Metal fatigue life prediction method and system based on deep learning small sample increment iterative training of self-attention mechanism

Abstract

The invention discloses a metal fatigue life prediction method and system based on deep learning small sample increment iterative training of a self-attention mechanism, and relates to the field of metal fatigue life prediction. The problems of model overfitting, low prediction precision, insufficient generalization and the like of an existing deep learning model due to the fact that an existing fatigue data sample is small are solved. The method comprises the following steps: performing model training and data derivation through original metal fatigue test data by utilizing a shallow neural network model to obtain an initial data set of a CNN-GRU-Attention hybrid neural network; constructing the neural network of a self-supervision mechanism, and setting parameters to obtain an optimal model; and 3, predicting the stress value of the fatigue life in the region, adding the stress value to original fatigue test data according to a sequence to construct an enhanced data set, repeating the steps 2 and 3 to execute the hybrid iterative training strategy, training a new optimal model, and terminating when the predicted cycle index reaches the cycle index corresponding to a predetermined fatigue limit.

Description

Technical Field

[0001] The present invention relates to the technical field of metal fatigue life prediction, and in particular to a metal fatigue life prediction method and system based on deep learning with small sample incremental iterative training based on a self-attention mechanism. Background Art

[0002] Fatigue is a serious form of failure in engineering, often causing sudden fractures without warning, resulting in serious safety hazards for equipment and personnel. Therefore, the fatigue life and fatigue strength of materials have always been important criteria for reliability and lifespan assessment in mechanical structure design.

[0003] Traditional fatigue performance research requires a large number of experimental samples and lengthy testing cycles to accurately determine the fatigue limit. Deep learning offers significant advantages in efficiency and cost control for predicting fatigue life in metallic materials. However, due to the generally small sample size of fatigue data, existing deep learning models often suffer from overfitting, low prediction accuracy, and insufficient generalization.

[0004] For the above reasons, there is an urgent need for a high-precision and generalizable metal fatigue life prediction method based on small samples to reduce the R&D cost and test cycle of hot materials in the industry and provide an efficient solution for exploring the fatigue performance of metal materials. Summary of the Invention

[0005] The present invention aims to solve the problems of overfitting, low prediction accuracy, and insufficient generalization in existing deep learning models caused by the generally small fatigue data samples in existing fatigue performance research.

[0006] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0007] Solution 1: The present invention proposes a metal fatigue life prediction method based on deep learning with small sample incremental iterative training based on self-attention mechanism, the method comprising the following steps:

[0008] Step 1: Use the shallow neural network model to train the model and realize data derivation through the original metal fatigue test data to obtain the initial data set D0 of the CNN-GRU-Attention hybrid neural network;

[0009] Step 2: Divide the initial data set D0 described in step 1 into a training set and a test set, construct a CNN-GRU-Attention hybrid neural network based on a self-supervisory mechanism, set parameters, use the training set for training, and use the test set to evaluate the trained model to obtain the optimal model M1;

[0010] Step 3: Use the optimal model M1 described in step 2 to predict the stress value corresponding to the fatigue life in the region, and add it to the original fatigue test data in sequence to construct an enhanced dataset D t , t = 1-n, n is the enhanced data set after the nth training;

[0011] Step 4: Enhanced dataset D constructed based on step 3 t Repeat steps 2 and 3 to execute the hybrid iterative training strategy and train the new optimal model M t+1 , t=1-n1, n1 is the optimal model after n1th training, and the optimal model M is used t+1 Generate a new augmented dataset D t+1 ,It ends when the predicted number of cycles reaches the number of cycles corresponding to the ,predetermined fatigue limit.

[0012] Furthermore, a preferred embodiment is provided, in step 1, the method for obtaining the initial data set D0 of the CNN-GRU-Attention hybrid neural network by using the shallow neural network model to perform model training and realize data derivation through the original fatigue test data is:

[0013] That is, one data point is collected every 1,000 cycles of the fatigue life of the predetermined material, and 971 sets of initial data in the cycle range of 30,000 to 1,000,000 are derived to form the initial training set D0.

[0014] Furthermore, a preferred implementation is provided, in which the construction of the CNN-GRU-Attention hybrid neural network based on the self-supervision mechanism in step 2 also includes the steps of preprocessing, dividing, and model configuring the original metal fatigue test dataset.

