A turbine blade dynamic stress field prediction method based on hybrid data training

Abstract

This invention discloses a method for predicting the dynamic stress field of turbine blades based on hybrid data training, relating to the field of turbine blade technology. The method involves acquiring a computational dataset calculated using a numerical model of blade dynamic stress and a measurement dataset obtained from experimental measurements of blade dynamic stress; acquiring an initial prediction model, which includes a first sub-model and a second sub-model; jointly training the first and second sub-models using the computational dataset based on a joint loss function to obtain a pre-trained model; the joint loss function is a weighted sum of the training losses of the first and second sub-models; freezing the weights in the pre-trained model whose absolute values ​​exceed an amplitude retention threshold, and fine-tuning the pre-trained model using the measurement dataset to obtain a blade dynamic stress field prediction model; and using this model to predict the blade's dynamic stress field in real time. This method achieves high-precision prediction of the blade's dynamic stress field.

Description

Technical Field

[0001] This application relates to the field of turbine blade technology, and in particular to a method for predicting the dynamic stress field of turbine blades based on mixed data training. Background Technology

[0002] With the increasing demand for higher efficiency and power in modern electric energy and industrial power fields, the load on turbine blades is constantly increasing. Ensuring the safe and reliable operation of turbine blades under complex loads is of great significance.

[0003] Blade vibration is a crucial component of turbine operational safety and reliability research. In engineering, the vibration response parameters (displacement, dynamic stress, etc.) of blades are often kept within allowable values ​​as a design and monitoring criterion. Dynamic stress is typically measured by attaching strain gauges to critical test locations on the blade surface. However, strain gauge measurement has unavoidable limitations: low measurement point survival rate under harsh industrial conditions; strain gauges and leads affecting blade safety; and the ability to obtain only localized, discrete data, with experiments unable to cover complex conditions, resulting in measured dynamic stress failing to reflect the overall distribution of the dynamic stress field. While numerical calculations can obtain full-field data, their accuracy cannot match that of experimental measurements.

[0004] Therefore, how to integrate high-value but sparse measured data with low-cost but wide-coverage simulation data to achieve accurate prediction of dynamic stress fields has become an urgent problem to be solved. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for predicting the dynamic stress field of turbine blades based on hybrid data training to address the above-mentioned technical problems.

[0006] The following technical solution is adopted in this specification: This specification provides a method for predicting the dynamic stress field of turbine blades based on mixed data training, including: Obtain the computational dataset calculated using the blade dynamic stress numerical model and the measurement dataset obtained from the blade dynamic stress test. Both the computational and measurement datasets include the turbine blade research parameters, as well as the corresponding dynamic strain and dynamic stress data. The blade research parameters include material parameters and load parameters. Obtain the initial prediction model; the initial prediction model includes a first sub-model and a second sub-model; the first sub-model is used to predict dynamic strain using blade study parameters, and the second sub-model is used to predict the dynamic stress field using material parameters and the dynamic strain predicted by the first sub-model; Based on the joint loss function, the first sub-model and the second sub-model are jointly trained using a computational dataset to obtain a pre-trained model; the joint loss function is a weighted sum of the training losses of the first sub-model and the second sub-model. The weights in the pre-trained model whose absolute values ​​are greater than the amplitude retention threshold are frozen, and the pre-trained model is fine-tuned using a measurement dataset to obtain a blade dynamic stress field prediction model. The dynamic stress field of the blade is predicted in real time using a blade dynamic stress field prediction model.

[0007] Optionally, the training process for jointly training the first sub-model and the second sub-model includes a fast training phase, a core optimization phase, and a fine-grained convergence phase, with the learning rate used to update the model parameters decreasing sequentially in the fast training phase, the core optimization phase, and the fine-grained convergence phase.

[0008] Optionally, freeze the weights in the pre-trained model whose absolute values ​​are greater than the magnitude retention threshold, including: Obtain the weight matrix of the pre-trained model; Weights in the weight matrix whose absolute value is greater than the magnitude retention threshold are set to the frozen state; the frozen state means that fine-tuning updates using the measurement dataset are not allowed. Weights in the weight matrix whose absolute value is less than or equal to the magnitude retention threshold are set to the active state; the active state indicates that fine-tuning updates can be performed using the measurement dataset.

[0009] Optionally, the fine-tuning loss function used when fine-tuning the pre-trained model with the measurement dataset. for: ; in, K This refers to the number of actual test measurement points; This is the position mask matrix. The dimension of the position mask matrix is ​​the same as the dimension of the dynamic stress field output by the model. The elements in the position mask matrix are marked as 1 at the measurement point position and 0 for the rest. The dynamic stress field predicted by the model; The dynamic stress measurement values ​​are mapped to the entire field, with non-measurement point locations set to 0. It represents the Hadamardi (or Hadama) stack; This represents the L2 norm.

[0010] Optionally, obtain the computational dataset calculated using the blade dynamic stress numerical model and the measurement dataset obtained from the blade dynamic stress test, including: Based on the actual service conditions of the blade and the corresponding three-dimensional high-fidelity geometric model of the blade-disk, a numerical model of blade dynamic stress is established; the numerical model of blade dynamic stress adopts a hybrid mesh generation strategy of denser mesh in key areas and sparser mesh in non-critical areas. The types of blade research parameters and the sampling space are determined. The Latin hypercube sampling method is used to sample the blade research parameters in the sampling space. The blade dynamic stress numerical model is used to perform numerical calculations on all sample points. The calculation results of blade dynamic strain and dynamic stress field are extracted. All calculation results and corresponding blade research parameters are integrated to obtain the calculation dataset. Determine and arrange strain gauge measurement points, and establish a dynamic strain data acquisition system; Based on the blade's service conditions and requirements, rated operating conditions and several typical operating conditions were selected as test conditions. Tests were conducted, and the dynamic strain at the blade strain gauge measurement points under each test condition was obtained through a dynamic strain data acquisition system. The dynamic stress at the corresponding measurement points was calculated, and all measurement results and corresponding test condition parameters were integrated to obtain a measurement dataset for blade dynamic stress test measurement. Each test condition corresponds to a blade research parameter.

[0011] Optionally, the strain gauge measurement points are determined and arranged, including: Based on the constraints of strain gauge arrangement and the dynamic strain distribution results of the blade calculated by the blade dynamic stress numerical model, the measurement points of the strain gauges are determined. The strain gauges are fixed to the measurement points of the blade by pasting or welding, thus completing the strain gauge measurement point arrangement for the blade dynamic strain test. Each strain gauge measurement point directly corresponds to a local position of the mesh in the blade dynamic stress numerical model, and there is a correspondence between the strain gauge measurement points and the blade mesh nodes.

