High-temperature S-N curve prediction method based on transfer learning, medium, equipment and product

Abstract

The invention provides a high-temperature S-N curve prediction method based on transfer learning, a medium, equipment and a product. The method comprises the following steps: acquiring fatigue performance data at different temperatures obtained through a fatigue performance test in advance; performing preset processing on the obtained fatigue performance data at different temperatures to obtain sample data at room temperature and sample data at high temperature; performing preliminary training of a gated loop network model by using the sample data at the room temperature to obtain a transfer learning source model; establishing a migration model based on the migration learning source model, and training the migration model by using the sample data at the high temperature to obtain a high-temperature S-N curve prediction model; and performing cyclic prediction by using the obtained high-temperature S-N curve prediction model, and outputting a high-temperature S-N curve. According to the scheme provided by the invention, the high-temperature S-N curve can be accurately predicted under the condition of an extremely small sample.

Description

Technical Field

[0001] The present invention relates to the field of control, and in particular to a high-temperature SN curve prediction method based on transfer learning, a medium, a device, and a program product. Specifically, it relates to a high-temperature SN curve prediction method based on a transfer learning-guided gated recurrent network, a computer-readable storage medium, a computing device, and a computer program product. Background Art

[0002] As a key structural component, the main shaft bearing of an aircraft engine is the core of the development of aviation technology at home and abroad in recent years. Its reliability directly affects the operating safety of the engine and even the entire aircraft. The main failure form of bearings in actual working conditions is fatigue failure. Bearing steel is a key factor in determining bearing performance. Accurately predicting the fatigue life of bearing steel is very important for the anti-fatigue design and performance evaluation of bearing steel.

[0003] A basic fatigue assessment method currently in common use is the SN curve method, which is used to describe the stress-life relationship of a standard specimen under certain loading conditions by fitting the stress data obtained from fatigue tests. However, the method of obtaining the SN curve through fatigue testing in related technologies, firstly, is costly and time-consuming, and usually involves a series of test parameters, making fatigue performance assessment more complicated; secondly, for some new materials, the published fatigue performance data is often limited, and the scarcity of data makes it impossible to make accurate predictions and assessments of their fatigue performance.

[0004] Therefore, how to provide a highly accurate and efficient high-temperature SN prediction method for bearing steel has become a technical problem that needs to be solved urgently in this field. Summary of the invention

[0005] The main purpose of the present invention is to overcome the defects of the above-mentioned related technologies and provide a high-temperature SN curve prediction method, medium, equipment and product based on transfer learning to solve the problem that the scarcity of data in the related technologies makes it impossible to make accurate predictions and evaluations of fatigue performance.

[0006] On the one hand, the present invention provides an SN curve prediction method based on transfer learning, including: obtaining fatigue performance data at different temperatures obtained in advance through fatigue performance tests; the fatigue performance data include: stress amplitude and cycle number; pre-processing the obtained fatigue performance data at different temperatures to obtain sample data at different temperatures; the sample data at different temperatures include: sample data at room temperature and sample data at high temperature; using the sample data at room temperature to perform preliminary training of a gated recurrent network model to obtain a transfer learning source model; the room temperature includes the temperature that can be reached indoors; the high temperature includes the temperature higher than room temperature; establishing a transfer model based on the transfer learning source model, and training the transfer model using the sample data at high temperature to obtain a high-temperature SN curve prediction model; using the obtained high-temperature SN curve prediction model to perform cyclic prediction and output a high-temperature SN curve.

[0007] Optionally, the fatigue performance test includes at least one of rotational bending fatigue, tension-compression fatigue and torsional fatigue.

[0008] Optionally, the fatigue performance data obtained at different temperatures are subjected to preset processing, including: interpolating the stress amplitude data in the fatigue performance data at different temperatures to unify the data format and expand the data volume; wherein the stress amplitude data obtained after the interpolation processing are equally spaced in the direction of the cycle number; standardizing the stress amplitude data obtained after the interpolation processing to convert them into dimensionless data; and serializing the stress amplitude data after the interpolation processing and the standardization processing, wherein the sequence length of the stress amplitude sequence is set according to the data volume.

[0009] Optionally, when performing interpolation processing, the cycle range is set to: the cycle range at room temperature is set to 9×10 4 Up to 1×10 7 ; and / or, the cycle range at 100°C is set to 7×10 4 Up to 1×10 7 ; and / or, the cycle range at 200°C is set to 1×10 5 Up to 1×10 7 ; and / or, the cycle range at 300°C is set to 3×10 4 Up to 1×10 7 ; and / or, the data volume range of the stress amplitude obtained after interpolation processing is set to 800 to 1500; and / or, the sequence length of the stress amplitude sequence is set to 10 to 80 data.

