Water turbine blade stress rapid simulation method based on ResNet-SE and incremental learning

Abstract

The invention discloses a water turbine blade stress rapid simulation method based on ResNet-SE and incremental learning, and belongs to the technical field of water turbine blade stress field research. The water turbine blade stress rapid simulation method comprises the steps that historical working condition parameters and corresponding stress field data are acquired, and a normalized historical sample library is formed; the method comprises the following steps: pre-training a ResNet-SE model constructed based on a ResNet model and an SE module to obtain a ResNet-SE pre-training model; after new working condition data are collected, historical samples are screened through model errors to serve as playback samples to be mixed with new samples; constructing a knowledge distillation framework based on a ResNet-SE pre-training model, and performing incremental training on the student model to obtain a water turbine blade stress field prediction model by taking simultaneous optimization of current task prediction precision and historical knowledge retention as targets; using the model to predict a water turbine blade stress field according to input real-time working condition parameters; according to the method, prediction errors can be reduced, the number of training iterations is reduced, high-precision stress field real-time simulation can still be maintained for a large number of incremental tasks, and powerful guarantee is provided for operation and design of power equipment.

Description

Technical Field

[0001] This invention relates to the field of stress field research for turbine blades, and in particular to a fast simulation method for turbine blade stress based on ResNet-SE and incremental learning. Background Technology

[0002] In traditional methods, finite element analysis (FEM) is a common approach for studying the stress field of turbine blades. It discretizes the three-dimensional physical model of the turbine blade and solves complex elasticity and stress balance equations to obtain the stress field distribution across the entire blade domain. Simulation methods based on machine learning have emerged in recent years. Some studies employ algorithms such as support vector machines (SVMs) and neural networks to train and model stress field data related to turbine blades. While these methods improve stress prediction efficiency to some extent, their reliance on fixed training datasets means that the completeness and regularity of the dataset are directly related to model performance. Traditional finite element analysis methods are computationally complex and time-consuming, with single-condition simulations often requiring hours to days. This makes it difficult to meet the engineering requirements of rapid iterative design of turbine blades and real-time stress monitoring during unit operation. Furthermore, during turbine operation, guide vane opening and head are key operating parameters, and their changes directly affect the flow field distribution around the runner, thereby altering the stress and strain state of the blades. Traditional turbine blade stress analysis methods are mostly based on numerical calculations or experimental tests under specific operating conditions, making it difficult to comprehensively cover the complex and ever-changing actual operating conditions. Moreover, they lack efficient and accurate methods for stress prediction under new operating conditions, resulting in poor flexibility and scalability.

[0003] Simulation techniques based on machine learning rely on fixed training datasets. When the operating conditions of the turbine change significantly or the blade design parameters are adjusted to produce new structural features, the model's generalization ability is insufficient and it cannot accurately predict complex stress distributions. It is necessary to collect a large amount of sample data again for full training, which not only results in high data collection and labeling costs but also long training cycles and poor timeliness.

[0004] Some existing technologies attempt to address the aforementioned problems by introducing incremental learning. Incremental learning, as a machine learning method that can learn from new data using existing knowledge models to continuously optimize model performance, provides a new approach to solving the problem of stress prediction for turbine blades under different operating conditions. However, these technologies often suffer from instability in the incremental training process and weak ability to remember historical data. This causes the model to easily forget historically valid features when adapting to new operating conditions and structures, thus affecting the overall prediction accuracy of the stress field of turbine blades under complex and variable operating conditions, making it difficult to meet the simulation reliability requirements of actual engineering. Summary of the Invention