[0012] Optionally, the method further includes: The measurement dataset is screened in multiple dimensions through a dual verification process. The dual verification includes signal integrity verification and repeatability verification. Signal integrity verification includes removing abnormal test samples with data interruption or amplitude abrupt change. Repeatability verification involves conducting multiple repeated tests on the test conditions, selecting several test samples with stable test conditions during the test process, and then calculating the average value as the test result to remove test samples with measurement data deviation caused by fluctuations in the test conditions.

[0013] Optionally, before training the model using the computational dataset and the measurement dataset, the method further includes: The computational and measurement datasets are normalized, and the normalized datasets are used for model training.

[0014] Optionally, the dynamic stress field of the blade is predicted in real time using a blade dynamic stress field prediction model, including: Acquire real-time monitoring data of the blades during the actual operation of the unit, i.e., the monitoring values ​​corresponding to the blade research parameters; The monitored values ​​were normalized using the same normalization parameters for both the computed and measured datasets. The normalized monitoring values ​​are input into the blade dynamic stress field prediction model to obtain the predicted blade dynamic stress field. The predicted dynamic stress field is inversely normalized and mapped onto the three-dimensional finite element mesh nodes of the blade to obtain the dynamic stress field distribution cloud map of the blade at the current moment.

[0015] Optionally, the first sub-model includes an input layer and a feature extraction layer, and the second sub-model includes a feature fusion layer, a feature reconstruction layer, and an output layer; the normalized monitoring values ​​are input into the blade dynamic stress field prediction model to obtain the predicted blade dynamic stress field, including: The normalized monitoring values ​​are input into the first sub-model of the blade dynamic stress field prediction model. After passing through the input layer and the feature extraction layer, the dynamic strain at the measurement point of the strain gauge on the blade surface is output. The normalized monitoring values ​​and the dynamic strain output from the first sub-model of the blade dynamic stress field prediction model are input into the second sub-model of the blade dynamic stress field prediction model. After passing through the feature fusion layer, feature reconstruction layer and output layer, the predicted blade dynamic stress field is obtained.

[0016] This specification provides a device for predicting the dynamic stress field of turbine blades based on hybrid data training, including: The first acquisition module is used to acquire the calculation dataset calculated by the blade dynamic stress numerical model and the measurement dataset measured by the blade dynamic stress test. Both the calculation dataset and the measurement dataset include the turbine blade research parameters, as well as the corresponding dynamic strain and dynamic stress data. The blade research parameters include material parameters and load parameters. The second acquisition module is used to acquire the initial prediction model. The initial prediction model includes a first sub-model and a second sub-model. The first sub-model is used to predict dynamic strain using blade research parameters, and the second sub-model is used to predict the dynamic stress field using material parameters and the dynamic strain predicted by the first sub-model. The pre-training module is used to jointly train the first sub-model and the second sub-model using a computational dataset based on the joint loss function to obtain a pre-trained model; the joint loss function is a weighted sum of the training losses of the first sub-model and the second sub-model. The fine-tuning module is used to freeze the weights in the pre-trained model whose absolute values ​​are greater than the magnitude retention threshold, and to fine-tune the pre-trained model using the measurement dataset to obtain the blade dynamic stress field prediction model. The prediction module is used to predict the dynamic stress field of the blade in real time using the blade dynamic stress field prediction model.

[0017] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the dynamic stress field of turbine blades based on mixed data training.

[0018] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described method for predicting the dynamic stress field of turbine blades based on mixed data training.

[0019] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: The method for predicting the dynamic stress field of turbine blades based on hybrid data training provided in this specification obtains computational and measurement datasets by constructing a numerical model of blade dynamic stress and conducting dynamic strain measurement experiments. An intelligent prediction model of the blade dynamic stress field is built based on deep learning methods. The model is trained and fine-tuned using hybrid data, incorporating high-precision measurement data to compensate for deviations between purely numerical calculation data and actual operating conditions. This fully leverages the synergistic effect of the two types of data, achieving rapid and high-precision prediction of parameters from materials and loads to the overall dynamic stress distribution of the blade. Compared to traditional methods, the improved technical approach of this invention offers advantages such as greater adaptability, higher prediction accuracy, faster response speed, higher feasibility for on-site deployment, and lower long-term operation and maintenance costs. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0021] Figure 1 A schematic diagram of the process for predicting the dynamic stress field of a turbine blade based on mixed data training, as provided in this specification; Figure 2 This document provides a schematic diagram of a two-step intelligent prediction model for the dynamic stress field of a turbine blade. Figure 3 This is a schematic diagram of an intelligent prediction model for the dynamic stress field of a turbine blade trained using mixed data, as provided in this specification. Figure 4 This specification provides a schematic diagram of a three-dimensional model of a compressor blade-disc using the method of the present invention; Figure 5 This specification provides a schematic diagram of a strain data acquisition system utilizing the method of the present invention; Figure 6 This specification provides a schematic diagram of the strain gauge measurement point arrangement for a compressor blade using the method of the present invention; Figure 7 This specification provides a schematic diagram of an intelligent prediction model for the dynamic stress field of a compressor blade using the method of the present invention. Figure 8A schematic diagram illustrating the prediction of dynamic stress in compressor blades using the method of the present invention, provided in this specification. Figure 9 This specification provides a schematic diagram of a computer device for implementing a method for predicting the dynamic stress field of turbine blades based on mixed data training.

[0022] Explanation of reference numerals in the attached figures: 1. Compressor blade-disc test piece; 2. Conductive slip ring; 3. Dynamic strain gauge; 4. Computer; 5. Biaxial strain gauge; 51. Grid wire; 52. Substrate. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0024] The limitations of existing strain gauge measurement methods include: (1) Effective strain gauge measurement depends on a stable installation environment and signal transmission channel, which can only be achieved under controlled laboratory conditions. In actual industrial settings, blades are usually under harsh conditions, the survival rate of measurement points is low, and the arrangement of strain gauges and leads will affect the performance and safety reliability of the blades. It is not feasible to arrange strain gauges on blades for real-time dynamic stress measurement; (2) "High-quality" test data depends on strict replication of operating conditions. It is extremely difficult to completely replicate the complex and variable operating conditions of industrial sites (such as variable load start-up and shutdown, extreme operating condition impact). Usually, the test can only cover a few typical operating conditions, and the operating condition coverage of the data is insufficient; (3) Due to the spatial limitations of the blade structure, the number of strain gauges that can be pasted on the blade is very limited, and it is impossible to achieve full coverage of the dynamic stress field on the blade surface. Only local discrete data can be obtained, which is difficult to reflect the overall distribution law of the dynamic stress field. However, strain gauge measurement data has an accuracy that is difficult to achieve by numerical calculation. How to effectively utilize these valuable and limited test data is an important issue.

