3D high-fidelity simulation data A 3D finite element model was used to simulate the nanoindentation process. The 3D model can more accurately reflect the three-dimensional mechanical behavior during indentation, and the model geometry closely matches the actual experiment. In the 3D model, a solid Berkovich indenter was used; however, a one-sixth of it was simplified for the analysis to improve computational efficiency (see appendix). Figure 4 Appendix Figure 5 These are a 1 / 6 symmetric finite element model of the three-dimensional Berkovich indenter and a top view of the stress contour plot. The solid indenter can effectively capture the local stress concentration effect near the tip and corner, resulting in more realistic and reliable simulation results, but the computational cost is higher. A total of 64 sets of three-dimensional simulation data were obtained in this stage, resulting in 64 sets of data vectors [n, E, σ]. y ,C,dP / dh,Wp / Wt】。

[0028] (5) Ensure high data independence By comparing the high-temperature experiment, 3D simulation and 2D simulation data, it is ensured that the three have high independence. High independence means that the three kinds of data have different characteristics and advantages in reflecting the mechanical behavior of materials, and can effectively complement each other. The high-temperature experiment data has the highest precision but limited quantity, the 3D data has moderate precision and moderate quantity, and the 2D data has limited precision but the highest calculation efficiency. This high independence provides a basis for subsequent multi-layer data fusion and model training.

[0029] Therefore, for 2D and 3D data, independent neural network models are trained respectively, as shown in FIGS. 1, 2, 3, 4, 5 and 6. Figure 6 Figure 7 Figure 8 Figure 9 The fitting and the Loss curve of the model training using the 2D and 3D data sets to train the DNN model are shown in FIGS. 1, 2, 3, 4, 5 and 6, respectively. From the Loss curve, the training of the model can be obtained, including convergence and whether there is overfitting and other problems. From the R 2 and MAPE values, it can be known whether the function fitting of the model is good and whether the precision of the model is acceptable. Subsequently, the confidence interval of the simulation data is compared to further check whether the fidelity conforms to the theoretical basis, as shown in FIGS. 7 and 8. Figure 10 From the area corresponding to the two fidelity data sets in the figure, although the 3D data is only 64, its fidelity is higher than that of the 367 2D data. Finally, it is confirmed that the fidelity gradient of the high-temperature experiment data > 3D simulation data > 2D simulation data.

[0030] (6) Features and targets Since the present application is an inversion work, the feature parameters are extracted from the load-displacement curve, and the material properties are inverted. Therefore, the three parameters C, dP / dh and Wp / Wt extracted from the load-displacement curve are the features in the embodiments, and n, E and σ y are the three turbine blade material parameter targets.

[0031] II. Residual neural network architecture construction (1) The residual network architecture includes a plurality of residual blocks, each residual block includes two fully connected layers, and uses a jump connection to realize residual learning. The residual block architecture is shown in FIG. 9. Figure 11 Figure 11 is a residual block connection between 2D and 3D. In addition, the network introduces a self-attention mechanism and a data enhancement technique to improve the robustness and prediction accuracy of the model. The specific steps are as follows: Define the input layer, and the input feature dimension is 3 (corresponding to the 3 features in the data set); Apply global normalization to standardize the input data to speed up the training process and improve the stability of the model; ​​​​Two residual blocks were constructed, each containing two fully connected layers with 64 and 32 neurons, respectively, using the He uniform initializer (kernel_initializer='he_uniform') and L2 regularization (kernel_regularizer=tf.keras.regularizers.l2(0.0003)). Batch normalization (BatchNormalization), ELU activation function (tf.keras.layers.ELU(alpha=1.0)), and Dropout layer (tf.keras.layers.Dropout(0.1)) were applied after each fully connected layer to prevent overfitting. The input of each residual block was directly passed to the output through a skip connection, and the output of the skip connection was added to the output of the fully connected layer using tf.keras.layers.Add() from the Tensorflow framework to form the final output of the residual block. A self-attention mechanism was introduced to reshape the output of the second residual block into a sequence form, and a multi-head self-attention mechanism was implemented using tf.keras.layers.MultiHeadAttention with num_heads=4 and key_dim=8. A Dropout layer was added to prevent overfitting. The output of the self-attention mechanism was flattened and concatenated with the output of the second residual block (tf.keras.layers.Concatenate()) to enhance the model's attention to important features. A fully connected layer with 16 neurons was added, using the He uniform initializer and L2 regularization, to provide more stable input for the output layer. The output layer contains one neuron that outputs the predicted value.

