A small sample turbine blade damage parameter prediction method based on meta learning

Abstract

This invention presents a method for predicting small-sample turbine blade damage parameters based on meta-learning, belonging to the field of turbine blade fatigue life assessment and prediction. Combining the load characteristics of typical positions on the turbine blade during service, the method treats each typical position at different cross-sectional heights as different service tasks. A meta-learning model is employed to effectively predict damage parameters at various positions on the turbine blade under different service times, improving blade utilization and reducing operating costs. Addressing the issue that turbine blade service data exhibits typical temporal correlation but has excessively short time series, an LSTM network is used as the base model. Complete time series samples from each typical position are packaged into a "pseudo-sample" for model training. This method utilizes meta-learning to solve the small-sample prediction problem while leveraging the temporal correlation of the samples to improve the model's prediction accuracy. This invention is applicable to the field of turbine blade fatigue life assessment and prediction, providing technical support for the reasonable scrapping of aero-engine turbine blades.

Description

Technical Field

[0001] This invention relates to a method for predicting damage parameters of turbine blades based on meta-learning, belonging to the field of turbine blade fatigue life assessment and prediction. Background Technology

[0002] Turbine blades are continuously exposed to high temperatures and subjected to complex stresses from the combustion gases and turbine rotation during aero-engine operation. This leads to ongoing microstructural damage within the turbine blades, ultimately resulting in blade failure. Accurate assessment of turbine blade damage is crucial for determining their suitability for use, ensuring the safe operation of the aircraft while fully utilizing turbine blade performance.

[0003] The current main method for determining turbine blade decommissioning involves preparing cross-sections of designated areas on the blade and qualitatively assessing its microstructure to determine its continued serviceability. With the development of computer image processing technology in recent years, some researchers have introduced image algorithms into the semi-quantitative extraction and measurement of turbine blade microstructure parameters, providing support for more refined and quantitative turbine blade decommissioning. However, this method still only determines the current state of the turbine blade and cannot effectively predict its future state, thus failing to meet the needs of a health management system (PHM).

[0004] With the continuous development of neural networks, their ability to analyze and extract from existing data and classify and predict unknown data is constantly improving. Many researchers have applied neural networks to fault detection and prediction in various equipment fields, such as monitoring equipment operating status through vibration signals of bearings and cutting tools. Due to their unique γ / γ′ phase structure, nickel-based superalloys exhibit a significant correlation between damage evolution and microstructure. Current image processing technology can effectively quantify and extract this microstructure, making the application of neural network technology to damage prediction of turbine blade materials highly feasible. However, due to the high cost of turbine blades and the specific nature of their application scenarios, damage data for turbine blades at different service times is difficult to obtain. In the absence of sufficient data, the accuracy of predicting damage parameters using neural networks is unlikely to meet practical application requirements.

[0005] To address the issue of low prediction accuracy in small sample problems, researchers have proposed various targeted neural network models. However, most of these are applied to fields such as image recognition and language processing, with relatively few neural networks specifically designed for regression prediction on small sample data. Furthermore, due to the limited number of turbine blade samples, damage data cannot be segmented using conventional time-series data processing methods. Among the many neural network models, meta-learning, with its unique training mode, has become a more suitable model for regression prediction on small sample data. By learning universally applicable meta-knowledge across multiple tasks, the model can converge quickly on new tasks, achieving high prediction accuracy. Summary of the Invention

[0006] To address the problems of existing turbine blade rejection criteria being overly conservative due to subjective factors and judgment methods limited to the current state, the main objective of this invention is to propose a small-sample turbine blade damage parameter prediction method based on meta-learning. Using an LSTM network as the basic model, a limited number of complete time series at each location of the blade are divided into pseudo-samples. Combined with meta-learning methods, the damage status of the turbine blade is effectively predicted based on the obtained small number of data samples, improving blade utilization, reducing blade operating costs, and providing technical support for the rational rejection of aero-engine turbine blades.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] This invention discloses a small-sample turbine blade damage parameter prediction method based on meta-learning. First, turbine blades at different locations under different service times are sliced ​​to obtain typical location samples. After sample processing, microstructure images are taken at different locations. Then, image processing techniques are used to extract the microstructure morphology at different locations and calculate damage parameters. By analyzing the characteristics of the damage data, a meta-learning model is constructed to predict the damage parameters, obtaining predicted damage parameter values ​​for each location on the turbine blade under different service times. This improves blade utilization, reduces operating costs, and provides technical support for the disposal of turbine blades.

