Residual fatigue life prediction method and calculation device

Abstract

The invention relates to a residual fatigue life prediction method, which comprises the following steps of: pairing hysteretic loop data of a test piece in different weeks in a fatigue test to obtain hysteretic loop data pairs, forming a training subset by a plurality of hysteretic loop data pairs of the test piece, combining the training subsets of a plurality of test pieces into a training total set, taking two pieces of hysteretic loop data in the hysteretic loop data pair and a cycle interval as an input set, and taking the residual fatigue life of the cycle where the later hysteretic loop data is located as an output set; training the machine learning algorithm model by using the training total set to obtain a prediction model; and predicting the residual fatigue life by using the prediction model according to the hysteretic loop data of the target piece. The method can effectively reduce the test data volume required by model training and improve the prediction accuracy. The invention further provides a computing device.

Description

Technical Field

[0001] This invention belongs to the field of reliability engineering, and specifically relates to a method and calculation device for predicting residual fatigue life. Background Technology

[0002] High-temperature alloy components have been widely used in aerospace and other fields, often in the manufacture of critical components that significantly impact safety performance, such as high-temperature engine parts. Therefore, accurate assessment and prediction of the fatigue performance of complex high-temperature alloy components is crucial for equipment safety. Conventional fatigue life prediction methods include nominal stress methods, local stress-strain methods, and energy methods. These methods typically predict fatigue life based on a single factor and often exhibit significant errors when predicting the fatigue life of structures with stress concentrations. Some approaches introduce artificial intelligence algorithms to predict the material's processing, microstructure, and properties; however, current algorithms generally rely heavily on the size of the training sample, resulting in insufficient robustness and generalization ability under small sample conditions. Furthermore, conducting large-scale fatigue tests on complex and expensive parts such as high-temperature alloys to expand the training sample is costly. Therefore, providing a method for predicting the fatigue life of parts using a limited number of samples is of significant importance. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting remaining fatigue life that can reduce the number of samples required for fatigue life prediction. This invention also provides a computing device.

[0004] According to one embodiment of the present invention, a method for predicting remaining fatigue life is provided, the method comprising the following steps:

[0005] Step a): Provide a fatigue test dataset, which includes hysteresis loop data of multiple test specimens in each cycle of the fatigue test;

[0006] Step b): Establish a training subset, which includes hysteresis loop data pairs from multiple different cycle periods in the same test piece. Each hysteresis loop data pair includes two hysteresis loop data pairs with a given cycle interval. Combine the training subsets of each test piece into a training set, which includes an input set X = {x1, x2, Δ} and an output set Y, where x1 is the hysteresis loop data pair with an earlier cycle period, x2 is the hysteresis loop data pair with a later cycle period, Δ is the cycle interval between x1 and x2, and Y is the remaining fatigue life of the cycle in which x2 is located.

[0007] Step c): Provide a machine learning algorithm model, input the training set into the machine learning algorithm model for training, and obtain a prediction model;

[0008] Step d): Provide hysteresis loop data of the target component and use the prediction model to predict the fatigue life of the target component.

[0009] This method expands the total number of training samples without increasing the actual experimental workload by extracting and pairing data from different cycles of the same test piece. This effectively solves the model accuracy defect caused by insufficient training samples when using machine learning to predict fatigue life.

[0010] Furthermore, in some embodiments, in step a), the hysteresis loop data includes the hysteresis loop area, damping ratio, and secant stiffness.

[0011] Furthermore, in some embodiments, step a) further includes the step of removing data from the fatigue test dataset that have excessive strain peak and valley values ​​or insufficient cycle counts.

[0012] Furthermore, in some embodiments, in step b), 100 ≤ Δ ≤ 500. Too small a week interval in the hysteresis loop data pairs will prevent the data from reflecting trends, while too large a week interval will result in insufficient samples to reduce the number of usable hysteresis loop data pairs.

[0013] Furthermore, in some embodiments, step b) further includes a step of normalizing the data in the training set. Normalization can eliminate dimensional differences between different data.

[0014] Furthermore, in some embodiments, step c) further includes a step of screening the machine learning algorithm model: providing multiple machine learning algorithm models and using K-fold cross-validation to select the machine learning algorithm model with the lowest mean squared error and the smallest variance.

[0015] Furthermore, in some embodiments, the plurality of machine learning algorithm models include all or a combination of SVW, XGBoost, random forest, and neural networks.

[0016] Furthermore, in some embodiments, in step c), the machine learning algorithm model is trained, with the loss function configured as Smooth L1 and the optimizer configured as Adam.

[0017] Furthermore, in some embodiments, the fatigue life prediction results made by the prediction model using the fatigue test dataset are within a 2x dispersion band compared to the actual results.