[0025] Therefore, there is an urgent need to develop a method and device for predicting the dynamic stress field of turbine blades based on hybrid data training. This method obtains a large amount of low-cost dynamic stress data of the blades through numerical calculations under all operating conditions, conducts dynamic stress tests on the blades, realizes dynamic stress measurement of turbine blades under typical operating conditions, obtains a small amount of high-precision experimental measurement data, and establishes an intelligent prediction model of the dynamic stress field of turbine blades based on deep learning. The model is trained using hybrid data from numerical calculations and experimental measurements. This method integrates low-cost, wide-coverage numerical calculation data with high-precision, limited experimental measurement data, and applies artificial intelligence methods for learning. It can achieve rapid and accurate prediction of the dynamic stress field of turbine blades in actual units, and has significant engineering application value.

[0026] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0027] Figure 1 This is a flowchart illustrating a method for predicting the dynamic stress field of turbine blades based on hybrid data training, as described in this specification. The method includes the following steps: S101, obtain the calculation dataset calculated by the blade dynamic stress numerical model and the measurement dataset measured by the blade dynamic stress test; both the calculation dataset and the measurement dataset include the turbine blade research parameters, as well as the corresponding dynamic strain and dynamic stress data; the blade research parameters include material parameters and load parameters.

[0028] In one embodiment, obtaining the computational dataset calculated using a blade dynamic stress numerical model and the measurement dataset obtained from blade dynamic stress experiments includes the following steps: S201. Based on the actual service conditions of the blade and the corresponding three-dimensional high-fidelity geometric model of the blade-disk, a numerical model of blade dynamic stress is established. The numerical model of blade dynamic stress adopts a hybrid mesh generation strategy of denser mesh in key areas and sparse mesh in non-critical areas.

[0029] A high-fidelity three-dimensional geometric model of the blade-disk was established, fully preserving complex and key structures such as the blade root groove and the blade surface, thus avoiding calculation errors caused by geometric simplification.

[0030] A hybrid meshing strategy of "refining in critical areas and sparsening in non-critical areas" is adopted. Hexahedral structured meshes are used to refine the mesh in areas prone to stress concentration, such as the blade root, leading and trailing edges, and wheel grooves. This meshing method ensures computational accuracy while improving numerical computation efficiency. Based on the actual service conditions of the blades and rotor, the boundary conditions and load application methods are determined, and a numerical model of blade dynamic stress is constructed to ensure that the dynamic stress of the blade under the actual working conditions can be accurately simulated.

[0031] S202. Determine the blade research parameter type and sampling space. Use the Latin hypercube sampling method to sample the blade research parameters in the sampling space. Then, use the blade dynamic stress numerical model to perform numerical calculations on all sample points. Extract the calculation results of blade dynamic strain and dynamic stress field. Integrate all calculation results and corresponding blade research parameters to obtain the calculation dataset.

[0032] Based on the influencing factors of the blade's dynamic stress field, material parameters and load parameters are selected as blade research parameters. X The sampling space is formed by determining the range of values ​​for each parameter based on the actual service and operating conditions of the blade.

[0033] A sample set was constructed using the Latin hypercube sampling method. Based on the blade dynamic stress numerical model, numerical calculations were performed on all samples to extract the blade dynamic strain and dynamic stress field calculation results. All results and corresponding sample research parameters were integrated to obtain the blade dynamic stress numerical calculation dataset.

[0034] The numerical calculation dataset for blade dynamic stress includes blade research parameters as model inputs. The dynamic strain field and dynamic stress field of the blade obtained through numerical calculation are used as labels. .

[0035] Optionally, when determining the range of values ​​for each blade research parameter, both typical and extreme working conditions should be considered to ensure the integrity and representativeness of the sampling space, thus guaranteeing the diversity of the dataset and the generalization ability of the model.

[0036] Optionally, a lightweight result extraction strategy can be adopted. When the result extraction is performed after the calculation of a single sample point, only the dynamic strain and dynamic stress data of the blade surface nodes are retained, which compresses the amount of single sample data, reduces the pressure of subsequent data processing, and solves the problems of data redundancy and large storage pressure in batch numerical calculations.

[0037] In this embodiment, the Latin hypercube sampling method can achieve uniform coverage of the entire parameter space with a limited number of sampling points. It can capture extreme tail condition samples that are easily missed by general random sampling, ensuring that the sampling space is not significantly biased, and providing high-quality and highly representative parameter combinations for numerical calculation.

[0038] S203. Determine and arrange strain gauge measurement points to establish a dynamic strain data acquisition system.

[0039] Design a blade dynamic stress measurement scheme: Design a blade test scheme that includes blade dynamic strain measurement and dynamic strain signal acquisition.

[0040] In the experiment, strain gauges were used to measure the dynamic strain of the blade during the test. In one embodiment, determining and arranging the strain gauge measurement points includes: determining the measurement point positions of the strain gauges based on the strain gauge arrangement constraints and the blade dynamic strain distribution results calculated by the blade dynamic stress numerical model; fixing the strain gauges to the measurement point positions on the blade by pasting or welding, thus completing the strain gauge measurement point arrangement for the blade dynamic strain test; each strain gauge measurement point directly corresponds to a local position in the blade dynamic stress numerical model, and there is a correspondence between the strain gauge measurement points and the blade mesh nodes.

[0041] A dynamic strain data acquisition system is established using strain gauges, conductive slip rings, a DC regulated power supply, a dynamic strain meter, and a computer.

[0042] S204. Based on the blade's service conditions and requirements, rated operating conditions and multiple typical operating conditions are selected as test conditions. Tests are conducted, and the dynamic strain at the blade strain gauge measurement points under each test condition is obtained through a dynamic strain data acquisition system. The dynamic stress at the corresponding measurement points is calculated, and all measurement results and corresponding test condition parameters are integrated to obtain a measurement dataset for blade dynamic stress test measurement. Each test condition corresponds to a set of blade research parameters.

[0043] The measurement dataset for the blade dynamic stress test includes the blade's operating parameters and material parameters under different operating conditions as model inputs. The dynamic strain and dynamic stress at the blade measuring points, measured by strain gauges, are used as labels. .

[0044] Optionally, to improve the accuracy of test data, a multi-dimensional screening of the measurement dataset is performed through dual verification. The dual verification includes signal integrity verification and repeatability verification. Signal integrity verification includes removing abnormal test samples with data interruption or amplitude abrupt change. Repeatability verification includes conducting multiple repeated tests on the test conditions, selecting several test samples with stable test conditions during the test process, and then calculating the average value as the test result to remove test samples with measurement data deviation caused by fluctuations in the test conditions.

[0045] Each strain gauge measurement point arranged on the test blade directly corresponds to a local part in the finite element mesh model of the blade numerical calculation. The correspondence between the strain gauge measurement point and the blade mesh node can be established. Therefore, the dynamic strain data obtained by the test at each strain gauge measurement point can be directly used for fine-tuning of the subsequent blade dynamic stress field prediction model.