[0032] A custom loss function was defined, combining Huber loss and mean squared error loss to balance the model's robustness to outliers and overall prediction accuracy, using the Adam optimizer.

[0033] (2) Residual preprocessing and data augmentation Preprocess residuals using the preprocess_residuals_v3 function. First, calculate the absolute value of the residuals and determine the clipping value based on the specified percentile. Then, use the tan function to smooth the residuals that exceed the clipping value to avoid the impact of outliers on model training. Finally, add small random noise (np.random.normal(0, 0.01, size=clipped_residuals.shape)) to enhance the robustness of the model. Use the data_augmentation function to augment the data. First, randomly select a portion of the samples based on the specified augmentation factor (augmentation_factor=0.3) and duplicate these samples. Then, add small random noise to the features of the duplicated samples, with a noise ratio of 0.05. Finally, combine the original data and augmented data to form the data fusion dataset.

[0034] III. Simulation data ratio optimization (1) Simulation data ratio optimization goal 2D data has high computational efficiency but limited accuracy; 3D data has high accuracy but high computational cost. To balance efficiency and accuracy while ensuring prediction accuracy, we need to find the optimal ratio of 2D and 3D data.

[0035] (2) Simulation data ratio optimization method Define the ratio range of 2D data and 3D data as np.arange(0.2, 1.0, 0.2). This means that the ratio of 2D data and 3D data will be searched from 20% to 80% with a step size of 0.2. To avoid randomness, the number of repeated training is set to 3. To find the ratio, supplement 184 groups of 3D data. For each ratio of 2D and 3D data, extract the corresponding proportion of data from the training set as training samples. Use these samples to train the base model (2D model and 3D model) and generate prediction values. Calculate the residual between the 2D prediction value and the interpolated 3D prediction value, and preprocess and augment the residual. Train the residual network model using the preprocessed residual and use the loss value on the validation set as the evaluation standard.

[0036] (3) Training and results In multiple repeated training, for each ratio of 2D and 3D data, train the base model and generate residuals. Train the residual network model using the residual and record the loss value on the validation set.Figure 12 The training and validation losses of the residual network show that the model training converges well. Repeating this process multiple times, the average loss value for each data ratio is calculated. By comparing the average loss values ​​under different ratios, the optimal 2D and 3D data ratio is determined.

[0037] Under the global performance characterization conditions of the turbine blades in this example, according to the attached... Figure 15 The results showed that when the residual network loss value was minimized, the ratio of 2D to 3D data was approximately 6:1, meaning that 2D data accounted for 6 / 7 of the total 2D+3D data. At this point, the residual network could learn the minimum residual connections. This ratio significantly reduced computational costs while maintaining prediction accuracy, providing a scientifically sound optimization strategy for characterizing the global mechanical properties of turbine blades.

[0038] IV. Verification of Turbine Blade Material Property Inversion The residual neural network model trained through the above steps successfully achieved accurate inversion of the global mechanical properties of the turbine blade. The results of the example can be found in the appendix. Figure 13 Appendix Figure 14 Analysis was conducted, including the attached Figure 13 This is a comparison of the residual distribution after the residual network is trained. It can be seen that after training, the data distribution is more symmetrically distributed on both sides of the baseline compared to the original residual distribution. (Attached) Figure 14 Based on the yield stress σ y For example, let's compare the prediction and correction results using 3D corrected data after the residual network is applied, and the prediction results using the original 2D data for training σ. y The performance improvement is evident. While the material hardening exponent n and elastic modulus E did not fit the model training results as well as yield stress on the original dataset, the improvement after residual network correction is clearly visible. Therefore, to further illustrate the advanced effectiveness of this method, yield stress is used as an example for explanation and result demonstration. Yield stress fits the original dataset quite well, proving that the model is not overfitting. The original R... 2 The value was 0.9996, which improved to 0.9998 after 3D correction, an increase of 0.0002. For the original highly fitted model, this improvement demonstrates the effectiveness of model correction. Validation results show that this method can effectively handle the global performance prediction of material hardening index, elastic modulus, and yield stress within the upper and lower threshold ranges, providing reliable technical support for predicting the mechanical properties of different parts of turbine blades.