[0018] According to another aspect of the present invention, a computing device is provided, the computing device including a memory and a processor, wherein the memory stores a computing program, and when the computing program is executed by the processor, it is capable of implementing the fatigue life prediction method provided in any of the foregoing embodiments. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of hysteresis loop data within one cycle in one embodiment;

[0020] Figure 2 This is a schematic diagram of hysteresis loop data extraction in one embodiment;

[0021] Figure 3 This is a schematic diagram of abnormal data in a fatigue test dataset in one embodiment;

[0022] Figure 4 This is a schematic diagram of the prediction results of the prediction model in one embodiment;

[0023] Figure 5 This is a comparison chart of the prediction results of the embodiment and the comparative embodiment in one example;

[0024] Figure 6 This is a comparison chart of the stability of the embodiment and the comparative model in one example.

[0025] The purpose of the above figures is to provide a detailed description of the invention so that those skilled in the art can understand the technical concept of the invention, and not to limit the invention. Detailed Implementation

[0026] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0027] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment herein. The phrase appearing in various places in the specification does not necessarily refer to the same embodiment, nor is it limited to mutually exclusive, independent, or alternative embodiments. Those skilled in the art will understand that the embodiments herein can be combined with other embodiments without structural conflict. In the description herein, "a plurality of" means at least two.

[0028] High-temperature alloy components have wide applications in aerospace and other fields. With technological advancements, critical equipment in the aerospace sector is evolving towards higher reliability, longer lifespan, and higher parameterization, placing higher demands on the fatigue resistance of complex high-temperature alloy components. Traditional fatigue life prediction methods mainly include nominal stress methods, local stress-strain methods, and energy methods. These methods typically predict fatigue life based on single factors such as stress, strain, and energy. However, fatigue life exhibits significant dispersion, with the same material showing marked differences in fatigue life across different test batches under the same processing conditions. Furthermore, traditional methods have large errors when predicting fatigue life of structures with stress concentrations, and most rely on material performance data such as SN curves and CN curves, which require extensive experimental work and significant costs. Currently, some technical solutions introduce machine learning to predict material processing, microstructure, and properties. However, the training effect of machine learning, especially deep learning, is highly dependent on the sample size. When the number of samples used for training is small, the robustness and generalization ability of the trained model are poor. For fatigue life prediction, especially for samples with high material costs and complex test specimen structures, it is difficult to accumulate data samples that can well meet the model training requirements by increasing the scale of the test. This greatly limits the use of machine learning for fatigue life prediction.

[0029] To address the aforementioned issues, embodiments of the present invention provide a method for predicting residual fatigue life. By extracting features from different time periods of the same fatigue test on a specimen, the number of samples is increased, effectively reducing the sample quantity requirement for model training.

[0030] The method includes the following steps:

[0031] Step a): Provide a fatigue test dataset, which includes hysteresis loop data for each cycle throughout the entire fatigue test cycle of multiple test specimens. In a preferred embodiment, the fatigue test dataset is filtered to remove sample data where strain peak and valley values ​​exceed the specified limits or the number of cycles is insufficient. Insufficient cycle count refers to data where the number of fatigue cycles at sample failure is significantly less than the normal level, such as exceeding the 3σ criterion or being shorter than the lower limit of the fatigue life of a qualified test specimen. Extract data from the fatigue test dataset to determine the hysteresis loop area, damping ratio, shear stiffness, elastic modulus, etc., for each hysteresis loop, as well as the remaining fatigue life for each cycle (i.e., total fatigue life minus the current cycle).

[0032] Step b) Establish a training subset by pairing data from different cycles of the same test piece into hysteresis loop data pairs. Merge the training subsets of each test piece to obtain the total training set, which includes an input set X = {x1, x2, Δ} and an output set Y, where x1 is the hysteresis loop data from the earlier cycle in the hysteresis loop data pair, x2 is the hysteresis loop data from the later cycle in the hysteresis loop data pair, Δ is the cycle interval between x1 and x2, and Y is the remaining fatigue life of the cycle containing x2. In a preferred embodiment, 100 ≤ Δ ≤ 500. Based on engineering experience, when the cycle interval is too small, it cannot accurately reflect the changing trend, affecting the training effect; while when the cycle interval is too large, it will result in too few usable hysteresis loop data pairs and insufficient samples. To eliminate the influence of dimensions, the data in the total training set is normalized.

[0033] Step c): Provide a machine learning algorithm model and train it using the training set to obtain a prediction model. In a preferred embodiment, the machine learning algorithm model selects the most suitable model from multiple different models using K-fold cross-validation. During training, a loss function and optimizer are set for the machine learning algorithm model. In a preferred embodiment, the loss function can be Smooth L1, and the optimizer can be Adam. The trained prediction model predicts fatigue life based on hysteresis loop data in the fatigue test dataset. The prediction results and actual fatigue life data are plotted as a scatter plot. The prediction results should be within a 2x scatter band, and in a further preferred embodiment, within a 1.5x scatter band.