[0046] Furthermore, based on the computational dataset of blade numerical calculations with sufficient data volume and the measurement dataset of high-precision blade experiments, a multi-source data fusion framework is established. The input and output format specifications for mutual adaptation between numerical calculation data and experimental measurement data are clarified. The input data is unified as a combination of all blade research parameters, such as material parameters + load parameters. The output data is defined differently according to type: the numerical calculation data outputs the dynamic strain and dynamic stress matrix of all nodes on the blade surface, and the experimental measurement data outputs the dynamic strain and dynamic stress vector of discrete measurement points.

[0047] In one embodiment, the computational dataset and the measurement dataset are normalized before model training is performed using the computational dataset and the measurement dataset. The normalized dataset is then used for model training.

[0048] A hybrid dataset including blade research parameters, dynamic strain, and dynamic stress data was established based on the normalized data.

[0049] Optionally, based on the characteristics of large quantity of numerical computation data but slightly insufficient accuracy and high accuracy of experimental measurement data but small quantity, the hybrid dataset is divided into numerical training set, numerical validation set, numerical test set, and experimental fine-tuning set. Among them, 70% of the computation dataset is used as the numerical training set, 20% as the numerical validation set, 10% as the numerical test set, and all of the measurement dataset is used as the experimental fine-tuning set, forming a hybrid dataset for blade dynamic stress field prediction that includes numerical samples and experimental samples. This satisfies the data scale requirements for model training while ensuring prediction accuracy through experimental data.

[0050] S102, Obtain the initial prediction model; the initial prediction model includes a first sub-model and a second sub-model; the first sub-model is used to predict dynamic strain using blade research parameters, and the second sub-model is used to predict the dynamic stress field using material parameters and the dynamic strain predicted by the first sub-model.

[0051] An initial prediction model for the dynamic stress field of a blade is constructed based on deep learning methods. Using the input and output parameters of a mixed dataset as a benchmark, the functions and parameters of each network layer are determined, and an initial prediction model including an input layer, a feature extraction layer, a feature fusion layer, a feature reconstruction layer, and an output layer is built to achieve an accurate mapping between the research parameter vector, the dynamic strain at the measurement point, and the dynamic stress field of the blade.

[0052] The initial prediction model adopts a two-step prediction structure, consisting of two sub-models. The first sub-model includes an input layer and a feature extraction layer, while the second sub-model includes a feature fusion layer, a feature reconstruction layer, and an output layer. The first sub-model takes the research parameters as input data and outputs the dynamic strain at the measurement points of the strain gauges on the blade surface after passing through the input layer and the feature extraction layer. The second sub-model takes the material parameters and the dynamic strain at the measurement points as input data and outputs the dynamic stress field on the blade surface after passing through the feature fusion layer, the feature reconstruction layer, and the output layer. Figure 2 As shown, Figure 2 This is a schematic diagram of the initial prediction model for the dynamic stress field of a turbine blade.

[0053] Appropriate deep learning models, such as deep neural networks, convolutional neural networks, and graph convolutional neural networks, can be selected based on the specific structural characteristics and service operation features of the blades being studied. Furthermore, the two sub-models can select different deep learning models, different loss functions, and different hyperparameter settings. Through highly flexible and customized prediction models, the data features can be fully explored and learned, thereby improving the accuracy of blade dynamic field prediction.

[0054] S103, based on the joint loss function, uses a computational dataset to jointly train the first sub-model and the second sub-model to obtain a pre-trained model; the joint loss function is a weighted sum of the training losses of the first sub-model and the second sub-model.

[0055] 70% of the dataset is used as the numerical training set, 20% as the numerical validation set, and 10% as the numerical test set.

[0056] Pre-training the initial prediction model of the dynamic stress field using a numerical training set allows for the selection of either MSE Loss or Smooth L1 Loss as the loss function for the two sub-models. The training loss of the first sub-model is calculated by combining the predicted dynamic strain at the measurement points with the calculated dynamic strain at the measurement points. Loss 1. The training loss of the second sub-model is calculated by predicting the dynamic stress field of the blade and calculating the dynamic stress field of the blade. Loss 2. Integration Loss 1 and Loss 2. Obtain the joint loss of the initial prediction model pre-training. The joint loss function is: In the formula, Loss train The joint loss of the initial prediction model, To lose weight, .

[0057] use Loss train The initial prediction model is trained using the Adam, Adammax, or AdamW optimizer. The initial learning rate is set to 0.001. After each iteration of training, the learning rate is reduced to one-tenth of its original value.

[0058] The pre-trained blade dynamic stress field prediction model was validated on a numerical validation set. To evaluate the model's prediction level of the blade dynamic stress field, two dimensionless evaluation metrics were defined: In the formula, FNAE is the average relative error of the dynamic stress field nodes, reflecting the level of the overall prediction deviation of the dynamic stress field; FNME is the maximum relative error of the dynamic stress field nodes, reflecting the deviation of the dynamic stress field prediction from the upper bound. N This represents the total number of grid nodes on the blade surface. For the first i The first sample m The calculated value of each node; For the first i The first sample m The predicted value of each node; , These are the maximum and minimum values ​​of the global dynamic stress field in the sample set.

[0059] After pre-training is completed, the FNAE, FNME, and coefficient of determination of all samples in the numerical test set are calculated to characterize the overall prediction performance of the test set.

[0060] In one embodiment, the initial prediction model adopts a phased training strategy. The training process of jointly training the first sub-model and the second sub-model includes a fast training phase, a core optimization phase, and a fine convergence phase. The learning rate used when updating the model parameters in the fast training phase, the core optimization phase, and the fine convergence phase decreases sequentially.

[0061] Specifically, a larger learning rate is used in the rapid training phase to allow the model parameters to quickly enter the neighborhood of the optimal solution. In the core optimization phase, the learning rate is appropriately reduced to deeply explore and learn the key features of the dynamic stress field. In the fine convergence phase, the learning rate is further reduced to make small adjustments to the model parameters to achieve stable and high-precision convergence of the prediction results. This strategy aims to synergistically improve the training efficiency and prediction accuracy of the model.

[0062] By jointly training the first and second sub-models using the dataset, the initial prediction model after training is determined as the pre-trained model.

[0063] S104: Freeze the weights in the pre-trained model whose absolute values ​​are greater than the amplitude retention threshold, and fine-tune the pre-trained model using the measurement dataset to obtain the blade dynamic stress field prediction model.

[0064] To address the sparse nature of blade test data and the uneven distribution of measurement points, and to prevent the model from forgetting the dynamic stress features learned during pre-training of the numerical computation dataset during fine-tuning, a physical feature-preserving parameter fine-tuning update strategy for multi-source blade data fusion is designed and adopted. Specifically, this strategy employs a dual-mask constraint fine-tuning approach. First, considering that the measurement dataset only covers a few specific locations on the blade, and that the measurement dataset is sparse measurement point data rather than full-field data, a location mask matrix with the same dimension as the dynamic stress field output by the model is constructed. In the position mask matrix In the process, the location where the measuring point is installed is marked as 1, and the other locations are marked as 0, in order to filter out invalid errors in the non-measuring point area when calculating the loss function during the model fine-tuning process.