[0009] This invention discloses a method for predicting small-sample turbine blade damage parameters based on meta-learning, comprising the following steps:

[0010] Step 1: Obtain microstructure images of typical locations on turbine blades at different service times;

[0011] Step 1.1: Slice the turbine blade at various typical locations, including but not limited to the tenon, blade root, 60% blade height, 77% blade height, and blade tip.

[0012] Step 1.2: Perform surface treatment on the turbine blade slices in order to obtain microscopic images. Surface treatment includes, but is not limited to, sample preparation, grinding, polishing, and etching.

[0013] Step 1.3: Scan the typical locations of turbine blade slices at different positions using a field emission scanning electron microscope and take multiple microscopic images at different locations. The typical locations on each turbine blade slice include, but are not limited to, the leading edge, trailing edge, blade base, and blade back.

[0014] Step 2: Calculate the microstructure damage parameters at typical locations on turbine blades under different service times;

[0015] Step 2.1: Crop the captured microscopic tissue images to the same size and perform preliminary screening to remove low-quality samples caused by problems with the surface treatment of the sections or imaging issues;

[0016] Step 2.2: Perform binarization processing on the images obtained from the initial screening in Step 2.1. The binarization processing methods include, but are not limited to, OTU binarization, local thresholding, and triangular thresholding.

[0017] Step 2.3: Input the binarized grayscale image into image processing software for measurement to obtain the microstructure morphology parameters at each location, including but not limited to the average area of ​​the reinforcing phase and the ratio of the long axis to the short axis of the reinforcing phase.

[0018] Step 2.4: Compare the microscopic morphology parameters at the tenon position of each turbine blade with the morphology parameters at each position of the turbine blade, make the ratio dimensionless, and define this ratio as the damage parameter.

[0019] Step 3: Using the damage parameters at various locations under different service times obtained in Step 2, establish a training dataset for the meta-learning model;

[0020] Step 3.1: Perform statistical filtering on multiple sets of data taken from various locations at different service times to remove obvious noise and reduce the impact of erroneous data on the model training results;

[0021] Step 3.2: Classify all data according to different locations and divide them into different training tasks. Each training task contains multiple sets of data for that location at all service times. Randomly divide the different training tasks into training sets and test sets according to a certain ratio.

[0022] Step 3.3: Arrange and group the data in each training task according to the service time series. Each group of data contains a damage parameter under all service times. According to a certain proportion, a part of the groups are divided into support sets and the rest are divided into query sets.

[0023] Step 3.4: Arrange the data in each group obtained in Step 3.3 according to the service time relationship, and package the several samples with complete time series data in each group into a "pseudo-sample".

[0024] Step 4: Merge the training data obtained in Step 3 into "pseudo-samples" as the smallest unit to build the training dataset of the basic model. Divide the training set and validation set according to the ratio. Adjust the model structure of the basic model based on the training results and determine the optimal model architecture under the current dataset.

[0025] Step 5: Apply the model architecture obtained in Step 4 to construct a meta-learning model for predicting damage parameters, and train the meta-learning model using the training set obtained in Step 3.

[0026] Step 5.1: Define the gradient update rule for the meta-learning model; The meta-learning model consists of two models with the same architecture: a base model and a meta-model, identical to the model obtained in Step 4; In each training task, the model parameters Φ of the meta-model are... 0 Assign θ to the base model, and let the base model train once on the support set S of the training task to obtain θ. 1 Using θ 1 Make predictions on the query set Q and obtain the loss function loss(θ). 1 loss(θ) 1 The formula for calculating ) is:

[0027]

[0028] In the formula, n is the number of samples participating in the training under the current task. The predicted value obtained under this model, The true label values ​​are for the training tasks. The total loss of the meta-learning model is derived from the sample loss for each task.