[0010] Optionally, the interpolation processing includes: interpolating the stress amplitude data using a linear interpolation method; and / or the normalization processing includes: mapping the range of the stress amplitude data to a range of 0 to 1 using a Min-Max normalization method.

[0011] Optionally, it also includes: dividing the sample data at different temperatures into training sets, validation sets and test sets according to different preset ratios, respectively, to obtain room temperature training sets, room temperature validation sets and room temperature test sets, as well as high temperature training sets, high temperature validation sets and high temperature test sets; using the sample data at room temperature to perform preliminary training and adjustment of the gated recurrent network model to obtain a transfer learning source model, including: constructing gated recurrent network models of different structures, and setting different numbers of neurons for the gated recurrent network models of different structures; using the room temperature training set to train the gated recurrent network models of different structures respectively; for the gated recurrent network models of different structures, by evaluating their performance on the room temperature validation set, determining the structure with the best performance and adjusting its hyperparameters; using the adjusted hyperparameters, reusing the room temperature training set and room temperature validation set to perform model training on the structure with the best performance to obtain a transfer learning source model; using the transfer learning source model to perform predictions on the room temperature test set to evaluate the performance of the transfer learning source model.

[0012] Optionally, the constructed gated recurrent network model includes a GRU layer, a fully connected layer and an output layer; wherein the number of neurons in the GRU layer is 20; the fully connected layer is a double-layer structure, the number of neurons in each layer is 20, and the number of neurons in the output layer is 1.

[0013] Optionally, a migration model is established based on the migration learning source model, including: establishing a target gated recurrent network model with the same network structure as the migration learning source model; copying the GRU layer of the migration learning source model to the GRU layer of the target gated recurrent network model to obtain a migration layer, and freezing the migration layer; randomly initializing the fully connected layer of the target gated recurrent network model and training the fully connected layer using the sample data at the high temperature to obtain a migration model.

[0014] Optionally, it also includes: dividing the sample data at different temperatures into a training set, a validation set and a test set according to different preset ratios, respectively, to obtain a room temperature training set, a room temperature validation set and a room temperature test set, as well as a high temperature training set, a high temperature validation set and a high temperature test set; using the obtained high temperature SN curve prediction model to perform cyclic prediction and output a high temperature SN curve, including: inputting the last stress amplitude sequence of the high temperature validation set with a length of L into the prediction model, and outputting the predicted first stress amplitude of the high temperature test set; sliding the window backward by a data length, inputting the last L-1 stress amplitudes of the high temperature validation set and the predicted first stress amplitude of the high temperature test set into the prediction model, and outputting the predicted second stress amplitude of the high temperature test set; and so on, cyclically outputting the prediction results of the prediction model for the high temperature test set until the target cycle number is reached to obtain a high temperature SN curve.

[0015] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the program implements the steps of any of the aforementioned methods when executed by a processor.

[0016] In another aspect, the present invention provides a computing device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the aforementioned methods when executing the program.

[0017] In yet another aspect, the present invention provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the steps of any of the aforementioned methods are implemented.

[0018] The technical solution according to the present invention has the following beneficial effects:

[0019] (1) This method is not limited by factors such as materials, processes, and service environment. It can predict high fatigue SN curves of bearings at various temperatures with high accuracy, demonstrating the high versatility of this method.

[0020] (2) Compared with traditional tests, this method can greatly reduce time and economic costs while ensuring high accuracy, and improve the efficiency of fatigue performance testing and evaluation.

[0021] (3) Compared with the SN curve prediction method that only uses a gated recurrent network, the gated recurrent network model based on transfer learning established by this method has a high tolerance for data volume, accurate prediction, and stable performance. It has higher application value and can provide reliable support for the design, manufacturing, and performance evaluation of high-end bearings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0023] Figure 1 It is a method schematic diagram of an embodiment of a high temperature SN curve prediction method provided by the present invention;

[0024] Figure 2 A schematic diagram showing a specific implementation of the steps of performing preliminary training of a gated recurrent network model based on fatigue performance data at room temperature to obtain a transfer learning source model;

[0025] Figure 3 Schematic diagram of the gated recurrent unit;

[0026] Figure 4 A schematic diagram of the SN curve prediction results at 100°C is shown;

[0027] Figure 5 A schematic diagram of the SN curve prediction results at 200°C is shown;

[0028] Figure 6 A schematic diagram of the SN curve prediction results at 300°C is shown;

[0029] Figure 7 A schematic diagram showing statistical results of model performance according to a specific example of the present invention is shown;

[0030] Figure 8 It is a method schematic diagram of a specific embodiment of the high temperature SN curve prediction method provided by the present invention;

[0031] Fig. 9 It is a schematic diagram of the SN curve cycle prediction method. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0034] Given the high-temperature service characteristics of some bearing steels, it is still a huge challenge to accurately predict their high-temperature fatigue performance using limited existing data, so this method still has room for improvement. As an emerging technical means, deep learning has gradually been applied to the field of materials research, and many traditional materials research methods are also constantly evolving and innovating by combining deep learning methods. However, there is currently no method that can solve the problem of high-temperature fatigue performance prediction of bearing steels with extremely small samples.