Claims

1. A data fusion driven turbine blade global mechanical property inversion method, characterized in that, The method comprises the following steps: (1) turbine blade global multi-fidelity data generation: 2D low-fidelity and 3D high-fidelity simulation data of the turbine blade are generated through a finite element simulation model, and data fusion is performed in combination with high-temperature nanoindentation experiment data to obtain a multi-fidelity data set with high independence and mutual complementation; (2) constructing a hierarchical correction residual neural network framework: a hierarchical correction residual network architecture for multi-fidelity data fusion is built, high-temperature experiment data are used to correct 3D simulation data, and the accuracy of 2D data is improved through optimized 3D data, so that the system error correction and fusion between different fidelity data are realized through a residual block learning mechanism; (3) data quantity ratio optimization: the ratio between 2D and 3D simulation data is optimized to minimize the loss function Loss, and the best data ratio for inversion work is determined; (4) turbine blade material property inversion: based on the fusion data set obtained by the trained hierarchical correction residual neural network model, inversion training is performed through a deep neural network to obtain the global mechanical properties of the turbine blade, including the elastic modulus, material hardening index and yield stress.

2. The method of claim 1, wherein the method is driven by data fusion. In the step (1), the high-temperature nanoindentation experiment needs to be performed under the high-temperature service condition of the turbine blade material (the high-temperature nanoindentation experiment temperature is set to 600-800 DEG C). ​ 3. The method of claim 2, wherein the method is driven by data fusion. The high-temperature nanoindentation experiment is performed on different parts of the turbine blade to obtain local mechanical property data from the blade root to the blade tip and from the leading edge to the trailing edge, so as to realize the characterization of the global mechanical properties of the turbine blade. ​ 4. The data fusion driven turbine blade global mechanical property inversion method of claim 1, wherein, In the step (1), the 2D simulation data are generated by a two-dimensional finite element model, and different indenter angles in the range of 50 DEG to 80 DEG are discretely simulated to simulate the nanoindentation process; the 3D simulation data are generated by a three-dimensional finite element model, and a six-simplified model of a Berkovich solid indenter is used for simulation, wherein the elastic-plastic parameters of the finite element simulation are in the range of: material hardening index 0.1-0.3, elastic modulus 100-230 GPa, and yield stress 500-1400 MPa.

5. The data fusion driven global mechanical performance inversion method of turbine blades according to claim 1, characterized in that, In the step (2), the hierarchical correction residual network comprises two layers of residual learning modules, and the residual network adopts a skip connection mode to learn residual mapping F(x) = H(x) - x, instead of directly learning the target function H(x). The core idea of the hierarchical correction residual network is described by the following mathematical expression: wherein, represents high temperature experimental data, represents 3D simulation data, represents 2D simulation data, and are correction functions learned by the first and second residual networks, respectively.

6. The data fusion driven turbine blade global mechanical property inversion method of claim 5, wherein, The hierarchical correction residual neural network comprises a multi-head attention mechanism structure after the second layer of residual learning modules, and the attention mechanism structure comprises a multi-head attention layer, a Dropout layer and a splicing layer.

7. The data fusion driven global mechanical performance inversion method of turbine blades according to claim 1, characterized in that, In the step (3), the data ratio optimization is to first train the residual neural network and optimize the hyperparameters, and then determine the optimal ratio of 2D simulation data and 3D simulation data by minimizing the loss function in the residual neural network training process, wherein the high-temperature experiment data are used as a fixed correction reference and do not participate in the ratio optimization.

8. The data fusion driven turbine blade global mechanical property inversion method of claim 7, wherein, The residual neural network adopts a self-defined loss function, and the loss function comprises a Huber loss term and a mean square error (MSE) loss term.

9. An electronic device, comprising: It comprises: (1) a memory for storing a computer program; (2) a processor configured to execute a computer program to cause the electronic device to perform the step of implementing a global mechanical performance nanoindentation inversion analysis of a turbine blade based on a high-temperature nanoindentation-multiple-fidelity simulation data fusion residual neural network according to any one of claims 1 to 8.

Images