[0029] Calculate the loss function loss(θ) 1 ) for θ 1 gradient G 1 This gradient is the update gradient of the meta-model. Multiplying this gradient by the learning rate of the meta-model and applying it to the meta-model parameters Φ yields the updated meta-model parameters Φ. 1 Complete the training for the current task;

[0030] In one round of training, the total loss of the meta-learning model when all training tasks are completed is:

[0031]

[0032] In the formula, m represents the total number of training tasks, and n represents the number of samples in the training tasks. Let i be the predicted value obtained by the model for sample i in task j. Let be the label value of sample i in task j.

[0033] The gradient of each parameter in the base model is calculated using this loss value. This gradient is then multiplied by the learning rate of the meta-model and applied to the meta-model parameters to update the meta-learning model for one epoch. The gradient descent formula for the meta-model in one epoch is:

[0034]

[0035] In the formula, Φ represents the parameters of the trained meta-model. 0 η represents the initial parameters of the meta-model, and η is the learning rate.

[0036] Step 5.2: Define the training data selection rules in the training process of the meta-learning model: Select training tasks without replacement from the divided training set, and put one training task into the meta-learning model for training each time until all training tasks are completed, that is, complete one epoch of training. Repeat this process until the set number of training cycles is reached, exit training, and return a trained meta-learning model.

[0037] Step 6: Use the test set to test the accuracy of the meta-learning model obtained in Step 5 and verify the model's reliability. If the accuracy meets the requirements, the model will be used to predict turbine blade damage parameters. If the accuracy does not meet the requirements, return to Step 4 and adjust the model until the accuracy meets the usage requirements.

[0038] Step 7: Apply the model tested in Step 6 to guide the maintenance and disposal of aero-engine turbine blades. Input the turbine blade data during actual service into the prediction model to predict the future damage state of the aero-engine turbine blades, thereby guiding the maintenance and replacement of turbine blades, improving turbine blade utilization, and reducing the operating cost of aero-engines.

[0039] Beneficial effects:

[0040] 1. The present invention discloses a small-sample turbine blade damage parameter prediction method based on meta-learning. It combines the load characteristics of turbine blades at various typical positions during service and regards each typical position at different cross-sectional heights of the turbine blade as different service tasks. The meta-learning model is used to effectively predict damage parameters, improve prediction accuracy, improve blade utilization, and reduce operating costs.

[0041] 2. The present invention discloses a small sample turbine blade damage parameter prediction method based on meta-learning. In view of the problem that turbine blade service data has typical temporal correlation but the time series is too short, multiple samples under the same time series are packaged into a "pseudo-sample" to participate in model training. The small sample data is put into the meta-learning model training and its temporal correlation is used to improve the prediction accuracy of the model. Attached Figure Description

[0042] Figure 1 This is a flowchart of a small-sample turbine blade damage parameter prediction method based on meta-learning disclosed in this invention;

[0043] Figure 2 The images show the microstructure of the turbine blades and a schematic diagram of the reinforcing phases in this embodiment.

[0044] Figure 3 This example uses the loss function data on the query set of the validation set in each epoch of the meta-learning model.

[0045] Figure 4 This is a comparison chart of the predicted and actual damage parameters of the meta-learning model for samples in the test set in this embodiment. Detailed Implementation

[0046] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The technical problems solved by the present invention and its beneficial effects are also described. It should be noted that the described embodiments are only intended to facilitate understanding of the present invention and do not constitute any limitation thereof.

[0047] This embodiment discloses a small-sample turbine blade damage parameter prediction method based on meta-learning, which predicts the damage parameters of a certain type of aero-engine turbine blade at different service times. The specific implementation steps are as follows:

[0048] Step 1: Obtain microstructure images of typical locations on turbine blades of a certain type of aero-engine at four different service times;

[0049] The specific steps include:

[0050] Step 1.1: Slice the turbine blades at typical locations under four different service times. The blades under different service times are cut at the same locations: tenon, blade root, 60% blade height, and blade tip. Each sample slice is 2 mm thick.