[0035] Gated Recurrent Unit (GRU) is a variant of Recurrent Neural Network (RNN), which is mainly used to deal with sequential or time problems. It is widely used in natural language processing, language translation, text classification and other fields. In practical applications, RNN has two serious shortcomings, namely gradient vanishing and gradient exploding, which make it unable to effectively handle dependencies in time series in some cases. The emergence of GRU aims to improve the gradient problem in RNN. The schematic diagram of its internal structure is shown in the figure below. Figure 2 Similar to RNN, GRU can also transfer information between hidden layers, but compared with RNN, GRU introduces two gating signals, namely the reset gate and the update gate. The two gating signals are used to control the network's memory and forgetting of time series information. Compared with RNN and other variants, GRU has a simple structure, excellent performance, faster model convergence speed, and is easier to train. The SN curve can be regarded as serialized data. Therefore, in theory, by using GRU to learn the rules of the early stage of SN data, it is possible to predict the curve in the later stage. In practical applications, it can reduce the sample demand for fatigue testing.

[0036] Transfer learning is a machine learning strategy that applies the knowledge or model learned in a certain task to the learning of another similar task, thereby improving the learning efficiency and performance of the model and achieving "learning from one example and applying it to other situations". Transfer learning is often used to solve problems such as data scarcity, difficulty in labeling, and low efficiency, and reduces learning costs by using existing labeled data and learned models. The fatigue performance of bearing steel at different temperatures can be regarded as related tasks. Therefore, by combining transfer learning strategies and gated recurrent networks, the guiding role of domain knowledge between SN curves at different temperatures can be fully utilized, and further accurate prediction of SN curves at different temperatures can be achieved with extremely small samples.

[0037] Figure 1 It is a schematic diagram of an embodiment of a method for predicting a high-temperature SN curve provided by the present invention.

[0038] like Figure 1 As shown, according to one embodiment of the present invention, the prediction method at least includes step S110, step S120, step S130, step S140 and step S150.

[0039] Step S110, obtaining fatigue performance data at different temperatures obtained in advance through fatigue performance tests.

[0040] Fatigue performance tests are performed at different temperatures in advance to obtain fatigue performance data at different temperatures. The fatigue performance data at different temperatures may specifically include: fatigue performance data at room temperature and fatigue performance data at high temperature. The fatigue performance test may specifically include: at least one of rotational bending fatigue, tension and compression fatigue, and torsional fatigue. The fatigue performance data may specifically include: stress amplitude σ and number of cycles N. The test material of the fatigue performance test may specifically be a metal material, such as bearing steel.

[0041] For example, M50 aero-engine bearing steel is used, which has good high temperature stability and wear resistance and is currently the main steel used in aero-engine bearings. The specific composition is shown in Table 1:

[0042] Table 1

[0043]

[0044] The bearing steel was subjected to rotary bending fatigue tests at room temperature and high temperatures of 100°C, 200°C, and 300°C, and 15 pairs of data points were obtained at different temperatures.

[0045] Step S120, performing preset processing on the acquired fatigue performance data at different temperatures to obtain sample data at different temperatures.

[0046] The preset processing may specifically include: interpolation processing, normalization processing and serialization processing. The sample data at different temperatures may specifically include: sample data at high temperature and sample data at room temperature. The room temperature includes the temperature that can be reached indoors, that is, normal temperature, such as 25°C. The high temperature includes a temperature higher than room temperature, such as 100°C, 200°C, and 300°C in the aforementioned experiment.

[0047] (1) interpolating the fatigue performance data obtained at different temperatures to unify the data format and expand the data volume;

[0048] Specifically, the stress amplitude σ data at different temperatures are interpolated to unify the data format and expand the data volume. Since GRU requires equal intervals in time series when processing time series data, and the data obtained through fatigue testing is relatively discrete and the intervals are unequal, it is necessary to unify the format of the input data. In a specific implementation, a linear interpolation method is used to unify the data format and expand the number of samples. The linear interpolation formula is as follows:

[0049]

[0050] Among them, (x, y) is the interpolation point to be sought, (x 0 ,y 0 ) and (x 1 ,y 1 ) are the known points on the left and right sides of the interpolation point.