[0034] Step d): Input the hysteresis loop data and cycle interval of the target part for which fatigue life prediction is required into the prediction model to predict the fatigue life of the target part.

[0035] Another embodiment of the present invention provides a computing device including a memory and a processor, wherein the memory stores a calculation program that, when executed by the processor, can implement the remaining fatigue life prediction method provided in any of the foregoing embodiments. In different embodiments, the computing device may be configured as a general-purpose computer, a dedicated computing device such as a microcontroller, or a cloud computing device or a virtual machine.

[0036] In a preferred embodiment, the remaining fatigue life of IN718 alloy parts is predicted using the remaining fatigue life prediction method provided in the foregoing embodiments.

[0037] First, a fatigue test dataset containing fatigue test data from 57 specimens is provided, including 19 specimens each with strain ratios of -1, 0, and 0.6. The dataset includes hysteresis loop data for each specimen in each cycle. The fatigue test dataset is then filtered, and the peak and valley values ​​of the isometric strain for each cycle in the fatigue test are plotted for each specimen. All sample cycle periods meet the requirements, but data with strain values ​​exceeding the tolerance and not conforming to the test design requirements need to be removed. For example, ... Figure 3 The peak and valley values ​​in some cycles of sample D shown were too low and did not meet the experimental design requirements, so they needed to be removed. After screening, a fatigue test dataset consisting of 52 samples was finally obtained.

[0038] Next, the hysteresis loop data for each cycle of each test piece are extracted, and the hysteresis loop characteristics are as follows: Figure 1 As shown, information such as hysteresis loop area, damping ratio, and secant stiffness are extracted from the data. The cycle number of the hysteresis loop and the remaining fatigue life in that cycle are recorded to establish a training subset for each test specimen. Here, the hysteresis loop area is the area of ​​the red region enclosed by the hysteresis loop, the secant stiffness Ed is the slope of the line connecting peak point A and peak point B, and the damping ratio λ is the hysteresis loop area / (4π × S). ΔAOC In triangle AOC, O is the midpoint of AB, and OC and AC are parallel to the σ-axis and ε-axis, respectively. Feature extraction is performed on the hysteresis loop of one cycle at one time period, such as... Figure 2 As shown, characteristic data such as maximum stress, minimum stress, maximum strain, minimum strain, and hysteresis loop area can be obtained. Hysteresis loop data extracted from different cycles of the same test piece are paired to obtain hysteresis loop data pairs (x1, x2, Δ), where x1 represents the hysteresis loop data from an earlier cycle, x2 represents the hysteresis loop data from a later cycle, and Δ is the cycle interval between x1 and x2. Establishing hysteresis loop data pairs effectively increases the sample size. Δ is set in the range of 100-500, and 2411 hysteresis loop data pairs are constructed from the fatigue test data of 52 test pieces to form a training set. The input set X = {x1, x2, Δ}, and the output set Y is the remaining fatigue life of the cycle containing x2. To eliminate dimensional differences and reduce model error, X is normalized; considering the large difference in fatigue life values, Y is logarithmically normalized.

[0039] Next, the machine learning algorithm model is trained. In different embodiments, different machine learning algorithm models such as SVW, XGBoost, random forest, or neural networks can be selected.

[0040] In a preferred embodiment, K-fold cross-validation is used to select the most suitable model. The selection method is as follows: First, the total training set is divided into K equal parts. Then, K loops are executed. In each loop, one part of the data is used as the test set, one part as the validation set, and the remaining K-2 parts as the training set. In each loop, the selected data is used to train different machine learning algorithm models, and the validation set data is used to evaluate the performance of each model. The mean squared error (MSE) is used to measure the predictive ability of different models. After K loops, the model with the lowest mean MSE and the smallest variance is selected as the prediction model.

[0041]

[0042]

[0043] Table 1 Comparison of experimental results for ANNs with different structures

[0044] In this embodiment, K=10, and different multi-branch artificial neural network (ANN) models are selected. The MSE results after the loop are shown in Table 1. Among them, the optimal model structure is 8-16-32-16, which is used as the prediction model.

[0045] Next, we set the hyperparameters for the prediction model. The training set was randomly divided into training, test, and validation sets. A suitable loss function was chosen. Since Smooth L1 Loss has a small gradient near outliers, making the model less sensitive to these outliers, and considering the strong dispersion of metal fatigue life with many outliers, Smooth L1 was chosen as the loss function to give the model higher tolerance for outliers and make the training more robust. Adam was chosen as the optimizer because it can adaptively adjust the learning rate of each parameter, thereby improving the model's convergence speed and generalization ability. The model was trained using the training and validation sets, with 200 epochs to ensure optimal training results. The epoch with the highest accuracy was then selected as the final prediction model.