[0005] This invention provides a fast simulation method for turbine blade stress based on ResNet-SE and incremental learning to overcome the above-mentioned technical problems.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A fast simulation method for turbine blade stress based on ResNet-SE and incremental learning includes: S1: Obtain historical operating condition parameters and corresponding turbine blade stress field distribution data, normalize the operating condition parameters, and form a historical sample library; S2: Construct a ResNet-SE model based on the ResNet model and the SE module, and pre-train the ResNet-SE model using a historical sample library to obtain the ResNet-SE pre-trained model; S3: Collect a certain amount of new working condition data as a new sample set, construct a sample selection method based on model error to screen out the replay sample set from the historical sample library, and form a mixed sample set together with the new sample set; S4: Construct a knowledge distillation framework based on the ResNet-SE pre-trained model. Use the mixed sample set to incrementally train the student model in the knowledge distillation framework with the goal of simultaneously minimizing the prediction error of the current task and the knowledge difference with the historical model, and obtain the incrementally trained turbine blade stress field prediction model. S5: Use the turbine blade stress field prediction model to predict the turbine blade stress field based on the input real-time operating parameters.

[0007] Furthermore, a knowledge distillation framework is constructed based on the ResNet-SE pre-trained model. Using the mixed sample set, with the goal of simultaneously minimizing the prediction error of the current task and the knowledge difference with historical models, the student model in the knowledge distillation framework is incrementally trained to obtain the incrementally trained turbine blade stress field prediction model, including: S41. Use the ResNet-SE pre-trained model as the teacher model and initialize the corresponding student model; S42. Construct a total loss function that includes distillation loss and prediction loss. Combine the mixed sample set and the total loss function to perform incremental training and updating of the student model, and obtain the incrementally trained student model as the turbine blade stress field prediction model.

[0008] Furthermore, a total loss function is constructed that includes distillation loss and prediction loss, including: Construct the total loss function as shown in formula (1). (1) in, For the total loss function, An adaptive weighting coefficient is used to control distillation losses. Predicted values ​​for the student model Compared with real samples The mean square error between them Predicted values ​​for the teacher model Compared with the student model's predicted values The mean square error between them.

[0009] Furthermore, the ResNet-SE model includes an input layer, a first ResNet network, a first SE module, a second ResNet network, a second SE module, and a fully connected layer connected in sequence. The first ResNet network is used to perform the first round of feature extraction and feature transformation on the initial feature tensor input to the input layer, and output the first feature tensor; The first SE module is used to perform a first round of feature compression, activation and recalibration on the input first feature tensor, and output a first modulation feature tensor; The second ResNet network is used to perform a second round of feature extraction and feature transformation on the input first modulation feature tensor, and output a second feature tensor; The second SE module is used to perform a second round of feature compression, activation, and recalibration on the input second feature tensor, and output a second modulation feature tensor; The fully connected layer is used to perform a linear transformation on the second modulation feature tensor to output the predicted stress field distribution of the turbine blades.

[0010] Furthermore, a sample selection method based on model error is constructed to screen replay sample sets from the historical sample library, which, together with the new sample set, form a mixed sample set, including: All samples in the historical sample database are sequentially input into the ResNet-SE pre-trained model to obtain the prediction result for each sample. The error value between the prediction result and the actual data is calculated, and the samples are sorted in descending order according to the error size to construct an error sequence. The sorted samples are divided into three error intervals based on three preset thresholds; Samples are taken at a set sampling ratio in three error intervals to obtain a playback sample set; The replay sample set is mixed with the new sample set to form a mixed sample set.

[0011] Furthermore, the initialized student model has the same structure as the ResNet-SE pre-trained model.