[0015] Furthermore, a preferred implementation is provided, in which the ratio of the training set to the test set in step 2 is 8:2.

[0016] Furthermore, a preferred embodiment is provided, in which the CNN-GRU-Attention hybrid neural network based on the self-supervision mechanism constructed in step 2 includes a single-layer 1-dimensional convolutional feature extraction module, a single-layer GRU module, a self-attention module and a self-supervised trend perception loss function module.

[0017] Furthermore, a preferred embodiment is provided, in which the enhanced dataset D constructed in step 3 is used in step 4. t ,Repeat steps 2 and 3 to execute the hybrid iterative training strategy, which includes the steps of the first stage iterative training, the second stage iterative training, and the continuous dynamic expansion stage.

[0018] Furthermore, a preferred embodiment is provided, wherein the first stage of iterative training is as follows: running M1 for 1000 epochs of training, predicting expansion to 5,000,000 cycles, generating 4,971 sets of new data, and constructing a first enhanced data set D1;

[0019] The second stage of iterative training: Use D1 to train the model M2 for 1000 epochs, and the prediction is extended to 10,000,000 cycles, generating 9,971 sets of data and constructing the second enhanced dataset D2;

[0020] Continuous dynamic expansion stage: A recursive prediction mechanism is used: D2 is input into the training to obtain M3 and then trained for 1000 epochs; the prediction is expanded to 19,971 sets of data to form D3, and the iteration is continued until 29,971 sets of final data sets are obtained.

[0021] Solution 2: A metal fatigue life prediction system based on deep learning with small sample incremental iterative training using a self-attention mechanism, the system comprising:

[0022] The initial data set acquisition module is used to use the shallow neural network model to train the model and realize data derivation through the original metal fatigue test data, and obtain the initial data set D0 of the CNN-GRU-Attention hybrid neural network;

[0023] The optimal model training module is used to divide the initial data set D0 described in the initial data set acquisition module into a training set and a test set, construct a CNN-GRU-Attention hybrid neural network based on a self-supervisory mechanism, set parameters, use the training set for training, and use the test set to evaluate the trained model to obtain the optimal model M1;

[0024] The enhanced dataset construction module is used to predict the stress value corresponding to the fatigue life in the region using the optimal model M1 described in the optimal model training module, and add it to the original fatigue test data in sequence to construct the enhanced dataset D t , t = 1-n, n is the enhanced data set after the nth training;

[0025] Fatigue limit prediction module is used to build a module based on the enhanced dataset. The constructed enhanced dataset D t , repeat the optimal model training module and the enhanced dataset construction module to execute the hybrid iterative training strategy to train the new optimal model M t+1 , t=1-n1, n1 is the optimal model after n1th training, and the optimal model M is used t+1 Generate a new augmented dataset D t+1 ,It ends when the predicted number of cycles reaches the number of cycles corresponding to the ,predetermined fatigue limit.

[0026] Solution 3: A computer device includes a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the method described in Solution 1.

[0027] Solution 4: A computer-readable storage medium storing a computer program, wherein the computer program implements the steps of the method described in Solution 1 when executed by a processor.

[0028] The present invention is beneficial in that:

[0029] The present invention proposes a metal fatigue life prediction method and system based on deep learning small sample incremental iterative training of self-attention mechanism, which has the characteristics of high efficiency, high precision and low cost when performing life prediction compared with traditional methods.

[0030] The metal fatigue life prediction method and system based on deep learning and small sample incremental iterative training of self-attention mechanism described in the present invention have significantly improved prediction accuracy under small sample data compared with traditional methods, and reduce a large number of high-cycle fatigue tests compared with traditional methods.

[0031] The metal fatigue life prediction method and system described in the present invention, which uses small sample fatigue test data to achieve fatigue life assessment, is based on deep learning and small sample incremental iterative training of self-attention mechanism.

[0032] The metal fatigue life prediction method and system described in the present invention using deep learning with small sample incremental iterative training based on the self-attention mechanism have been verified to show good fatigue life prediction accuracy for the same metal under different temperatures, KT values, R values and other conditions.