[0046] The final prediction model is used to predict the remaining fatigue life of the data in the test set. The model's performance is then verified, and the actual data of the test set and the prediction results of the final prediction model are plotted as a scatter plot. Figure 4 As shown, the prediction results all fall within 1.5 times the dispersion zone, indicating high accuracy.

[0047] In the comparative example, no hysteresis loop data pairs were established; fatigue test data from the same batch of 52 test pieces were directly used as training samples for model training. The box plot comparing the prediction results of the comparative example and the implementation example is shown below. Figure 5As shown in the figure, examples 1 and 2 are comparative examples, and example 3 is an example. Comparative example 1 uses early fatigue hysteresis loop data and fatigue life for model training, while comparative example 2 uses hysteresis loop data at half the fatigue life for model training. In the figure, the bars represent standard deviation, the upper and lower limit lines represent extreme values, and the broken line represents the corresponding mean MSE. It can be seen that compared to the comparative examples, the example has better convergence and higher generalization.

[0048] The final prediction model from the example and the prediction model from the comparative example were each used to perform 10 predictions, and the results are as follows: Figure 6 As shown, the bar chart represents the standard deviation, and the limit lines at the top and bottom of the bar chart represent extreme values. It can be seen that when repeated predictions are performed, the prediction results of Comparative Example 2 exhibit significant fluctuations, with both extreme values ​​and standard deviations being much larger than those of the Example. The MSE fluctuation range of the Example's prediction results is around 0.01, and no abnormal discrete points appear, indicating that the surface model has far superior stability compared to the Comparative Example.

[0049] The purpose of the above embodiments is to provide a further detailed description of the present invention in conjunction with the accompanying drawings, so that those skilled in the art can understand the technical concept of the present invention. Within the scope of the present invention, optimization or equivalent substitution of the method steps involved, as well as combination of implementation methods in different embodiments without causing a conflict of principles, all fall within the protection scope of the present invention.

Claims

1. A method for predicting remaining fatigue life, characterized in that, Includes the following steps: Step a): Provide a fatigue test dataset, which includes hysteresis loop data of multiple test specimens in each cycle of the fatigue test; Step b): Establish a training subset, which includes hysteresis loop data pairs from multiple different cycles in the same test piece, and the hysteresis loop data pairs include two hysteresis loop data pairs spaced apart by a given number of cycles; The training subsets of each test piece are combined into a training set, which includes an input set X = {x1, x2, Δ} and an output set Y, where x1 is the hysteresis loop data with an earlier cycle in the hysteresis loop data pair, x2 is the hysteresis loop data with a later cycle in the hysteresis loop data pair, Δ is the interval between x1 and x2, and Y is the remaining fatigue life of the cycle in which x2 is located. Step c): Provide a machine learning algorithm model, input the training set into the machine learning algorithm model for training, and obtain a prediction model; Step d): Provide hysteresis loop data of the target component and use the prediction model to predict the fatigue life of the target component.

2. The method for predicting remaining fatigue life according to claim 1, characterized in that, In step a), the hysteresis loop data includes the hysteresis loop area, damping ratio, and secant stiffness.

3. The method for predicting remaining fatigue life according to claim 1 or 2, characterized in that, Step a) further includes the step of removing data from the fatigue test dataset that have excessive strain peak and valley values ​​or insufficient cycle counts.

4. The method for predicting remaining fatigue life according to claim 1 or 2, characterized in that, In step b), 100 ≤ Δ ≤ 500.

5. The method for predicting remaining fatigue life according to claim 1 or 2, characterized in that, Step b) further includes a step of normalizing the data in the training set.

6. The method for predicting remaining fatigue life according to claim 1 or 2, characterized in that, Step c) further includes a step of screening the machine learning algorithm model: providing multiple machine learning algorithm models and using K-fold cross-validation to select the machine learning algorithm model with the lowest mean squared error and the smallest variance.

7. The method for predicting remaining fatigue life according to claim 6, characterized in that, The plurality of machine learning algorithm models include all or a combination of SVW, XGBoost, random forest, and neural networks.

8. The method for predicting remaining fatigue life according to claim 1 or 2, characterized in that, In step c), the machine learning algorithm is trained, with the loss function configured as Smooth L1 and the optimizer configured as Adam.

9. The method for predicting remaining fatigue life according to claim 1 or 2, characterized in that, The fatigue life prediction results made by the prediction model using the fatigue test dataset are within a 2x dispersion band compared to the actual results.

10. A computing device, comprising a memory and a processor, characterized in that, The memory stores a calculation program that, when executed by the processor, can implement the remaining fatigue life prediction method as described in any one of claims 1 to 9.

Images