[0051] Before performing linear interpolation, it is necessary to set the range of different temperature cycle times N according to the obtained fatigue performance test data. For example, the cycle range of the room temperature SN curve is set to 9×10 4 Up to 1×10 7 The cycle range of the SN curve at 100°C is set to 7×10 4 Up to 1×10 7 The cycle range of the SN curve at 200°C is set to 1×10 5 Up to 1×10 7 The cycle range of the SN curve at 300°C is set to 3×10 4 Up to 1×10 7 . The stress amplitude data obtained after interpolation processing are equally spaced in the direction of the cycle number. The amount of stress amplitude data obtained after interpolation processing can be set to 800 to 1500. For example, the amount of stress amplitude σ data samples obtained by linear interpolation is set to 800.

[0052] (2) Standardizing the stress amplitude data obtained after interpolation and converting them into dimensionless data;

[0053] Gated recurrent units are sensitive to the range of input data, so the data generated by linear interpolation needs to be standardized. Data standardization refers to scaling the original data and converting it into dimensionless data. The standardized data set is more standardized and is conducive to the convergence of the model. Common data standardization methods include: Min-Max standardization, log function conversion, z-score standardization, etc.

[0054] Min-Max standardization, also known as deviation standardization, is based on linear transformation of the original data and then mapping it to the interval [0,1]; the principle of log function conversion is to normalize the original data through a log function with a base of 10; z-score standardization is also called standard deviation standardization, and the original data after processing conforms to the standard normal distribution with a mean of 0 and a variance of 1. In this embodiment, the Min-Max standardization method is selected, and its conversion formula is as follows:

[0055]

[0056] Among them, X max is the maximum value of sample data under a certain dimension, X min is the minimum value of the sample data under a certain dimension. The data standardization step is implemented with the help of the MinMaxScaler function in the scikit-learn function library.

[0057] (3) The stress amplitude data after interpolation and standardization are serialized, wherein the sequence length of the stress amplitude sequence is set according to the amount of data.

[0058] The stress amplitude σ needs to be input into the gated recurrent network in the form of serialized data. The network learns the time dependency in the sequence and predicts the later trend of the SN curve. The sequence length L can be flexibly set according to the number of samples. For example, if L is set to 50, the model outputs the predicted value of the target stress amplitude by learning the 50 stress amplitudes before the stress amplitude corresponding to the target cycle.

[0059] Step S130, using the sample data at room temperature to perform preliminary training on a gated recurrent network model to obtain a transfer learning source model.

[0060] In order to avoid overfitting and underfitting of the model and improve the accuracy of the model, the sample data obtained at different temperatures are first divided into training sets, validation sets and test sets according to different preset ratios, respectively, to obtain room temperature training sets, room temperature validation sets and room temperature test sets, as well as high temperature training sets, high temperature validation sets and high temperature test sets. That is, the SN curve data at different temperatures are divided into training sets, validation sets and test sets according to different ratios. Preferably, the front segment data is used as the training set, the middle segment data is used as the validation set, and the back segment data is used as the test set. Since the cycle stress gradually decreases from front to back in the curve, the cycle life gradually increases. Therefore, the network is trained using the front segment stress data, and then the back segment data is output, which can reduce the time cost of the experiment. The source model of the transfer learning is pre-trained using the room temperature SN curve data, and the division ratio of the training set, validation set and test set is set to 4:1:5; the high temperature SN curve data is used to retrain the fully connected layer of the transfer model, and the corresponding division ratio is set to 2:1:7. Table 2 shows an example of the percentage of the training set, validation set and test set in the sample data at different temperatures according to the present invention.

[0061] Table 2

[0062]

[0063] The training set is used to perform preliminary training on the network. After the preliminary training is completed, the validation set is used to verify the network performance, adjust the network's hyperparameters, obtain the best model, and finally predict the test set to evaluate the network performance. In practical applications, according to the embodiment of the present invention, the corresponding temperature annotated data with sufficient sample size can be selected to train the source model of the migration model, and then the obtained model can be applied to the SN curve prediction task of other temperatures with the help of transfer learning.

[0064] Figure 2 A schematic diagram showing a specific implementation of the steps of performing preliminary training of a gated recurrent network model based on fatigue performance data at room temperature to obtain a transfer learning source model.

[0065] like Figure 2 As shown, in a specific implementation, step S130 includes: step S131, step S132, step S133, step S134 and step S135.

[0066] Step S131, constructing gated recurrent network models with different structures, and setting different numbers of neurons for the gated recurrent network models with different structures.