[0012] Beneficial Effects: This invention provides a fast simulation method for turbine blade stress based on ResNet-SE and incremental learning. In the pre-training stage, data normalization is used to eliminate the influence of dimensions and accelerate model convergence. A ResNet-SE model is constructed by combining ResNet and SE modules. The deep residual structure in ResNet solves the gradient vanishing problem, and the SE module enhances the extraction of key features, improving the accuracy of stress field mapping. In the incremental learning stage, a sample selection method based on model error is constructed to accurately select high-value samples to build a replay sample library, improving the model's generalization ability to uncertain samples. An incremental learning strategy based on knowledge distillation achieves efficient transfer of new and old knowledge, effectively avoiding catastrophic forgetting of the model. Compared with traditional methods, this method achieves breakthroughs in prediction accuracy under new operating conditions, model adaptation speed, and retention of historical knowledge, significantly reducing prediction errors and training iterations. Even when facing a large number of incremental tasks, it can still maintain high-precision real-time stress field simulation, providing strong support for the operation and design of power equipment. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 The flowchart of the fast simulation method for turbine blade stress based on ResNet-SE and incremental learning provided by this invention; Figure 2 The data processing block diagram for the simulation of the stress field of a water turbine blade based on ResNet and incremental learning provided by this invention; Figure 3 The structure diagram of the ResNet-SE model provided by this invention; Figure 4 This is a diagram showing the stress field prediction results of the initial model of this embodiment for a newly added working condition; Figure 5 The image shows the stress field prediction results of the updated model for the same newly added working condition in this embodiment of the invention. Figure 6 This is a diagram showing the stress field prediction results of the updated model for a certain initial working condition in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] This embodiment provides a fast simulation method for turbine blade stress based on ResNet-SE and incremental learning, such as... Figure 1 As shown, it includes: S1: Obtain historical operating condition parameters and corresponding turbine blade stress field distribution data, normalize the operating condition parameters, and form a historical sample library; S2: Construct a ResNet-SE model based on the ResNet model and the SE module, and pre-train the ResNet-SE model using a historical sample library to obtain the ResNet-SE pre-trained model; S3: Collect a certain amount of new working condition data as a new sample set, construct a sample selection method based on model error to screen out the replay sample set from the historical sample library, and form a mixed sample set together with the new sample set; S4: Construct a knowledge distillation framework based on the ResNet-SE pre-trained model. Use the mixed sample set to incrementally train the student model in the knowledge distillation framework with the goal of simultaneously minimizing the prediction error of the current task and the knowledge difference with the historical model, and obtain the incrementally trained turbine blade stress field prediction model. S5: Use the turbine blade stress field prediction model to predict the turbine blade stress field based on the input real-time operating parameters.

[0017] Specifically, such as Figure 2 As shown, it includes the following two stages: The model pre-training stage based on historical sample databases is as follows: First, the samples are standardized. Second, a pre-trained model is constructed and trained using the sample data, employing stacked residual blocks to capture multi-level features of operating parameters (inlet pressure). Then, a Squeeze-and-Excitation module is introduced after the residual blocks, generating channel attention weights through global average pooling, fully connected layers, and a sigmoid activation function to automatically enhance important feature channels and suppress irrelevant information. Finally, a fully connected layer is used to directly map the stress field distribution of the turbine blades, yielding the prediction results. In the incremental learning phase, firstly, a sample selection method based on model error is adopted. By quantifying the error between the model's prediction results and the actual data, samples that play a key role in improving model performance are accurately located and replayed to obtain a replayed sample set, which is then combined with the new sample set to form a mixed sample. Subsequently, a teacher-student model and distillation loss are introduced during incremental training. The pre-trained model is set as the teacher model, and the loss weight coefficient is dynamically adjusted based on the prediction error to train the student model. The model is then updated using gradient descent. Through knowledge distillation and data replay mechanisms, the student model can adapt to the distribution characteristics of new working condition data while reviewing knowledge of old working condition data, achieving a better balance in learning new and old tasks, and obtaining the final prediction model. Water turbines and auxiliary equipment are important hydroelectric devices. Water flow drives the turbine blades, converting the energy of the water flow into rotational mechanical energy to drive the generator to produce electricity. Turbine blades are crucial components of the turbine. To ensure the normal and stable operation of the turbine, the blade condition needs to be checked regularly. The method proposed in this invention provides a core driver for intelligent condition monitoring and predictive maintenance of turbine blades, upgrading traditional blade monitoring, which relies on shutdown inspections, into an online, real-time, and predictive intelligent health management system.