[0065] Then, in order to retain the weights that carry the main dynamic stress field features in the pre-trained model during the fine-tuning process, an intra-layer sparse fine-tuning strategy based on the gradient mask algorithm is designed. By constructing the gradient mask matrix of network layer parameters, the weight parameters that carry the main physical features (parameters with large absolute weight values) are frozen, and only some parameters are released for learning the measurement dataset.

[0066] Optionally, freezing weights in the pre-trained model whose absolute values ​​are greater than the magnitude retention threshold includes: obtaining the weight matrix of the pre-trained model; setting weights in the weight matrix whose absolute values ​​are greater than the magnitude retention threshold to a frozen state; the frozen state indicates that fine-tuning updates using the measurement dataset are not allowed; setting weights in the weight matrix whose absolute values ​​are less than or equal to the magnitude retention threshold to an active state; the active state indicates that fine-tuning updates using the measurement dataset are allowed.

[0067] Specifically, for each trainable network layer in the pre-trained model l Its weight matrix is ​​denoted as Construct a weight matrix Binary parametric mask matrices of the same dimension Set an amplitude preservation threshold. Identify The absolute value is greater than The parameters representing the main physical characteristics are labeled as the "frozen state," that is, their corresponding... Position set to 0; Recognition The absolute value is less than or equal to The parameters are marked as "active states", that is, their corresponding The position is set to 1, allowing it to be fine-tuned and updated using measurement data.

[0068] Read the weights of the final pre-trained model These parameters were used as initial weights and loaded into the blade dynamic stress field prediction model. The model was then fine-tuned using the measurement dataset, incorporating the blade research parameters from the measurement dataset. Input the model to obtain the predicted dynamic stress field. .

[0069] To address the sparsity of the measurement data, an output position mask matrix is ​​used. Calculate the effective error between the predicted values ​​and the measured values, and define the loss function when fine-tuning the pre-trained model using the measurement dataset. Unlike the full-field loss in the pre-training phase, the fine-tuning loss only calculates the prediction error at the measurement point location. The fine-tuning loss function... for: in, K This refers to the number of actual test measurement points; This is the position mask matrix. The dimension of the position mask matrix is ​​the same as the dimension of the dynamic stress field output by the model. The elements in the position mask matrix are marked as 1 at the measurement point position and 0 for the rest. The dynamic stress field predicted by the model; The dynamic stress measurement values ​​are mapped to the entire field, with non-measurement point locations set to 0. It represents the Hadamardi (or Hadama) stack; This represents the L2 norm.

[0070] Based on fine-tuning loss function Calculate the original gradient of each layer parameter Before updating the parameters, the parameter gradient mask matrix is ​​used. Filter the gradient: Then, the optimizer is used to update the model parameters: In the formula, For the model number The gradient matrix of the weight parameters of the layer can be finely adjusted and updated. To fine-tune the learning rate, t and t +1 indicates the current time and the next time. for t Time Model The weight matrix of the layer, for t +1 time model Layer weight matrix.

[0071] The fine-tuning training of the blade dynamic stress field prediction model uses the Adam, Adammax, or AdamW optimizer to fine-tune the learning rate. Set it to one-tenth of the final learning rate of the pre-training.

[0072] This embodiment designs a fine-tuning and updating strategy for physical feature-preserving parameters of blade multi-source data fusion (a double-mask constraint fine-tuning strategy) during the fine-tuning process. This is achieved by constructing a position mask matrix of strain gauge measurement points on the blade surface. The prediction model performs error backpropagation only on high-precision measurement data, eliminating the interference of unmeasured regions on model parameter updates. Through an intra-layer sparse fine-tuning strategy, a gradient mask matrix of network layer parameters is constructed, allowing a portion of parameters to be fine-tuned and updated in each layer of the model. This not only solves the problem that small sample experimental data can easily lead to catastrophic forgetting in deep learning models, but also adapts to the specific distribution of real physical fields through subtle corrections, achieving a smooth transition and improved accuracy from simulation calculation to experimental measurement.

[0073] Based on the above training process, such as Figure 3 As shown, Figure 3 A schematic diagram illustrating the training of a dynamic stress field prediction model for turbine blades using mixed data.

[0074] S105 uses a blade dynamic stress field prediction model to predict the blade's dynamic stress field in real time.

[0075] Specifically, the final trained blade dynamic stress field prediction model is deployed to the edge computing unit or cloud monitoring platform of the turbine, and connected to the real-time monitoring data during the actual operation of the unit to achieve real-time prediction of the full-field dynamic stress of the blade under service conditions.

[0076] In one embodiment, the dynamic stress field of a blade is predicted in real time using a blade dynamic stress field prediction model, including: acquiring real-time monitoring data of the blade during actual operation of the unit, i.e., monitoring values ​​corresponding to the blade research parameters; performing the same normalization processing on the monitoring values ​​using the normalization parameters of the computational dataset and the measurement dataset; inputting the normalized monitoring values ​​into the blade dynamic stress field prediction model to obtain the predicted blade dynamic stress field; performing inverse normalization processing on the predicted dynamic stress field and mapping it onto the three-dimensional finite element mesh nodes of the blade to obtain the dynamic stress field distribution cloud map of the blade at the current moment.

[0077] In one embodiment, the normalized monitoring values ​​are input into the blade dynamic stress field prediction model to obtain the predicted blade dynamic stress field. This includes: inputting the normalized monitoring values ​​into the first sub-model of the blade dynamic stress field prediction model, passing through an input layer and a feature extraction layer, and outputting the dynamic strain at the measurement point of the strain gauge on the blade surface; inputting the normalized monitoring values ​​and the dynamic strain output by the first sub-model of the blade dynamic stress field prediction model into the second sub-model of the blade dynamic stress field prediction model, passing through a feature fusion layer, a feature reconstruction layer, and an output layer to obtain the predicted blade dynamic stress field.

[0078] The dynamic stress field prediction model for turbine blades is built using a data-driven deep learning method. It is obtained through pre-training and fine-tuning training using mixed data. It can achieve fast and accurate end-to-end prediction of dynamic stress field from blade material parameters and load parameters to the entire dynamic stress field of the blade surface without complex modeling and time-consuming numerical calculations. It can be used for online real-time monitoring of blade dynamic stress field in industrial sites, and can also support further blade safety assessment, life damage calculation, blade vibration control, etc.

[0079] Furthermore, once the intelligent prediction model for the dynamic stress field of turbine blades is trained, it only needs to collect monitoring data and material parameters from the unit's operation section. Without the need to deploy measuring equipment such as strain gauges, the dynamic stress field of the blades can be obtained. This can effectively reduce the difficulty and cost of on-site monitoring, assist on-site operation and maintenance, and ensure the safe operation of the unit.