[0033] The present invention is also applicable to the interdisciplinary field of deep learning and titanium alloy fatigue life prediction, and specifically relates to a method for predicting the fatigue limit of small sample metal materials based on deep learning and achieving high-precision fatigue life through a hybrid iterative strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of a metal fatigue life prediction method based on deep learning with small sample incremental iterative training based on self-attention mechanism as described in Implementation Method 1.

[0035] Figure 2 This is a block diagram of the CNN-GRU-Attention hybrid neural network model training system based on self-supervised learning and self-attention mechanism as described in embodiment one.

[0036] Figure 3This is a schematic diagram of the CNN-GRU-Attention hybrid neural network structure described in Implementation Method 1.

[0037] Figure 4 This is a schematic diagram of the results of the model training visualization evaluation system described in Implementation 11.

[0038] Figure 5 This is a diagram showing the high cycle fatigue life prediction results of the titanium alloy specimen described in the eleventh embodiment. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the implementation methods of this application clearer, the technical solutions in the implementation methods of this application will be clearly and completely described below in combination with the drawings in the implementation methods of this application. Obviously, the described implementation methods are only part of the implementation methods of this application, not all of the implementation methods.

[0040] Implementation method 1: This implementation method proposes a metal fatigue life prediction method based on deep learning with small sample incremental iterative training based on self-attention mechanism, the method comprising the following steps:

[0041] Step 1: Use the shallow neural network model to train the model and realize data derivation through the original metal fatigue test data to obtain the initial data set D0 of the CNN-GRU-Attention hybrid neural network;

[0042] Step 2: Divide the initial data set D0 described in step 1 into a training set and a test set, construct a CNN-GRU-Attention hybrid neural network based on a self-supervisory mechanism, set parameters, use the training set for training, and use the test set to evaluate the trained model to obtain the optimal model M1;

[0043] Step 3: Use the optimal model M1 described in step 2 to predict the stress value corresponding to the fatigue life in the region, and add it to the original fatigue test data in sequence to construct an enhanced dataset D t , t = 1-n, n is the enhanced data set after the nth training;

[0044] Step 4: Enhanced dataset D constructed based on step 3 t , repeat steps 2 and 3 to execute the hybrid iterative training strategy to train the new optimal model M t+1 , t=1-n1, n1 is the optimal model after n1th training, and the optimal model M is used t+1 Generate a new augmented dataset D t+1 ,It ends when the predicted number of cycles reaches the number of cycles corresponding to the ,predetermined fatigue limit.

[0045] Implementation 2: This implementation further limits the metal fatigue life prediction method based on deep learning with small sample incremental iterative training using a self-attention mechanism described in Implementation 1. In step 1, a shallow neural network model is used to train the model and derive data using the original fatigue test data to obtain the initial data set D0 of the CNN-GRU-Attention hybrid neural network. The method is as follows:

[0046] That is, one data point is collected every 1,000 cycles of the fatigue life of the predetermined material, and 971 sets of initial data in the cycle range of 30,000 to 1,000,000 are derived to form the initial training set D0.

[0047] Implementation method three. This implementation method further limits the metal fatigue life prediction method based on deep learning small sample incremental iterative training of self-attention mechanism described in implementation method one. The construction of the CNN-GRU-Attention hybrid neural network based on the self-supervision mechanism in step two also includes the steps of preprocessing, dividing, and model configuration of the original metal fatigue test data set.

[0048] Implementation method 4. This implementation method further limits the metal fatigue life prediction method based on deep learning small sample incremental iterative training of self-attention mechanism described in implementation method 1. The ratio of the training set and the test set described in step 2 is 8:2.

[0049] Implementation method five. This implementation method further limits the metal fatigue life prediction method of deep learning with small sample incremental iterative training based on self-attention mechanism described in implementation method one. The CNN-GRU-Attention hybrid neural network based on self-supervision mechanism constructed in step two includes a single-layer 1-dimensional convolution feature extraction module, a single-layer GRU module, a self-attention module and a self-supervised trend perception loss function module.

[0050] Implementation 6. This implementation is a further limitation of the metal fatigue life prediction method based on deep learning with small sample incremental iterative training based on self-attention mechanism described in Implementation 1. In step 4, based on the enhanced dataset D constructed in step 3, t ,Repeat steps 2 and 3 to execute the hybrid iterative training strategy, which includes the steps of the first stage iterative training, the second stage iterative training, and the continuous dynamic expansion stage.