[0067] Figure 3 This is a schematic diagram of the principle of the gated recurrent unit. Figure 3As shown in the figure, compared with RNN, GRU adds two gating signal settings, namely reset gate r and update gate z, h is the hidden state of stress amplitude, is the candidate hidden state of the stress amplitude. In GRU, the long-term and short-term dependencies in the stress amplitude sequence can be captured by the reset gate and update gate. Specifically, the reset gate determines how much of the input information at the current moment is combined with the input information at the previous moment. The candidate hidden state in GRU can be calculated by resetting the gate r. This enables the capture of short-term dependencies in the stress amplitude series, and the calculation method is as follows:

[0068] r r =σ(W r ·[x t ,h t-1 ])

[0069]

[0070] Among them, W r is the weight matrix of the reset gate, W is the weight matrix of the candidate hidden state, x t is the stress amplitude σ at the current moment, h t-1 is the hidden state of σ at the previous moment, ⊙ is the Hadamard product. σ is the Sigmoid function, whose output range is [0, 1], and tanh is the hyperbolic tangent function, whose output range is [-1, 1].

[0071] These two are two common activation functions in machine learning, and their expressions are as follows:

[0072]

[0073]

[0074] The update gate determines how much of the input information of the previous moment of the network is saved to the current moment. The hidden state h in the GRU can be calculated through the update gate z, thereby capturing the long-term dependency relationship in the amplitude sequence. The calculation method is as follows:

[0075] z t =σ(W z ·[x t ,h t-1 ])

[0076]

[0077] Among them, W z is the weight matrix of the updated gate, and ht is the hidden state of σ at the current moment. With the help of two gating signals, GRU can accurately capture important patterns in the stress amplitude sequence.

[0078] In order to find a better network structure and improve network performance, different structures of gated recurrent network models are set in this embodiment. The network uniformly includes a GRU layer, a fully connected layer and an output layer, where the different structures refer to a single-layer structure of the fully connected layer and a double-layer structure of the fully connected layer. Different numbers of neurons in the fully connected layer are set for the gated recurrent network models with different structures. The number of neurons in the GRU layer is uniformly set to 20, and the output is 1. The specific network structure established is shown in Table 3:

[0079] Table 3

[0080]

[0081] The gated recurrent networks with different structures were preliminarily trained. The model optimizer used Adam, the learning rate was set to 0.001, and the loss function was Mean_Square_Error. By comparing the model performance, when the number of neurons was equal, the performance of the double-layer structure was slightly better than that of the single-layer structure. Therefore, considering comprehensively, the preferred transfer learning source model structure in the present invention is 20×20×20×1.

[0082] Step S132, using the room temperature training set to train the gated recurrent network models of different structures respectively.

[0083] The training set of room temperature SN curves is used to perform preliminary training on gated recurrent networks with different structures.

[0084] Step S133, for the gated recurrent network models with different structures, determine the structure with the best performance by evaluating their performance on the room temperature validation set, and adjust its hyperparameters.

[0085] The performance of the network is evaluated on the validation set of the room temperature curve, and the network hyperparameters are adjusted; by observing the performance of the model on the validation set after training, the network structure and other hyperparameters of the network can be adjusted. For models with different structures, after training on the training set, by comparing their performance on the validation set, the best performing structure can be selected, and the hyperparameters can be adjusted to obtain the best hyperparameters. The evaluation index of the model can be, for example, the determination coefficient R 2 , mean absolute error (MAE) and root mean square error (RMSE).

[0086] Step S134, using the adjusted hyperparameters, reusing the room temperature training set and the room temperature validation set to perform model training on the structure with the best performance to obtain a transfer learning source model.

[0087] After determining the best-performing structure and adjusting the hyperparameters, the adjusted hyperparameters are used to re-use the room temperature training set and the room temperature validation set to train the model of the best-performing structure, thereby obtaining the best transfer learning source model.

[0088] Step S135, using the transfer learning source model to perform prediction on the room temperature test set to evaluate the performance of the transfer learning source model.

[0089] Specifically, the prediction performance of the model is evaluated using a test set of sample data at room temperature.

[0090] Step S140, establishing a migration model based on the migration learning source model, and training the migration model using the sample data under the high temperature to obtain a high temperature SN curve prediction model.

[0091] In a specific implementation, a target gated recurrent network model with the same network structure as the transfer learning source model is established; the GRU layer of the transfer learning source model is copied to the GRU layer of the target gated recurrent network model to obtain a transfer layer, and the transfer layer is frozen so that it does not participate in the subsequent model training; the fully connected layer of the target gated recurrent network model is randomly initialized, and the fully connected layer is trained using the sample data at the high temperature to obtain the transfer model.