[0080] In a specific embodiment, the prediction method provided by the present invention is used to predict the dynamic stress field of a gas turbine compressor blade, as follows: 1. A three-dimensional model of a gas turbine compressor blade-disc is shown below. Figure 4 As shown, a meshing strategy of denser meshing in critical areas and sparser meshing in non-critical areas is adopted. Stress concentration areas such as blade root, blade leading and trailing edges, and wheel grooves are denser with hexahedral structured meshes.

[0081] Based on the actual operating characteristics of the compressor, a numerical calculation model of the dynamic stress field of the compressor blade is established by setting displacement constraints, periodic symmetry constraints, blade root groove contact boundary conditions and rotational speed, and airflow excitation force load.

[0082] 2. The blade material density, elastic modulus, Poisson's ratio, rotational speed, and airflow excitation force were selected as the research parameters. Based on the actual service conditions of the blade, the range of values ​​for each parameter was determined to form the sampling space, as shown in Table 1.

[0083] Table 1 Sampling space of research parameters The Latin hypercube sampling method was used to sample various research parameters in the sampling space to obtain a sample set of 2000. The research parameters of each sample were input into the blade dynamic stress numerical model for numerical sampling calculation. The dynamic strain and dynamic stress calculation results of the blade were extracted. All results and corresponding sample research parameters were integrated to obtain the blade dynamic stress numerical calculation dataset.

[0084] 3. Design the blade test scheme. The strain data acquisition system mainly consists of strain gauges, conductive slip rings, a DC regulated power supply, a dynamic strain gauge, and a computer. For example... Figure 5As shown, the strain data acquisition system specifically includes a compressor blade-disc test piece 1, a conductive slip ring 2, a dynamic strain gauge 3, a computer 4, and a biaxial strain gauge 5. The biaxial strain gauge 5 includes a grid wire 51 and a substrate 52. The strain gauge uses a 1 / 4 bridge connection and is attached to the part of the blade with a large strain amplitude. The strain gauge wire is led out through the central hole to the slip ring, which is connected to the dynamic strain gauge. The slip ring outputs a voltage signal to the data acquisition unit, which processes the received dynamic signal and transmits the result to the computer for subsequent data analysis.

[0085] To more intuitively display the physical quantity information of the mesh nodes on the blade surface, the three-dimensional curved surface of the blade is unfolded from the leading edge and laid out as a two-dimensional plane. The positions of the strain gauge measurement points correspond one-to-one with the nodes of the numerical calculation model. The strain gauges on the blade surface are arranged as follows: Figure 6 As shown, the strain gauge measuring points on the suction surface of the blade are numbered #1 and #2, and the measuring point on the pressure surface of the blade is numbered #3.

[0086] 4. Referring to the rated operating conditions and typical operating conditions of the blade in service, 50 sets of parameters were selected as test conditions; test results were carried out, and measurement data of the blade strain gauge measuring points under each test condition were obtained through a dynamic strain data acquisition system. The measurement signals were preprocessed through signal integrity verification and test repeatability verification. The dynamic strain data of the measuring point position can be obtained using the preprocessed measurement signals, and then the dynamic stress data of the measuring point can be calculated. All test measurement results and corresponding test condition parameters were integrated to obtain the blade dynamic stress test measurement dataset.

[0087] 5. Standardize the format of the numerical calculation data and experimental measurement data of blade dynamic stress. The input data is a combination of blade material parameters and load parameters, and the output data is blade dynamic strain and dynamic stress. The numerical calculation data is the dynamic strain and dynamic stress matrix of all nodes on the blade surface, and the experimental measurement data is the dynamic strain and dynamic stress vector of discrete measurement points. At the same time, normalize the input and output data to establish a hybrid dataset containing blade research parameters, dynamic strain and dynamic stress field data.

[0088] The hybrid dataset is divided into a numerical training set, a numerical validation set, and an experimental fine-tuning set. 70% of the numerical computation data is used as the numerical training set, 20% as the numerical validation set, and 10% as the numerical test set. All experimental measurement data is used as the experimental fine-tuning set, forming a hybrid dataset for predicting the dynamic stress field of the blade, which includes numerical samples and experimental samples.

[0089] 6. Construct a blade dynamic stress field prediction model based on convolutional neural networks, such as... Figure 7 As shown.

[0090] The first sub-model is the dynamic strain prediction model at the blade measuring point (NN1), which adopts a 2D CNN structure. The input is material parameters and load parameters, and the output is the dynamic strain at the measuring point. ReLU is used as the activation function, and a batch normalization layer is set after each convolutional layer to stabilize the training process, speed up the convergence speed, and reduce the sensitivity to weight initialization.

[0091] The second sub-model is the blade dynamic stress field prediction model (NN2), which also adopts a 2D CNN structure. The input is the dynamic strain of the measurement point node and the elastic modulus of the blade material output by NN1, and the output is the dynamic stress field of the blade surface. ReLU is used as the activation function, and a batch normalization layer is set after each convolutional layer. However, since there are fewer input variables and more output variables, it is necessary to increase the dimensionality of the data. In order to avoid information loss, the CNN network of NN2 does not set a pooling layer.

[0092] The dynamic stress field prediction model was pre-trained using the numerical training set in the mixed dataset. The loss function for NN1 was selected as mean square error (MSE), and the loss function for NN2 was selected as Huber Loss. The loss of NN1 was calculated by predicting the dynamic strain at the measuring points and calculating the dynamic strain at the measuring points. Loss 1. Calculate NN2 loss by predicting the blade dynamic stress field and calculating the blade dynamic stress field. Loss 2. Integration Loss 1 and Loss 2. Obtain the joint loss of the pre-trained prediction model. Loss train .

[0093] Use Loss train The dynamic stress field prediction model for the blade was trained using the Adama optimizer. The initial learning rate was set to 0.001, and the training was iterated for 300 steps. Every 100 steps of the training was iterated, and the learning rate was reduced to one-tenth of the original rate.

[0094] The pre-trained blade dynamic stress field prediction model was validated on a numerical validation set, and the model's prediction level of blade dynamic stress field was evaluated using FNAE and FNME.

[0095] After pre-training is completed, the FNAE, FNME and determination coefficients of all samples in the numerical test set are calculated to characterize the overall prediction performance of the test set, as shown in Table 2.

[0096] Table 2 Overall prediction results of the intelligent prediction model test suite for dynamic stress field of compressor blades As can be seen from the table, the maximum FNAE is 5.340 × 10⁻⁶. -3 The maximum FNME is 1.826 × 10 -2The mean coefficient of determination for the predicted dynamic stress field of the test set samples was 0.9963, indicating that the error between the predicted field and the actual field is very small, and the prediction of the dynamic stress field of the blade is relatively accurate.

[0097] 7. Using the experimental fine-tuning set in the mixed dataset, the physical feature preservation parameter fine-tuning update strategy for blade multi-source data fusion provided by this invention is used to fine-tune the blade dynamic stress field prediction model. Specifically, it is a fine-tuning strategy with dual mask constraints.