[0051] Implementation 7. This implementation further defines the metal fatigue life prediction method based on deep learning with small sample incremental iterative training using a self-attention mechanism described in Implementation 6. The first stage of iterative training is as follows: M1 is run for 1000 epochs, and the prediction is extended to 5,000,000 cycles, generating 4,971 sets of new data and constructing a first enhanced dataset D1.

[0052] The second stage of iterative training: Use D1 to train the model M2 for 1000 epochs, and the prediction is extended to 10,000,000 cycles, generating 9,971 sets of data and constructing the second enhanced dataset D2;

[0053] Continuous dynamic expansion stage: A recursive prediction mechanism is used: D2 is input into the training to obtain M3 for 1000 epochs; the prediction is expanded to 19971 sets of data to form D3, and the iteration is continued until 29971 sets of final data sets are obtained.

[0054] Embodiment 8: This embodiment provides a metal fatigue life prediction system based on deep learning with small sample incremental iterative training based on a self-attention mechanism, the system comprising:

[0055] The initial data set acquisition module is used to use the shallow neural network model to train the model and realize data derivation through the original metal fatigue test data, and obtain the initial data set D0 of the CNN-GRU-Attention hybrid neural network;

[0056] The optimal model training module is used to divide the initial data set D0 described in the initial data set acquisition module into a training set and a test set, construct a CNN-GRU-Attention hybrid neural network based on a self-supervisory mechanism, set parameters, use the training set for training, and use the test set to evaluate the trained model to obtain the optimal model M1;

[0057] The enhanced dataset construction module is used to predict the stress value corresponding to the fatigue life in the region using the optimal model M1 described in the optimal model training module, and add it to the original fatigue test data in sequence to construct the enhanced dataset D t , t = 1-n, n is the enhanced data set after the nth training;

[0058] Fatigue limit prediction module is used to build a module based on the enhanced dataset. The constructed enhanced dataset D t , repeat the optimal model training module and the enhanced dataset construction module to execute the hybrid iterative training strategy to train the new optimal model M t+1 , t=1-n1, n1 is the optimal model after n1th training, and the optimal model M is used t+1 Generate a new augmented dataset D t+1,It ends when the predicted number of cycles reaches the number of cycles corresponding to the ,predetermined fatigue limit.

[0059] Implementation method nine: This implementation method proposes a computer device including a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the method described in any one of implementation methods one to seven.

[0060] Implementation 10: This implementation proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the method described in any one of Implementations 1 to 7 are implemented.

[0061] Implementation 11: This implementation provides an example, which is used to explain the above implementations 1 to 10. Specifically, the example is as follows:

[0062] See also Figure 1 and Figure 5 This embodiment specifically includes the following steps:

[0063] Step 1: Use the shallow neural network model to train the model and realize data derivation through the original fatigue test data to obtain the initial data set D0;

[0064] Step 2: Standardize the initial dataset D0 and use a sliding time window to construct a time series-based dataset D with a window length of 15 and a step size of 1. x0 ; At the same time, select the stress target value corresponding to the next moment of the sliding time window to construct D y0 ; x represents the data set of the input network, and y represents the predicted target stress data set corresponding to each set of input data;

[0065] Step 3: D in step 2 x0 and D y0 Divide into training set and test set in a ratio of 8:2;

[0066] Step 4: Build a CNN-GRU-Attention hybrid neural network based on self-attention and self-supervision mechanism, and use the training set in step 3 to train the model, and the test set to evaluate the model, and obtain the optimal model M1 through training;

[0067] Step 5: Use the optimal model described in step 4 to predict the stress values corresponding to fatigue life within a certain range, and add them to the original data set in sequence to construct an enhanced data set D1;

[0068] Step 6: Repeat steps 2, 3, 4, and 5 based on the new enhanced dataset to perform a hybrid iterative training strategy and train a new optimal model M.i , and use the optimal model M i Generate a new augmented dataset D i ,i=1-n represents the i-th training, which is terminated when the predicted number of cycles reaches the number of cycles corresponding to the specified fatigue limit.

[0069] In step 1, the operating conditions in the original fatigue test data include stress ratio, stress concentration factor and fatigue life, and 30,000-30,000,000 cycles are specified as the high-cycle fatigue life range of the titanium alloy specimen.