[0092] Specifically, a prediction model of SN curves at different temperatures (target gated recurrent network model) is established. The network structure is the same as the migration source model, consisting of a GRU layer, a fully connected layer, and an output layer. The number of neurons is also the same as the migration source model. For example, the number of neurons in the GRU layer is 20, the fully connected layer is a two-layer structure, the number of neurons in each layer is 20, and the number of output layers is 1. The GRU layer of the obtained transfer learning source model is copied to the corresponding layer of the above model, which is called the transfer layer; the fully connected layer of the above model is randomly initialized to obtain the migration model.

[0093] During the training process of the migration model, the GRU layer is kept frozen so that it does not participate in subsequent training. The fully connected layer of the migration model is trained using sample data at high temperatures, and the network hyperparameters are adjusted. Finally, a high-temperature SN curve prediction model is obtained. The specific training process is the same as the training process of the aforementioned migration source model.

[0094] Step S150, using the obtained high-temperature SN curve prediction model to perform cyclic prediction and output a high-temperature SN curve.

[0095] Specifically, the high-temperature SN curve may include, for example, SN curves at 100°C, 200°C, and 300°C, that is, the prediction results of the SN curves at 100°C, 200°C, and 300°C are output on the test set, respectively. Because the test data of the late stage of the curve is unknown in actual applications, and only very little high stress data is available, the output strategy of the present invention on the test set is: input the last stress amplitude sequence of length L of the high-temperature verification set into the prediction model, and output the first stress amplitude of the predicted high-temperature test set; input the last L-1 stress amplitudes of the high-temperature verification set and the predicted first stress amplitude of the high-temperature test set into the prediction model, and output the second stress amplitude of the predicted high-temperature test set; and so on, cyclically output the prediction results of the prediction model for the high-temperature test set until the target cycle is reached to obtain the high-temperature SN curve.

[0096] The method of applying the cyclic output outputs the prediction results of the test set, so the first result of the test set should be the prediction of the last L data in the input validation set, and then the window moves back to generate the 2nd to nth results in turn. The improvement here takes into account the training and output logic of the GRU network, implements the sample data in the front section of the GRU learning curve to obtain prediction capabilities, and outputs all the data in the back section of the curve.

[0097] For example, when the length of the stress amplitude sequence is L = 50, the last L data of the validation set are input into the network as a sequence to obtain the first stress amplitude of the test set. At the same time, the window slides backward by one unit, and the last 49 stress amplitudes of the validation set and the predicted first stress amplitude of the test set are input into the network to obtain the second stress amplitude of the test set. On this basis, the prediction results of the network for the test set are cyclically output until the target cycle number, and the high-temperature SN curve is obtained.

[0098] Fig. 9 A schematic diagram showing the use of a prediction model for prediction. Fig. 9 As shown, assuming that the stress amplitude sequence length L = 2, points 1 and 2 are the last two points of the validation set, the model reads the last two points of the validation set (i.e., points 1 and 2) and predicts the first point of the test set, i.e., point 3, and then the window slides back one (data) point, the model reads the last point of the validation set (i.e., point 2) and the first point of the test set predicted, the second point of the test set predicted, i.e., point 4, and then the window slides back one (data) point, the model reads the first two points of the predicted test set (i.e., points 3 and 4), the third point of the test set predicted, i.e., point 5, and so on.

[0099] The front data divided according to the preset ratio is the training set, which has high cycle stress σ and short cycle life N; the back data is the test set, which has low cycle stress σ and longer cycle life N. Therefore, if the trend of the front part of the learning curve can be predicted by the trend of the back part of the curve, a complete SN curve can be obtained. Therefore, only fatigue tests under high stress are needed to reduce the time required for the test. At the same time, compared with the test under the full stress range, the number of samples required for fatigue tests can be reduced.

[0100] Using the coefficient of determination R 2 , Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) are used to evaluate the performance of the model. The corresponding calculation methods are shown in Table 4:

[0101] Table 4

[0102]

[0103] Among them, y i,true is the true value, y i,pre is the predicted value, for y i,true The average value of .

[0104] Figures 4 to 6 A schematic diagram of the prediction results of SN curves at different temperatures based on the above parameters is shown. Figure 4 A schematic diagram of the SN curve prediction results at 100°C is shown; Figure 5 A schematic diagram of the SN curve prediction results at 200°C is shown; Figure 6 A schematic diagram of the SN curve prediction results at 300°C is shown. In order to demonstrate the advantages of this method, the prediction results of the migration gated recurrent network are compared with the results of the non-migration gated recurrent network. As can be seen from the figure, the method proposed in the present invention can realize the prediction of SN curves at different temperatures under extremely small sample conditions, has extremely high accuracy, and can improve data utilization. At the same time, this method can achieve accurate predictions for different test temperature conditions and has good versatility. Compared with the method that does not use transfer learning, the migration model proposed by this method has a smaller data volume requirement and is more stable, which can further reduce costs and improve efficiency.