[0018] The fast simulation method based on ResNet-SE and incremental learning proposed in this invention, with its excellent adaptability and scalability, is also applicable to the physical field simulation scenarios of other equipment, providing a general and efficient solution for the safe operation and optimized design of equipment.

[0019] In a specific embodiment, the scheme for obtaining historical operating condition parameters and corresponding turbine blade stress field distribution data, and normalizing the operating condition parameters to form a historical sample library is as follows: Based on the unidirectional fluid-structure interaction method, a high-fidelity fluid-structure interaction simulation model of a water turbine is created, and the stress distribution of the water turbine blades under different operating parameters is calculated and obtained, thereby constructing a priori dataset of the water turbine. In the stress field simulation of turbine blades, there are significant differences in the dimensions and numerical distribution of operating parameters. Imbalanced feature scales can lead to an imbalance in the model's sensitivity to different features during training, resulting in training bias and reduced prediction accuracy. Therefore, a normalization preprocessing method is employed to eliminate differences in data dimensions, avoid model training bias caused by differences in feature scales, and obtain a historical sample database.

[0020] In a specific embodiment, the scheme for constructing a ResNet-SE model based on the ResNet model and the SE module, and pre-training the ResNet-SE model using a historical sample database to obtain the ResNet-SE pre-trained model is as follows: The ResNet-SE model is as follows: Figure 3 As shown, it includes an input layer, a first ResNet network, a first SE module, a second ResNet network, a second SE module, and a fully connected layer connected in sequence. The first ResNet network is used to perform the first round of feature extraction and feature transformation on the initial feature tensor input to the input layer, and output the first feature tensor; The first SE module is used to perform a first round of feature compression, activation and recalibration on the input first feature tensor, and output a first modulation feature tensor; The second ResNet network is used to perform a second round of feature extraction and feature transformation on the input first modulation feature tensor, and output a second feature tensor; The second SE module is used to perform a second round of feature compression, activation, and recalibration on the input second feature tensor, and output a second modulation feature tensor; The fully connected layer is used to perform a linear transformation on the second modulation feature tensor to output the predicted stress field distribution of the turbine blades. The ResNet-SE model is pre-trained using a historical sample database to obtain the ResNet-SE pre-trained model; In this solution, both the ResNet network and the SE module are mature models in machine learning and neural networks. Those skilled in the art know how to train them to obtain pre-trained models, so the training process will not be described in detail.

[0021] The basic idea of ​​residual blocks is to add the input features to the output features of the convolutional layer, adding a skip connection to the traditional feedforward convolutional network. This allows the network to learn a residual mapping from input features to output features, rather than directly learning the mapping. This avoids feature degradation that occurs in deep networks. The network consists of two convolutional layers and a skip connection. To overcome the vanishing gradient problem in deep neural networks, the remaining part of the network allows for effective training, improving the overall training efficiency of the network while ensuring that the optimal solution is not compromised. The core idea of ​​the SE module lies in achieving dynamic recalibration of feature channels through compression (Squeeze) and excitation (Excitation). The compression operation uses global average pooling to compress the spatial dimension of each feature channel into a single statistic, generating a channel description vector. Specifically, given an input feature map... in It is the number of channels. and These are the height and width of the feature map, respectively. Global average pooling divides the feature map of each channel... Compressed into scalar Through this operation, the SE module can capture the global context information of each channel and generate a length of [length missing]. The channel description vector; The activation operation learns the importance weights of each channel through a fully connected neural network. First, the channel description vector... After passing through a bottleneck layer (a fully connected layer with dimensionality reduction capabilities), the number of channels is reduced from... Reduce to / ,in This is the reduction ratio (usually set to 16) to reduce computational complexity. Next, after a ReLU activation function, a second fully connected layer restores the number of channels to [previous value]. Finally, the weight values ​​are normalized to the range [0,1] using the Sigmoid activation function; The weight vector obtained from the excitation Applied to the original input feature map This allows for the recalibration of each channel. Specifically, for each channel... Its recalibrated feature map Through this operation, the SE module can enhance the response to important channels and suppress the influence of unimportant channels, thereby improving the feature representation ability and prediction performance of the entire network.