[0070] In step 1, a shallow neural network model is used to train the model and derive data using the original fatigue test data to obtain an initial data set that can be used by the hybrid neural network. That is, one data point is collected every 1,000 cycles of the material fatigue life, and 971 groups of initial data in the cycle range of 30,000 to 1,000,000 are derived to form the initial training set D.

[0071] In steps 2 and 3, in order to complete the preparation for model training, the sliding window cutting and standardization preprocessing of the time series data and the construction and training configuration of the CNN-GRU-Attention model should be implemented, including: Data preprocessing: Read the CSV data with timestamps, that is, the metal fatigue dataset, use StandardScaler for global standardization, generate time series samples with sliding windows (time_step = 15 historical steps, forecast_step = 1 forecast step), and achieve continuous sliding (no data overlap) through scroll_window = 1; Data partitioning: The ratio of training set to test set is 8:2, and the last 20% of the data is retained as the test set; Model configuration: Use a custom self-supervised trend-aware loss function, configure the Adam optimizer, and the evaluation indicators include MSE / MAE / R 2 ;

[0072] In step 4, the CNN-GRU-Attention hybrid neural network architecture includes: feature extraction module, GRU module, self-attention module and self-supervised trend perception loss function module.

[0073] In step 4, the feature extraction module includes a one-dimensional convolution layer, a dropout layer and a flatten layer.

[0074] In step 4, the 1D convolutional layer is set with 32 filters, each with a filter size of 3, and the activation function uses the ReLU function with fast convergence speed and low computational overhead. To prevent overfitting of the deep neural network and increase the sparsity of the model, a Dropout layer is added after the convolutional layer, and the dropout probability is set to 0.2. Finally, the multi-channel data is flattened through the flatten layer to increase the data dimension to meet the GRU input requirements.

[0075] In step 4, since the complexity of the metal fatigue life prediction sample data is relatively low, the GR module adopts a single-layer structure to save computing and memory overhead. At the same time, in order to capture more hidden layer time series information, return_sequences = True is set;

[0076] In step 4, the prediction of metal fatigue life mainly focuses on the survival rate under high cycle fatigue, and the number of high cycle fatigue failure cycles exceeds 10 4 to 10 5 In order to obtain better long sequence prediction capabilities, this model adopts a self-attention mechanism with stronger global information capture capabilities. It contains a trainable weight matrix W∈R d*d , trainable bias vector b∈R d , trainable attention parameter vector u∈R d*1 The specific calculation formula is:

[0077]

[0078] Input sequence: X∈R B*T*d , B is the batch size, T is the sequence length, d is the embedding dimension, where the original attention score is:

[0079] A=tanh(X*W+b)*u

[0080] Where A∈R B*T*1 , A is the attention score for each time step, and the raw attention scores are normalized along the column sequence dimension:

[0081] α=softmax(A,axis=1)

[0082] Apply attention weights to the input sequence X and aggregate the sequence

[0083]

[0084] Among them, С∈R B*d , C is the output context vector.

[0085] In step 4, the self-supervised trend perception loss function module introduces a trend penalty term based on the MAE basic loss function and the prior logic of fatigue life. When the model behavior does not conform to the prior logic of fatigue life, the loss is penalized. The calculation formula is:

[0086]

[0087] B is the batch size, T is the time step, is the true value of the b-th sample at time t, is the predicted value of the b-th sample at time t, and λ is the trend penalty weight.

[0088] In step 4, the CNN-GRU-Attention hybrid neural network model can implement a model training process with an intelligent training control mechanism (early stopping / learning rate adjustment / model saving), specifically including:

[0089] EarlyStopping: terminates training early when validation loss continues to fail to improve and automatically restores the best weights; ReduceLROnPlateau: decays the learning rate after validation loss stagnates; ModelCheckpoint: continuously saves the complete model with the lowest validation loss (best_model.h5);

[0090] In step six, the specific hybrid iterative training strategy includes:

[0091] The first stage of iterative training: run M1 for 1000 epochs of training, expand the prediction to 5000000 cycles, generate 4971 sets of new data, and construct the first enhanced data set D1;

[0092] Second stage iterative training: Use D1 to train model M2 (1000 epochs), expand the prediction to 10,000,000 cycles, generate 9971 sets of data, and construct the second enhanced dataset D2;

[0093] Continuous dynamic expansion phase: Using a recursive prediction mechanism: a. Input D2 into training to obtain M3 (1000 epochs), b. Prediction is expanded to 19,971 sets of data to form D3, c. Continue iterating until a final data set of 29,971 sets is obtained;

[0094] Termination condition: The method terminates when the predicted number of cycles reaches the number of cycles corresponding to the specified fatigue limit.