[0051] Step 1.2: To address the issue of the small size of the turbine blade sample and the difficulty in processing it, the cut turbine blade was first cold-mounted for sample preparation. Then, the sample surface was polished sequentially using sandpaper of different grits. After the sample surface was free of obvious scratches, diamond polishing paste was used in conjunction with a wool polishing head to polish the sample surface. The polished surface was then etched with an etching solution. After etching, the surface was dried with a hair dryer.

[0052] Step 1.3: Using a field emission scanning electron microscope (FET), scan the leading edge, trailing edge, blade base, and blade base position of the tenon and turbine blade slices at various heights. The magnification is 5000x, and the image resolution is 1024*768 pixels. Take 5 to 10 microstructure images at different locations. The microstructure images are shown below. Figure 2 As shown.

[0053] Step 2: Extract microscopic damage parameters at typical locations on the model turbine blades at different service times;

[0054] The specific steps include:

[0055] Step 2.1: Use a Python program to crop the captured microscopic tissue images to a 512*512 window size. During the cropping process, the captured images are initially screened based on conditions such as scratches and carbides. In addition, the parameter column needs to be cropped to avoid affecting the binarization effect during subsequent image binarization processing.

[0056] Step 2.2: Perform binarization processing on the images obtained from the initial screening in Step 2.1. First, use OTU binarization to process the images in batches. Since there are large differences in brightness, contrast between the enhanced phase and the matrix in different locations of the images, local thresholding is used for some images with poor binarization processing results. By adjusting the window size and binarization threshold of the local binarization processing, grayscale images with better processing results are obtained.

[0057] Step 2.3: The grayscale image obtained in Step 2.2 is entered into the image processing software Image Pro Plus for measurement to obtain the average area of ​​the reinforcing phase and the aspect ratio parameter of the reinforcing phase at each typical location of the turbine blade;

[0058] Step 2.4: Since the tenon position is exposed to a low temperature during service, the morphology of the strengthening phase hardly changes with the increase of service time. Therefore, the strengthening phase parameter at the tenon position is regarded as the initial state of the turbine blade. The strengthening phase parameters at other positions are compared with the strengthening phase data at the tenon to obtain a dimensionless parameter, which is defined as the damage parameter.

[0059] Step 3: Calculate the damage parameters at various locations under different service times, and establish a training dataset for the meta-learning model;

[0060] The specific steps include:

[0061] Step 3.1: Divide the damage parameter dataset into 12 training tasks according to 12 locations: leaf root, 60% leaf height position, leading edge of leaf tip, trailing edge, leaf base, and leaf back. Each training task contains all damage parameter data for that location in all time series. Then, divide the entire dataset into training set, validation set and test set in a ratio of 7:3:2.

[0062] Step 3.3: In each training task, the damage data is organized according to the service time. Five data points are taken from each time point, and each complete time series is a group. A total of five groups are divided into support set and query set according to a 3:2 ratio.

[0063] Step 3.4: Arrange the data in each group obtained in Step 3.3 according to the service time relationship. In each group, the data are arranged in the order of 0h, 700h, 1400h, and 1900h. Set the sample group with complete time series data as a "pseudo sample".

[0064] Step 4: Due to the limited amount of data and its time-series correlation, a smaller LSTM model is considered to reduce the risk of overfitting while utilizing the temporal relationships of the data. The training data obtained in Step 3 is merged into "pseudo-samples" as the smallest unit to construct the training dataset for the basic model. The training and validation sets are divided in an 8:4 ratio. The model structure of the basic model is adjusted based on the training results to obtain the optimal model architecture for the current dataset. Finally, a small LSTM network model with 32 neurons and a single hidden layer is constructed as the basic model.

[0065] Step 5: Apply the model architecture obtained in Step 4 to construct a meta-learning model for predicting damage parameters, and train the meta-learning model using the training set obtained in Step 3.