[0098] First, construct a position mask matrix with the same dimensions as the dynamic stress field output by the model. In the mask matrix In the diagram, the location where the measuring point is installed is marked as 1, and the other locations are marked as 0.

[0099] Then, using an intra-layer sparse fine-tuning strategy of the present invention, a layer with the weight matrix is ​​constructed. Binary parametric mask matrices of the same dimension Weight parameters for each layer of the model Sort the parameters from largest to smallest, and use the top 60% of the largest values ​​as the amplitude retention threshold. Identify The absolute value is greater than The parameters representing the main physical characteristics are labeled as the "frozen state," that is, their corresponding... Position set to 0; Recognition The absolute value is less than or equal to The parameters are marked as "active states", that is, their corresponding The position is set to 1, allowing it to be fine-tuned and updated using experimental data.

[0100] The model is fine-tuned using the fine-tuning training method in step S104 with the experimental fine-tuning set in the mixed dataset.

[0101] After completing the fine-tuning training, the compressor blade dynamic stress field prediction model was tested. It only takes 0.1 seconds to predict the dynamic stress field. The dynamic stress field prediction results are as follows: Figure 8 As shown, the predicted dynamic stress field cloud map ((a)), the actual dynamic stress field cloud map ((b)), and the absolute error cloud map ((c)) are plotted respectively. It can be seen that the predicted dynamic stress distribution is highly consistent with the actual distribution, and the absolute error is about two orders of magnitude smaller than the actual dynamic stress value, realizing the rapid and accurate prediction of the dynamic stress field of the compressor blade.

[0102] 8. Deploy the blade dynamic stress field prediction model, fine-tuned based on experimental data, to the online monitoring platform of the gas turbine. Integrate real-time monitoring data of load parameters during actual unit operation, retrieve compressor blade material parameters, and obtain real-time monitoring data of load parameters during actual unit operation through the unit's distributed control system or condition monitoring system. Combine these data to construct the current moment's dynamic stress field prediction model. t Input feature vector Using the normalization parameters of the mixed dataset to Perform the same normalization process to obtain the standard input for the prediction model; then normalize the input... Input the finely tuned blade dynamic stress field prediction model to quickly predict the blade dynamic stress field. ,right Inverse normalization is performed and mapped onto the 3D finite element mesh nodes of the blade, thus reconstructing the current moment. t Dynamic stress field distribution cloud map of the lower blade.

[0103] The method for predicting the dynamic stress field of turbine blades based on hybrid data training provided by this invention also has the following advantages: The intelligent prediction model for blade dynamic stress field adopts a two-step prediction structure, which decomposes the dynamic stress field prediction into dynamic strain prediction at local measurement points and dynamic stress prediction across the entire field. Different deep learning models, loss functions, and hyperparameter settings can be selected for the two sub-models according to the specific characteristics of the blade structure and service operation. Through a highly flexible and customized prediction model, the model fully explores and learns data features, thereby improving the accuracy of blade dynamic field prediction.

[0104] Furthermore, compared to pure experimental data-driven schemes and pure numerical calculation data-driven schemes, the hybrid data-driven approach provided by this invention significantly reduces data acquisition costs while ensuring high accuracy of the prediction model. It first uses a large amount of numerical calculation data with low accuracy to pre-train the model, and then uses high-precision experimental measurement data to fine-tune the model. This approach retains the global feature learning results of the blade dynamic stress field from the numerical data, and further improves the prediction accuracy of the model through fine-tuning training with experimental data.

[0105] Furthermore, the present invention designs and employs a physical feature preservation parameter fine-tuning update strategy for multi-source data fusion of blades. Its dual-mask strategy has a unique synergistic effect. The position mask matrix solves the spatial mismatch problem between the numerical simulation data of dynamic stress field and the sparse experimental measurement data, enabling the model to directly use sparse and unevenly distributed experimental data for learning and training. The network layer parameter gradient mask matrix solves the catastrophic forgetting problem that is easily caused by fine-tuning of small sample experimental data, realizing a smooth transition from simulation calculation to experimental measurement and improving accuracy.

[0106] Furthermore, the gradient mask matrix of the network layer parameters of the present invention enables sparse fine-tuning within the layer. Compared with the traditional method of freezing part of the network layer as a whole, it allows the correction information of the experimental data to penetrate into each layer of the prediction model. It can more finely calibrate the difference between numerical simulation and actual operation while preserving the physical mechanism of numerical simulation, thereby achieving higher model prediction accuracy under small sample experimental data.

[0107] Furthermore, the intelligent prediction method provided by this invention eliminates the need to install strain gauges, wires, and other measuring devices on turbine blades in industrial settings. It only requires the use and acquisition of easily obtainable material and load parameters, thus solving the problems of high-temperature detachment, signal interference, and damage to aerodynamic shape caused by traditional strain gauge measurements in industrial applications.

[0108] In summary, the present invention provides a novel technical means for real-time monitoring and research of dynamic stress in turbine blades based on hybrid data training, which has significant engineering implications and broad application prospects.

[0109] The execution subject of the methods provided in this specification can be a server, which can be a server set up on a business platform, or a device such as a desktop computer or laptop computer that can execute the solutions in this specification.

[0110] When applying the turbine blade dynamic stress field prediction method based on hybrid data training provided in this manual, it is not necessary to rely on... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.

[0111] The above describes one or more embodiments of the method for predicting the dynamic stress field of turbine blades based on mixed data training. Based on the same idea, this specification also provides a corresponding device for predicting the dynamic stress field of turbine blades based on mixed data training, which includes: The first acquisition module is used to acquire the calculation dataset calculated by the blade dynamic stress numerical model and the measurement dataset measured by the blade dynamic stress test. Both the calculation dataset and the measurement dataset include the turbine blade research parameters, as well as the corresponding dynamic strain and dynamic stress data. The blade research parameters include material parameters and load parameters. The second acquisition module is used to acquire the initial prediction model; the initial prediction model includes a first sub-model and a second sub-model; the first sub-model is used to predict dynamic strain through blade research parameters, and the second sub-model is used to predict dynamic stress field through material parameters and dynamic strain. The pre-training module is used to jointly train the first sub-model and the second sub-model using a computational dataset based on the joint loss function to obtain a pre-trained model; the joint loss function is a weighted sum of the training losses of the first sub-model and the second sub-model. The fine-tuning module is used to freeze the weights in the pre-trained model whose absolute values ​​are greater than the magnitude retention threshold, and to fine-tune the pre-trained model using the measurement dataset to obtain the blade dynamic stress field prediction model. The prediction module is used to predict the dynamic stress field of the blade in real time using the blade dynamic stress field prediction model.

[0112] Specific limitations regarding the turbine blade dynamic stress field prediction device based on hybrid data training can be found in the limitations of the turbine blade dynamic stress field prediction method based on hybrid data training mentioned above, and will not be repeated here. Each module in the aforementioned turbine blade dynamic stress field prediction device based on hybrid data training can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0113] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A method for predicting the dynamic stress field of turbine blades based on mixed data training is provided.