[0022] Therefore, this scheme leverages the advantages of deep residual structures in ResNet networks, enhancing the network's feature extraction capability for combinations of input parameters by stacking residual blocks; then, it combines the advantage of the SE attention mechanism in adaptively adjusting the weights of different channels in the network structure to improve the network's attention to stress-sensitive region features and enhance the accuracy of stress field mapping.

[0023] In a specific embodiment, the following scheme involves collecting a certain amount of new operating condition data as a new sample set, constructing a sample selection method based on model error to screen out a replay sample set from the historical sample library, and combining it with the new sample set to form a hybrid sample set: Reacquire new working condition data and calculate the corresponding stress data. Normalize the new working condition data to obtain a new sample set. When the operating conditions of a water turbine change, the stress field distribution varies significantly under different external conditions. When incremental training is performed using data from the new operating conditions, catastrophic forgetting can easily occur. Therefore, this invention constructs a sample selection method based on model error, replaying some historical data to alleviate the forgetting problem in the incremental learning process at the data level and maintain the model's predictive performance under old operating conditions. This sampling method uses sample prediction error as a quantitative indicator to accurately measure the model's adaptability to the current operating conditions. The larger the error value, the higher the uncertainty of the model when processing that sample, and it also reflects the model's cognitive blind spots under the corresponding operating conditions. Such samples contain key information that the model has not yet fully learned. By prioritizing the selection of high-error samples for training, it is possible to effectively fill the model's cognitive gaps, improve the model's generalization ability, and enhance its adaptability and predictive accuracy to unknown scenarios. The proposed sampling algorithm selects mean squared error as the evaluation standard for model error. The specific sampling steps are as follows: Error calculation and sorting: All samples in the historical sample library are sequentially input into the pre-trained model to obtain the prediction result of each sample. The error value between the prediction result and the actual data is calculated, and the samples are sorted in descending order according to the error magnitude to construct an error sequence. The sorted samples are divided into three error intervals based on three preset thresholds; Specifically, this scheme pre-sets three thresholds based on the distribution of the error sequence to divide the error range into "high error interval", "medium error interval" and "low error interval"; the specific thresholds can be determined by those skilled in the art based on the actual situation, and no specific numerical limit is imposed. Samples are taken at a set sampling ratio in three error intervals to obtain a playback sample set; Specifically, different sampling strategies are adopted for different error ranges. In the "high error range," samples are selected with a higher sampling ratio; in the "medium error range," samples have strong representativeness of the working conditions and are sampled at a moderate ratio; in the "low error range," samples are sampled at a lower ratio to avoid introducing too many simple samples that would affect training efficiency. Finally, the samples extracted from each range together form the playback sample set.

[0024] The sampling ratio within each interval can be determined by those skilled in the art based on the required number of samples, the number of training rounds, or actual requirements. This application does not impose any restrictions, as long as the sampling ratios of the several intervals are significantly different. The replay sample set is mixed with the new sample set to form a mixed sample set.

[0025] In this scheme, the hierarchical sampling strategy based on model error can accurately select high-value samples to build a replay sample library, thereby improving the model's generalization ability to uncertain samples.