[0105] Figure 7 The following is a schematic diagram showing the statistical results of the model performance according to a specific example of the present invention. To avoid contingency during operation, all the following results are the average values ​​obtained after 5 independent runs of the network. It can be seen that the migration gated recurrent network of the present invention performs much better than the non-migration gated recurrent network. The left ordinate is used to represent the determination coefficient R 2The right ordinate is used to represent the mean absolute error MAE and the root mean square error RMSE, where the R 2 The R of the 200℃ migration network on the test set is 98.1%, MAE is 1.3MPa, and RMSE is 1.2MPa. 2 The R of the 300℃ migration network on the test set is 99.0%, MAE is 0.78MPa, and RMSE is 0.73MPa. 2 It is 95.2%, MAE is 1.72MPa, and RMSE is 2.63MPa.

[0106] Figure 8 It is a method schematic diagram of a specific embodiment of the high temperature SN curve prediction method provided by the present invention.

[0107] like Figure 8 As shown in the figure, data is obtained through fatigue tests at different temperatures, and the stress amplitude σ of the test data at different temperatures is interpolated to unify its data format and expand the data volume. The data obtained after interpolation is preprocessed, and then the processed stress amplitude σ is serialized according to the requirements of the gated cyclic network for the input data shape. The above data set is divided into training set, validation set and test set according to a certain ratio, where the training set is used to train the network, the validation set is used to adjust the hyperparameters, and the test set is used to evaluate the network prediction performance. Gated cyclic network models with different structures are established, and the network is preliminarily trained using the training set of the room temperature SN curve. At the same time, the performance of the network is verified using the validation set, the hyperparameters of the network are adjusted, and the model is retrained using the training set and validation set to obtain the best source model for transfer learning, and transfer learning is performed to obtain the SN curve prediction model at high temperature at different temperatures; the obtained model is used for cyclic prediction, and finally the complete high temperature fatigue SN curve at different temperatures is output.

[0108] The present invention also provides a storage medium corresponding to the high-temperature SN curve prediction method, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0109] The present invention also provides a computing device corresponding to the prediction method of the high-temperature SN curve, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the aforementioned methods when executing the computer program.

[0110] The present invention also provides a computer program product corresponding to the high-temperature SN curve prediction method, including a computer program, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0111] Based on this, the solution provided by the present invention uses only a small amount of sample data to train the gated recurrent network model based on the transfer learning method to obtain the SN curve prediction model at high temperature; it can realize the transfer of domain knowledge between different service temperatures, so as to make accurate predictions of high-temperature SN curves under extremely small sample conditions. By comparison, it is found that compared with the network without transfer learning, the method proposed by the present invention has a smaller sample size requirement and higher accuracy.

[0112] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored as one or more instructions or codes on a computer-readable medium or transmitted via a computer-readable medium. Other examples and implementations are within the scope and spirit of the present invention and the appended claims. For example, due to the nature of software, the functions described above may be implemented using software executed by a processor, hardware, firmware, hard wiring, or a combination of any of these. In addition, each functional unit may be integrated into a processing unit, each unit may exist physically separately, or two or more units may be integrated into one unit.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0114] The units described as separate components may or may not be physically separated, and the components of the control device may or may not be physical units, that is, they may be located in one place or distributed in multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0115] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the relevant technology or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0116] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A high temperature SN curve prediction method based on transfer learning, characterized in that: include: Obtain fatigue performance data at different temperatures obtained in advance through fatigue performance tests; The fatigue performance data include: stress amplitude and cycle number; the fatigue performance data at different temperatures include: fatigue performance data at room temperature and fatigue performance data at high temperature; Preset processing is performed on the fatigue performance data obtained at different temperatures to obtain sample data at different temperatures; the sample data at different temperatures include: sample data at room temperature and sample data at high temperature; The sample data at room temperature is used to perform preliminary training on a gated recurrent network model to obtain a transfer learning source model; the room temperature includes a temperature that can be reached indoors; the high temperature includes a temperature higher than room temperature; Establishing a migration model based on the migration learning source model, and training the migration model using the sample data at the high temperature to obtain a high temperature SN curve prediction model; The obtained high-temperature SN curve prediction model is used to perform cyclic prediction and output a high-temperature SN curve.