[0095] This embodiment provides a CNN-GRU-Attention hybrid neural network based on self-supervised learning and self-attention mechanism and an incremental iterative training method for high-cycle fatigue life prediction of alloys with small sample size, such as Figure 1As shown, the specific steps include:

[0096] Step 1: Preliminary processing of traditional fatigue test data of titanium alloy specimens under different working conditions, capturing fatigue life data points within 1,000,000 cycles to form a shallow neural network training dataset. It is stipulated that one data point is collected every 1,000 cycles. Therefore, a shallow neural network model is used to train the model using the original fatigue test data points, and data derivation is achieved within the range of 30,000-1,000,000 cycles. 971 sets of data are obtained for each working condition. The data under all working conditions are trained to obtain the initial training dataset D of the CNN-GRU-Attention hybrid neural network.

[0097] Step 2: Input the training data set into the CNN-GRU-Attention hybrid neural network and obtain the optimal model M through training;

[0098] The data set is randomly divided into training set and test set, 80% of which is training set and 20% is test set, and a convolutional neural network model is constructed, such as Figure 2 、 Figure 3 As shown, the parameters include 32 convolution kernels of size 3*1, regularization configuration Dropout (0.2), and activation function ReLU. The model is trained with the training set, and the test set is used to evaluate the trained model. Its indicators are MSE / RMSE / MAE / MAPE / R 2 ,Visual evaluation system such as Figure 4 shown.

[0099] Step 3: Start the hybrid neural network iterative training strategy in stages. First, run M for 1000 epochs and predict the stress value corresponding to the fatigue life within a certain range to obtain the enhanced data set D1.

[0100] Step 4: Repeat steps 2 and 3, observing the evaluation report at each stage, gradually increasing the prediction range, and continuously expanding the enhanced data set. When the predicted number of cycles reaches the specified fatigue limit, that is, the number of cycles corresponding to 30,000,000, the prediction is terminated.

[0101] The fatigue life predicted by the CNN-GRU-Attention hybrid neural network based on self-supervised learning and self-attention mechanism and an incremental iterative training method for high-cycle fatigue life prediction of small sample alloys is compared with the traditional fatigue test fitting SN curve. Figure 5As shown, the results show that the life prediction performed by a CNN-GRU-Attention hybrid neural network based on self-supervised learning and self-attention mechanism and an incremental iterative training method for high-cycle fatigue life prediction of small sample alloys proposed in the present invention has the characteristics of high efficiency, high precision and low cost compared with traditional methods.

[0102] Those skilled in the art will understand that the above description is only a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of the present disclosure may be combined or coupled in various ways, even if such a combination or coupling is not explicitly described in the present disclosure. It is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art may still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

[0103] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, the present invention is intended to include such changes and modifications as fall within the scope of the claims and their equivalents.

Claims

1. A metal fatigue life prediction method based on deep learning with small sample incremental iterative training based on self-attention mechanism, characterized by: The method comprises the following steps: Step 1: Use the shallow neural network model to train the model and realize data derivation through the original metal fatigue test data to obtain the initial data set D0 of the CNN-GRU-Attention hybrid neural network; Step 2: Divide the initial data set D0 described in step 1 into a training set and a test set, construct a CNN-GRU-Attention hybrid neural network based on a self-supervisory mechanism, set parameters, use the training set for training, and use the test set to evaluate the trained model to obtain the optimal model M1; Step 3: Use the optimal model M1 described in step 2 to predict the stress value corresponding to the fatigue life in the region, and add it to the original fatigue test data in sequence to construct an enhanced dataset D t , t = 1-n, n is the enhanced data set after the nth training; Step 4: Enhanced dataset D constructed based on step 3 t , repeat steps 2 and 3 to execute the hybrid iterative training strategy to train the new optimal model M t+1 , t=1-n1, n1 is the optimal model after n1th training, and the optimal model M is used t+1 Generate a new augmented dataset D t+1 ,It ends when the predicted number of cycles reaches the number of cycles corresponding to the ,predetermined fatigue limit.