[0026] In a specific embodiment, a knowledge distillation framework is constructed based on the ResNet-SE pre-trained model. Using the mixed sample set, with the goal of simultaneously minimizing the prediction error of the current task and the knowledge difference with historical models, the student model in the knowledge distillation framework is incrementally trained to obtain the incrementally trained turbine blade stress field prediction model. S41. Use the ResNet-SE pre-trained model as the teacher model and initialize the corresponding student model; the initialized student model has the same structure as the ResNet-SE pre-trained model. Specifically, in the knowledge distillation process, the teacher model is a pre-trained old model, and the student model is the current model to be updated. The knowledge transfer between the two is achieved through a designed loss function. This invention innovatively designs the loss function into two parts: the first part is the student model prediction loss, used to measure the model's learning effect; the second part is the distillation loss of the teacher-student model, used for the transfer learning of old knowledge. S42. Construct a total loss function that includes distillation loss and prediction loss, as shown in formula (2). (2) in, For the total loss function, An adaptive weighting coefficient is used to control distillation losses. Predicted values ​​for the student model Compared with real samples The mean square error between them Predicted values ​​for the teacher model Compared with the student model's predicted values The mean square error between them; To balance the learning priorities of new and old tasks, this invention employs an adaptive weight adjustment strategy: dynamically adjusting the weights of each component of the loss function based on the prediction error. When the learning error is high, the weight of the current task loss is automatically increased, prompting the model to prioritize learning new knowledge; when the distillation loss is high, the weight of the distillation loss is increased to strengthen the model's memory of historical knowledge, thereby effectively avoiding catastrophic forgetting and ensuring that the model maintains excellent performance under both new and old conditions.

[0027] The student model is incrementally trained and updated by combining the mixed sample set and the total loss function, and the incrementally trained student model is used as the stress field prediction model for turbine blades. Specifically, let all trainable parameters in the ResNet-SE model be... The above loss function can be expressed as: Based on the above loss function and the model parameters of the Adam optimizer, training stops when the loss converges or the preset maximum number of iterations is reached, and the trained student model is obtained.

[0028] Updating model parameters using gradient descent and the Adam optimizer based on the loss function is a common technique used in neural network training by those skilled in the art. Therefore, the specific training and update process will not be described in detail, as those skilled in the art know how to perform the update and training.

[0029] Example: Mixed-flow turbines are widely used in various hydropower stations due to their significant advantages, such as a wide range of adaptable heads, compact structure, and high runner efficiency. However, in actual operation, the turbine runner blades are subjected to fluid-structure interaction for extended periods, which can easily lead to stress concentration, excessive deformation, and other problems. This can result in blade cracks, fractures, and other malfunctions, seriously threatening the safe and stable operation of the turbine unit and causing huge economic losses and safety hazards.

[0030] This embodiment selects a certain type of mixed-flow turbine as the research object. Its core structural parameters are as follows: runner diameter 2.43m, runner blades 12, movable guide vanes and fixed guide vanes both 18; operating parameters range: maximum head 184.0m, rated head 149.5m, minimum head 145.7m, rated flow rate 52.3m³ / s, rated speed 272.7r / min. The runner blades are made of stainless steel with the following material parameters: density 8g / cm³, elastic modulus 210GPa, yield strength 600MPa, and Poisson's ratio 0.3.

[0031] Numerical simulations were performed to obtain blade stress data under the operating conditions of 70% guide vane opening and a characteristic head range of [145.7m-184.0m] (initial operating condition), resulting in a total of 150 samples. Of these, 105 samples (head range of [145.7m-172.51m], representing a new operating condition) were selected as the dataset for pre-training the model; and 45 samples (head range of [172.51m-184.0m) were selected for updating and validating the model during incremental learning.

[0032] The results of training and updating using this invention are as follows: Figures 4-6 As shown: Figure 4 The image shows the predicted stress field of the blade under a new operating condition by the pre-trained model (before incremental update). As can be seen from the image, the maximum stress on the blade is significantly beyond the normal range under the operating conditions of the turbine blade. The vertical axis in the image represents the von Mises stress. Figure 5The image shows the prediction results of the incremental model for the blade stress field under the new working conditions. As can be seen from the image, the error of the stress field prediction results after the model update is significantly lower than that before the model update, indicating that the performance of the model is significantly improved after being updated and trained with the new working condition data. The updated model maintains good performance under the initial conditions, such as Figure 6 As shown, the predicted stress field of the blades is within the normal range. The above demonstrates the effectiveness of the incremental learning method proposed in this invention.