[0114] This instruction manual also provides Figure 9 The schematic diagram of the computer device shown is as follows: Figure 9 At the hardware level, the computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 A method for predicting the dynamic stress field of turbine blades based on mixed data training is provided.

[0115] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0116] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for predicting the dynamic stress field of turbine blades based on mixed data training, characterized in that, include: Obtain the computational dataset calculated using the blade dynamic stress numerical model and the measurement dataset obtained from the blade dynamic stress test. Both the computational and measurement datasets include the research parameters of the turbine blades, as well as the corresponding dynamic strain and dynamic stress data; Blade research parameters include material parameters and load parameters; Obtain the initial prediction model; the initial prediction model includes a first sub-model and a second sub-model; the first sub-model is used to predict dynamic strain using blade study parameters, and the second sub-model is used to predict the dynamic stress field using material parameters and the dynamic strain predicted by the first sub-model; Based on the joint loss function, the first sub-model and the second sub-model are jointly trained using a computational dataset to obtain a pre-trained model; the joint loss function is a weighted sum of the training losses of the first sub-model and the second sub-model. The weights in the pre-trained model whose absolute values ​​are greater than the amplitude retention threshold are frozen, and the pre-trained model is fine-tuned using a measurement dataset to obtain a blade dynamic stress field prediction model. The dynamic stress field of the blade is predicted in real time using a blade dynamic stress field prediction model. Freeze the weights in the pre-trained model whose absolute value is greater than the magnitude retention threshold, including: Obtain the weight matrix of the pre-trained model; Weights in the weight matrix whose absolute value is greater than the magnitude retention threshold are set to the frozen state; the frozen state means that fine-tuning updates using the measurement dataset are not allowed. In the weight matrix, weights whose absolute values ​​are less than or equal to the magnitude retention threshold are set to the active state; the active state indicates that fine-tuning updates can be performed using the measurement dataset. The fine-tuning loss function used when fine-tuning the pre-trained model using the measurement dataset. for: ； in, K This refers to the number of actual test measurement points; This is the position mask matrix. The dimension of the position mask matrix is ​​the same as the dimension of the dynamic stress field output by the model. The elements in the position mask matrix are marked as 1 at the measurement point position and 0 for the rest. The dynamic stress field predicted by the model; The dynamic stress measurement values ​​are mapped to the entire field, with non-measurement point locations set to 0. It represents the Hadamardi (or Hadama) stack; This represents the L2 norm.

2. The method according to claim 1, characterized in that, The training process for jointly training the first and second sub-models includes a fast training phase, a core optimization phase, and a fine-grained convergence phase. The learning rate used when updating the model parameters decreases sequentially in the fast training phase, the core optimization phase, and the fine-grained convergence phase.

3. The method according to claim 1, characterized in that, Obtain the computational dataset calculated using the blade dynamic stress numerical model and the measurement dataset obtained from blade dynamic stress experiments, including: Based on the actual service conditions of the blade and the corresponding three-dimensional high-fidelity geometric model of the blade-disk, a numerical model of blade dynamic stress is established; the numerical model of blade dynamic stress adopts a hybrid mesh generation strategy of denser mesh in key areas and sparser mesh in non-critical areas. The blade research parameters and sampling space are determined. The Latin hypercube sampling method is used to sample the blade research parameters in the sampling space. The blade dynamic stress numerical model is used to perform numerical calculations on all sample points. The dynamic strain and dynamic stress field calculation results of the blade are extracted. All calculation results and corresponding blade research parameters are integrated to obtain the calculation dataset. Determine and arrange strain gauge measurement points, and establish a dynamic strain data acquisition system; Based on the blade's service conditions and requirements, rated operating conditions and several typical operating conditions were selected as test conditions. Tests were conducted, and the dynamic strain at the blade strain gauge measurement points under each test condition was obtained through a dynamic strain data acquisition system. The dynamic stress at the corresponding measurement points was calculated, and all measurement results and corresponding test condition parameters were integrated to obtain a measurement dataset for blade dynamic stress test measurement. Each test condition corresponds to a blade research parameter.

4. The method according to claim 3, characterized in that, Determine and arrange strain gauge measurement points, including: Based on the constraints of strain gauge arrangement and the dynamic strain distribution results of the blade calculated by the blade dynamic stress numerical model, the measurement points of the strain gauges are determined. The strain gauges are fixed to the measurement points of the blade by pasting or welding, thus completing the strain gauge measurement point arrangement for the blade dynamic strain test. Each strain gauge measurement point directly corresponds to a local position of the mesh in the blade dynamic stress numerical model, and there is a correspondence between the strain gauge measurement points and the blade mesh nodes.

5. The method according to claim 3, characterized in that, The method further includes: The measurement dataset is screened in multiple dimensions through a dual verification process. The dual verification includes signal integrity verification and repeatability verification. Signal integrity verification includes removing abnormal test samples with data interruption or amplitude abrupt change. Repeatability verification involves conducting multiple repeated tests on the test conditions, selecting several test samples with stable test conditions during the test process, and then calculating the average value as the test result to remove test samples with measurement data deviation caused by fluctuations in the test conditions.

6. The method according to claim 1, characterized in that, The method further includes, prior to training the model using the computational dataset and the measurement dataset: The computational and measurement datasets are normalized, and the normalized datasets are used for model training.

7. The method according to claim 1, characterized in that, The dynamic stress field of the blade is predicted in real time using a blade dynamic stress field prediction model, including: Acquire real-time monitoring data of the blades during the actual operation of the unit, i.e., the monitoring values ​​corresponding to the blade research parameters; The monitored values ​​were normalized using normalization parameters from both the computational and measurement datasets. The normalized monitoring values ​​are input into the blade dynamic stress field prediction model to obtain the predicted blade dynamic stress field. The predicted dynamic stress field is inversely normalized and mapped onto the three-dimensional finite element mesh nodes of the blade to obtain the dynamic stress field distribution cloud map of the blade at the current moment.

8. The method according to claim 7, characterized in that, The first sub-model includes an input layer and a feature extraction layer, and the second sub-model includes a feature fusion layer, a feature reconstruction layer, and an output layer. The normalized monitoring values ​​are input into the blade dynamic stress field prediction model to obtain the predicted blade dynamic stress field, including: The normalized monitoring values ​​are input into the first sub-model of the blade dynamic stress field prediction model. After passing through the input layer and the feature extraction layer, the dynamic strain at the measurement point of the strain gauge on the blade surface is output. The normalized monitoring values ​​and the dynamic strain output from the first sub-model of the blade dynamic stress field prediction model are input into the second sub-model of the blade dynamic stress field prediction model. After passing through the feature fusion layer, feature reconstruction layer and output layer, the predicted blade dynamic stress field is obtained.

Images