2. The method according to claim 1, characterized in that The fatigue performance test includes at least one of rotational bending fatigue, tension and compression fatigue and torsional fatigue.

3. The method according to claim 1, characterized in that The fatigue performance data obtained at different temperatures are processed by preset methods, including: The stress amplitude data in the fatigue performance data at different temperatures are interpolated respectively to unify the data format and expand the data volume; wherein the stress amplitude data obtained after the interpolation processing are distributed at equal intervals in the direction of the cycle number; The stress amplitude data obtained after interpolation processing is standardized and converted into dimensionless data; The stress amplitude data after interpolation and standardization are serialized, wherein the sequence length of the stress amplitude sequence is set according to the amount of data.

4. The method according to claim 3, characterized in that When performing interpolation processing, set the cycle range to: The cycle range at room temperature was set to 9×10 4 Up to 1×10 7 and / or, The cycle range at 100°C is set to 7×10 4 Up to 1×10 7 and / or, The cycle range at 200°C is set to 1×10 5 Up to 1×10 7 and / or, The cycle range at 300°C is set to 3×10 4 Up to 1×10 7 and / or, The data amount range of the stress amplitude obtained after interpolation processing is set to 800 to 1500; and / or, The sequence length of the stress amplitude sequence is set to 10 to 80 data.

5. The method according to claim 3, characterized in that: The interpolation processing includes: performing interpolation processing on the stress amplitude data using a linear interpolation method; and / or, The normalization process includes: using a Min-Max normalization method to map the range of the stress amplitude data to a range of 0 to 1.

6. The method according to claim 1, characterized in that Also includes: Dividing the sample data at different temperatures into a training set, a validation set, and a test set according to different preset ratios, respectively, to obtain a room temperature training set, a room temperature validation set, and a room temperature test set, as well as a high temperature training set, a high temperature validation set, and a high temperature test set; The sample data at room temperature is used to perform preliminary training and adjustment on the gated recurrent network model to obtain a transfer learning source model, including: Constructing gated recurrent network models with different structures, and setting different numbers of neurons for the gated recurrent network models with different structures; Using the room temperature training set to train the gated recurrent network models with different structures respectively; For the gated recurrent network models with different structures, determine the structure with the best performance by evaluating their performance on the room temperature validation set, and adjust their hyperparameters; Using the adjusted hyperparameters, reusing the room temperature training set and the room temperature validation set to perform model training on the structure with the best performance to obtain a transfer learning source model; The transfer learning source model is used to make predictions on the room temperature test set to evaluate the performance of the transfer learning source model.

7. The method according to claim 6, characterized in that The constructed gated recurrent network model includes a GRU layer, a fully connected layer and an output layer; Among them, the number of neurons in the GRU layer is 20; the fully connected layer is a two-layer structure, the number of neurons in each layer is 20, and the number of neurons in the output layer is 1.

8. The method according to any one of claims 1 to 7, characterized in that: Establishing a migration model based on the migration learning source model includes: Establishing a target gated recurrent network model with the same network structure as the transfer learning source model; Copying the GRU layer of the transfer learning source model to the GRU layer of the target gated recurrent network model to obtain a transfer layer, and freezing the transfer layer; The fully connected layer of the target gated recurrent network model is randomly initialized, and the fully connected layer is trained using the sample data at the high temperature to obtain a migration model.

9. The method according to any one of claims 1 to 7, characterized in that: Also includes: Dividing the sample data at different temperatures into a training set, a validation set, and a test set according to different preset ratios, respectively, to obtain a room temperature training set, a room temperature validation set, and a room temperature test set, as well as a high temperature training set, a high temperature validation set, and a high temperature test set; The obtained high temperature SN curve prediction model is used to perform cyclic prediction and output a high temperature SN curve, including: Inputting the last stress amplitude sequence of length L of the high temperature validation set into the prediction model, and outputting the first stress amplitude of the predicted high temperature test set; The window slides backward by a data length, the last L-1 stress amplitudes of the high temperature verification set and the predicted first stress amplitude of the high temperature test set are input into the prediction model, and the predicted second stress amplitude of the high temperature test set is output; By analogy, the prediction results of the prediction model for the high-temperature test set are output cyclically until the target cycle number is reached to obtain a high-temperature SN curve.

10. A storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps of any method described in claims 1-9 are implemented.

11. A computing device, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory and executable on the processor, wherein the steps of the method according to any one of claims 1 to 9 are implemented when the processor executes the program.

12. A computer program product, characterized in that The invention comprises a computer program, which implements the steps of any one of the methods of claims 1 to 9 when being executed by a processor.

Images