[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fast simulation method for turbine blade stress based on ResNet-SE and incremental learning, characterized in that, include: S1: Obtain historical operating condition parameters and corresponding turbine blade stress field distribution data, normalize the operating condition parameters, and form a historical sample library; S2: Construct a ResNet-SE model based on the ResNet model and the SE module, and pre-train the ResNet-SE model using a historical sample library to obtain the ResNet-SE pre-trained model; S3: Collect several new working condition data as a new sample set, construct a sample selection method based on model error to screen out the replay sample set from the historical sample library, and form a mixed sample set together with the new sample set; S4: Construct a knowledge distillation framework based on the ResNet-SE pre-trained model. Use the mixed sample set to incrementally train the student model in the knowledge distillation framework with the goal of simultaneously minimizing the prediction error of the current task and the knowledge difference with the historical model, and obtain the incrementally trained turbine blade stress field prediction model. S5: Use the turbine blade stress field prediction model to predict the turbine blade stress field based on the input real-time operating parameters.

2. The method for rapid simulation of turbine blade stress based on ResNet-SE and incremental learning according to claim 1, characterized in that, A knowledge distillation framework is constructed based on the ResNet-SE pre-trained model. Using the hybrid sample set, with the goal of simultaneously minimizing the prediction error of the current task and the knowledge difference with historical models, the student model in the knowledge distillation framework is incrementally trained to obtain the incrementally trained turbine blade stress field prediction model, including: S41. Use the ResNet-SE pre-trained model as the teacher model and initialize the corresponding student model; S42. Construct a total loss function that includes distillation loss and prediction loss. Combine the mixed sample set and the total loss function to perform incremental training and updating of the student model, and obtain the incrementally trained student model as the turbine blade stress field prediction model.

3. The method for rapid simulation of turbine blade stress based on ResNet-SE and incremental learning according to claim 2, characterized in that, Construct a total loss function that includes distillation loss and prediction loss, including: Construct the total loss function as shown in formula (1). （1） in, For the total loss function, An adaptive weighting coefficient is used to control distillation losses. For the predicted values ​​of the student model Compared with real samples The mean square error between them Predicted values ​​for the teacher model Compared with the student model's predicted values The mean square error between them.

4. The method for rapid simulation of turbine blade stress based on ResNet-SE and incremental learning according to claim 1, characterized in that, The ResNet-SE model includes an input layer, a first ResNet network, a first SE module, a second ResNet network, a second SE module, and a fully connected layer connected in sequence. The first ResNet network is used to perform the first round of feature extraction and feature transformation on the initial feature tensor input to the input layer, and output the first feature tensor; The first SE module is used to perform a first round of feature compression, activation and recalibration on the input first feature tensor, and output a first modulation feature tensor; The second ResNet network is used to perform a second round of feature extraction and feature transformation on the input first modulation feature tensor, and output a second feature tensor; The second SE module is used to perform a second round of feature compression, activation, and recalibration on the input second feature tensor, and output a second modulation feature tensor; The fully connected layer is used to perform a linear transformation on the second modulation feature tensor to output the predicted stress field distribution of the turbine blades.

5. The method for rapid simulation of turbine blade stress based on ResNet-SE and incremental learning according to claim 1, characterized in that, A sample selection method based on model error is constructed to filter replay sample sets from the historical sample library, which, together with the new sample set, form a mixed sample set, including: All samples in the historical sample database are sequentially input into the ResNet-SE pre-trained model to obtain the prediction result for each sample. The error value between the prediction result and the actual data is calculated, and the samples are sorted in descending order according to the error size to construct an error sequence. The sorted samples are divided into three error intervals based on three preset thresholds; Samples are taken at a set sampling ratio in three error intervals to obtain a playback sample set; The replay sample set is mixed with the new sample set to form a mixed sample set.

6. The method for rapid simulation of turbine blade stress based on ResNet-SE and incremental learning according to claim 2, characterized in that, The initialized student model has the same structure as the ResNet-SE pre-trained model.

Images