[0066] The specific steps for constructing a meta-learning model include:

[0067] Step 5.1: Establish a model-independent meta-learning model and define the gradient update rule for the meta-learning model as follows: Two training models are set up in the meta-learning model, namely the base model and the meta-model. During the model learning process, the training task is specified by an external network. Each time a training task is input, the current model parameters Φ of the meta-model M are updated. i Cloning it into the base model, at which point the base model B and the meta-model have the same model parameters Φ. iThen, the base model B is trained using the support set from the training task. After one training iteration, model B′ is obtained. The error Loss(B′) of model B′ on the query set is calculated. The mean squared error is used as the error function for the base model on a single training task, and the formula is defined as:

[0068]

[0069] In the formula, n is the number of samples participating in the training. The predicted value obtained under this model, This represents the true label value for the training task.

[0070] Therefore, the loss function of the meta-model is:

[0071]

[0072] In the formula, m represents the total number of training tasks, and n represents the number of samples in the training tasks. Let i be the predicted value obtained by the model for sample i in task j. Let be the label value of sample i in task j.

[0073] Then, the gradient G of each parameter in model B′ is calculated using this loss value. i Multiply this gradient by the meta-model's learning rate (set to 0.005) and apply it to the meta-model's parameters to achieve parameter updates through meta-learning. Repeat the above training process until the model reaches a good level of accuracy. The formula for the descent of model parameters is:

[0074]

[0075] In the formula, Φ represents the parameters of the trained meta-model. 0 η represents the initial parameters of the meta-model, and η is the learning rate.

[0076] Step 5.2: Determine the training rules for the training tasks: Select training tasks without replacement from the divided training set. Select one training task at a time and put it into the meta-learning model for training until all training tasks are completed, i.e., one epoch of training is completed. Repeat this process until the last epoch is reached, then exit training and return a trained meta-learning model.

[0077] Step 5.3: Apply the model architecture obtained in Step 4 to the meta-learning model to establish the meta-learning model. Train the meta-learning model using the training set obtained in Step 3 and examine the model accuracy after each epoch. The results show that when the training cycle is 10, the mean squared error of the meta-learning model on the training set reaches 0.016228. The loss function curve of the query set in the validation set is shown below. Figure 3As shown, the mean square error on the test set is .

[0078] Step 6: Use the test set to test the accuracy of the model obtained in Step 5. The results show that the mean squared error of the model on the test set is 0.018111. The prediction results for the test set are as follows: Figure 4 As shown, this method achieves a significant improvement compared to the mean squared error of 0.031697 obtained using LSTM with the same amount of data. This embodiment discloses a small-sample turbine blade damage parameter prediction method based on meta-learning, which can improve prediction accuracy, increase blade utilization, and reduce operating costs, providing technical support for the rational disposal of aero-engine turbine blades.

[0079] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting small-sample turbine blade damage parameters based on meta-learning, characterized in that: Includes the following steps, Step 1: Obtain microstructure images of typical locations on turbine blades at different service times; Step 2: Calculate the microstructure damage parameters at typical locations on turbine blades under different service times; The implementation method for step 2 is as follows: Step 2.1: Crop the captured microscopic tissue images to the same size and perform preliminary screening to remove low-quality samples caused by problems with the surface treatment of the sections or problems with the shooting. Step 2.2: Perform binarization processing on the images obtained from the initial screening in Step 2.