2. The metal fatigue life prediction method based on deep learning and small sample incremental iterative training of self-attention mechanism according to claim 1 is characterized in that: In step 1, the shallow neural network model is used to train the model and realize data derivation through the original fatigue test data. The method to obtain the initial data set D0 of the CNN-GRU-Attention hybrid neural network is as follows: That is, one data point is collected every 1,000 cycles of the fatigue life of the predetermined material, and 971 sets of initial data in the cycle range of 30,000 to 1,000,000 are derived to form the initial training set D0.

3. The metal fatigue life prediction method based on deep learning and small sample incremental iterative training of self-attention mechanism according to claim 1 is characterized in that: In step 2, constructing a CNN-GRU-Attention hybrid neural network based on a self-supervisory mechanism also includes the steps of preprocessing, partitioning, and model configuration of the original metal fatigue test dataset.

4. The metal fatigue life prediction method based on deep learning and small sample incremental iterative training of self-attention mechanism according to claim 1 is characterized in that: The ratio of the training set to the test set described in step 2 is 8:

2.

5. The metal fatigue life prediction method based on deep learning and small sample incremental iterative training of self-attention mechanism according to claim 1 is characterized in that: The CNN-GRU-Attention hybrid neural network based on the self-supervision mechanism described in step 2 includes a single-layer 1-dimensional convolutional feature extraction module, a single-layer GRU module, a self-attention module and a self-supervised trend perception loss function module.

6. The metal fatigue life prediction method based on deep learning and small sample incremental iterative training of self-attention mechanism according to claim 1 is characterized in that: In step 4, the enhanced dataset D constructed in step 3 is t ,Repeat steps 2 and 3 to execute the hybrid iterative training strategy, which includes the steps of the first stage iterative training, the second stage iterative training, and the continuous dynamic expansion stage.

7. The metal fatigue life prediction method based on deep learning and small sample incremental iterative training of self-attention mechanism according to claim 6 is characterized in that: The first stage of iterative training: run M1 for 1000 epochs of training, predict and extend it to 5000000 cycles, generate 4971 sets of new data, and construct the first enhanced data set D1; The second stage of iterative training: Use D1 to train the model M2 for 1000 epochs, and the prediction is extended to 10,000,000 cycles, generating 9,971 sets of data and constructing the second enhanced dataset D2; Continuous dynamic expansion stage: A recursive prediction mechanism is used: D2 is input into the training to obtain M3 and then trained for 1000 epochs; the prediction is expanded to 19,971 sets of data to form D3, and the iteration is continued until 29,971 sets of final data sets are obtained.

8. A metal fatigue life prediction system based on deep learning with small sample incremental iterative training based on self-attention mechanism, characterized by: The system comprises: The initial data set acquisition module is used to use the shallow neural network model to train the model and realize data derivation through the original metal fatigue test data, and obtain the initial data set D0 of the CNN-GRU-Attention hybrid neural network; The optimal model training module is used to divide the initial data set D0 described in the initial data set acquisition module into a training set and a test set, construct a CNN-GRU-Attention hybrid neural network based on a self-supervisory mechanism, set parameters, use the training set for training, and use the test set to evaluate the trained model to obtain the optimal model M1; The enhanced dataset construction module is used to predict the stress value corresponding to the fatigue life in the region using the optimal model M1 described in the optimal model training module, and add it to the original fatigue test data in sequence to construct the enhanced dataset D t , t = 1-n, n is the enhanced data set after the nth training; Fatigue limit prediction module is used to build a module based on the enhanced dataset. The constructed enhanced dataset D t , repeat the optimal model training module and the enhanced dataset construction module to execute the hybrid iterative training strategy to train the new optimal model M t+1 , t=1-n1, n1 is the optimal model after n1th training, and the optimal model M is used t+1 Generate a new augmented dataset D t+1 ,It ends when the predicted number of cycles reaches the number of cycles corresponding to the ,predetermined fatigue limit.

9. A computer device comprising a memory and a processor, characterized in that A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images