1. The binarization processing methods include, but are not limited to, OTU binarization, local thresholding, and triangular thresholding. Step 2.3: Input the binarized grayscale image into image processing software for measurement to obtain the microstructure morphology parameters at each location, including but not limited to the average area of ​​the reinforcing phase and the ratio of the long axis to the short axis of the reinforcing phase. Step 2.4: Compare the microscopic morphology parameters at the tenon position of each turbine blade with the morphology parameters at each position of the turbine blade, make the ratio dimensionless, and define this ratio as the damage parameter; Step 3: Using the damage parameters at various locations under different service times obtained in Step 2, establish a training dataset for the meta-learning model; The implementation method for step 3 is as follows: Step 3.1: Perform statistical filtering on multiple sets of data taken from various locations at different service times to remove obvious noise and reduce the impact of erroneous data on the model training results; Step 3.2: Classify all data according to different locations and divide them into different training tasks. Each training task contains multiple sets of data for that location at all service times. Randomly divide the different training tasks into training sets and test sets according to a certain ratio. Step 3.3: Arrange and group the data in each training task according to the service time series. Each group of data contains a damage parameter under all service times. According to a certain proportion, a part of the groups are divided into support sets and the rest are divided into query sets. Step 3.4: Arrange the data in each group obtained in Step 3.3 according to the service time relationship, and package the several samples with complete time series data in each group into a "pseudo-sample"; Step 4: Merge the training data obtained in Step 3 into "pseudo-samples" as the smallest unit to build the training dataset of the basic model. Divide the training set and validation set according to the ratio. Adjust the model structure of the basic model based on the training results and determine the optimal model architecture under the current dataset. Step 5: Apply the model architecture obtained in Step 4 to construct a meta-learning model for predicting damage parameters, and train the meta-learning model using the training set obtained in Step 3. The implementation method for step 5 is as follows: Step 5.1: Define the gradient update rule for the meta-learning model; The meta-learning model consists of two models with the same architecture: a base model and a meta-model, identical to the model obtained in Step 4; In each training task, the model parameters Φ of the meta-model are... 0 Assign it to the base model, and let the base model be trained once on the support set S of the training task to obtain ,use Make predictions on the query set Q and obtain the loss function loss( loss( The formula for calculating ) is: In the formula, n is the number of samples participating in the training under the current task. The predicted value obtained under this model, The true label value for the training task; The total loss of the meta-learning model is obtained from the sample loss under each task. Calculate the loss function (loss( )right gradient G 1 This gradient is the update gradient of the meta-model. Multiplying this gradient by the learning rate of the meta-model and applying it to the meta-model parameters Φ yields the updated meta-model parameters Φ. 1 Complete the training for the current task; In one round of training, the total loss of the meta-learning model when all training tasks are completed is: In the formula, m represents the total number of training tasks, and n represents the number of samples in the training tasks. Let i be the predicted value obtained by the model for sample i in task j. Let be the label value of sample i in task j; The gradient of each parameter in the base model is calculated using this loss value. This gradient is then multiplied by the learning rate of the meta-model and applied to the meta-model parameters to update the meta-learning model for one epoch. The gradient descent formula for the meta-model in one epoch is: In the formula These are the parameters of the trained meta-model. These are the initial parameters of the meta-model. The learning rate; Step 5.2: Define the training data selection rules in the training process of the meta-learning model: Select training tasks with replacement from the divided training set, and put one training task into the meta-learning model for training each time until all training tasks are completed, that is, complete one epoch of training. Repeat this process until the set number of training cycles is reached, exit training, and return to the trained meta-learning model. Step 5.3: Apply the model architecture obtained in Step 4 to the meta-learning model to construct the meta-learning model; The meta-learning model is trained using the training set obtained in step 3, and the model training parameters are adjusted using the validation set. Step 6: Use the test set to test the accuracy of the meta-learning model obtained in Step 5 and verify the model's reliability. If the accuracy meets the requirements, the model can be used to predict turbine blade damage parameters. If the accuracy does not meet the requirements, return to Step 4 and adjust the model until the accuracy meets the usage requirements.

2. The method for predicting small-sample turbine blade damage parameters based on meta-learning as described in claim 1, characterized in that: It also includes step 7, applying the model tested in step 6 to guide the maintenance and scrapping of aero-engine turbine blades; inputting turbine blade data during actual service into the prediction model to predict the damage state of the aero-engine turbine blades at future moments, thereby guiding the maintenance and replacement of turbine blades, improving the utilization rate of turbine blades, and reducing the operating cost of aero-engines.

3. The method for predicting small-sample turbine blade damage parameters based on meta-learning as described in claim 1, characterized in that: The implementation method for step 1 is as follows: Step 1.1: Slice the turbine blade at various typical locations, including but not limited to the tenon, blade root, 60% blade height, 77% blade height, and blade tip. Step 1.2: Perform surface treatment on the turbine blade slices to obtain microscopic images. Surface treatment includes, but is not limited to, sample preparation, grinding, polishing, and etching. Step 1.3: Scan the typical locations of turbine blade slices at different positions using a field emission scanning electron microscope and take multiple microscopic images at different locations. The typical locations on each turbine blade slice include, but are not limited to, the leading edge, trailing edge, blade base, and blade back.

Images