Method for predicting fatigue life of cylinder cover material heat engine based on self-supervised generative adversarial network
A self-supervised generative adversarial network addresses the inefficiencies of existing methods by using small sample data to predict cylinder head material fatigue life, enhancing prediction accuracy and efficiency.
Patent Information
- Application Number
- CN202411038989.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-12-03
AI Technical Summary
Existing methods for predicting the thermal mechanical fatigue life of cylinder head materials require large amounts of training data and suffer from low prediction efficiency and accuracy, especially under small sample conditions.
A self-supervised generative adversarial network is used to predict the thermal mechanical fatigue life of cylinder head materials, utilizing a limited number of thermomechanical fatigue tests to obtain representative small sample data, with a generator and discriminator built under the framework of a generative adversarial network to improve learning efficiency and accuracy.
The method provides accurate and efficient prediction of thermal mechanical fatigue life, reducing costs and increasing efficiency in cylinder head material research and development, with improved prediction accuracy under small sample conditions.
Smart Images

Figure CN119066950A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of internal combustion engine cylinder heads, and in particular relates to a method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network. Background Art
[0002] As one of the most complex structural components in the engine that bears the most severe mechanical and thermal loads, the load-bearing characteristics of the engine cylinder head material directly affect its fatigue strength reliability. With the increase in diesel engine power density and combustion pressure, the service conditions of the cylinder head are becoming increasingly severe, which greatly increases the probability of fatigue damage under high temperature and high pressure conditions, and also makes the fatigue failure problem more significant. Understanding the fatigue behavior and fatigue life of the cylinder head material under the action of thermal-mechanical coupling is of great significance to improving the durability and reliability of the cylinder head service.
[0003] At present, the more mature cylinder head materials are mainly cast iron materials and aluminum alloy materials. These two types of materials have the advantages of excellent casting performance and good comprehensive mechanical properties. Since the cylinder head fire surface is subjected to periodic thermal shock loads, the thermal mechanical fatigue behavior of the material under pulse loads and the related life prediction have become the focus of attention. Obtaining the temperature and stress-strain resistance of the cylinder head material over time is an important basis for material research and development and cylinder head component assessment and evaluation, and is also a necessary prerequisite for cylinder head fatigue life prediction. The cost of sampling the cylinder head material is high, and the thermomechanical fatigue test is time-consuming and complicated. Therefore, research on the use of limited small samples to establish an accurate life prediction model for cylinder head materials has gradually become the mainstream.
[0004] With the development of data-driven algorithms, the use of machine learning and deep learning methods to predict fatigue life under complex conditions has become a focus of attention. Currently, many researchers have used machine learning frameworks, regression models, etc. to predict the notch fatigue life of polycrystalline alloys and the fatigue life of aluminum alloys in the fields of aerospace and major equipment, as well as to identify key characteristic parameters such as mechanical properties and load characteristics. However, the commonly used machine learning algorithms are mainly supervised learning algorithms, which require a large amount of experimental data as support and have poor training effects under small sample conditions. Self-supervised learning algorithms do not require labeled data, can self-generate labels, and learn based on these labels, which improves the learning efficiency under small sample conditions. By using a self-supervised learning algorithm based on a generative adversarial network, the accuracy of the prediction model can be improved under small sample conditions. Summary of the invention
[0005] In view of this, the present invention aims to propose a method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network, so as to solve the problems of large amount of training data required, low prediction efficiency and low prediction accuracy in existing fatigue life prediction methods.
[0006] To achieve the above object, the technical solution of the present invention is achieved as follows:
[0007] A method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network comprises the following steps:
[0008] Step 1: Based on the periodic thermal shock load characteristics of the cylinder head fire surface, with mechanical and thermal loads as boundary inputs, the strain amplitude and temperature are tested under the same / opposite phase pulse cyclic loading conditions to determine the thermal mechanical fatigue behavior of the cylinder head material, so as to establish a life prediction relationship;
[0009] Based on the test results of the thermomechanical fatigue behavior of cylinder head materials, including sample data with temperature series and strain series under half-life cycles as input features and logarithmic fatigue life as output features;
[0010] Step 2: Normalize the sample data obtained in step 1;
[0011] The normalization formula is:
[0012] In the formula, x represents sample data, is the sample data variance, σ(x) is the sample data standard deviation, x nor is the normalized sample feature;
[0013] Step 3: Based on the gated neural network, a generator is constructed using the temperature sequence T(t) and strain sequence ε(t) under the half-life cycle normalized in step 2 to predict the fatigue life of the cylinder head material under different load conditions;
[0014] The generator loss function of the generative adversarial network is:
[0015]
[0016] In the formula, D represents the discriminator, G represents the generator, Represents the discriminator's expectation of the judgment result of the virtual sample;
[0017] Step 4: Based on the gated neural network, the temperature series T(t), strain series ε(t) and logarithmic fatigue life lg(N) under the half-life cycle normalized in step 2 are used. t ), build a discriminator to determine whether the input sample is a real test sample or a virtual sample;
[0018] The discriminator loss function of the generated adversarial network is defined as follows:
[0019]
[0020] In the formula, D represents the discriminator, G represents the generator, Represents the discriminator's expectation of the judgment result of the real sample, Represents the discriminator's expectation of the judgment result of the virtual sample;
[0021] Step 5: The training process of the generator and the discriminator is carried out alternately under the framework of the generative adversarial network model. The discriminator continuously judges the real samples and virtual samples and propagates the judgment results to the discriminator and the generator through error back propagation;
[0022] Step 6: The generator improves the generated results based on the error;
[0023] Step 7: The prediction accuracy is evaluated by the root mean square error;
[0024] The evaluation formula is:
[0025] In the formula, y i,pre is the logarithmic fatigue life predicted by the model, y i,exp Logarithmic fatigue life obtained from the experiment.
[0026] Furthermore, the implementation method of step 1 is:
[0027] Step 1.1: The strain amplitude range of the thermal mechanical fatigue test of the cylinder head material is 0.3%-0.8%, and the temperature cycle range is designed to be heated from 100°C to the test temperature value, and the strain and temperature are loaded in a triangular waveform in positive and negative phases respectively;
[0028] Step 1.2: Under each temperature cycle, control different strain amplitudes and conduct 4-6 groups of tests respectively.
[0029] Furthermore, the implementation method of step 3 is:
[0030] The generator is built under the gated neural network framework and includes two layers of gated recurrent unit layers. It receives the temperature sequence T(t) and the strain sequence ε(t) data, extracts features through the gated recurrent unit, and then obtains the logarithmic life prediction value lg(N t ).
[0031] Furthermore, the implementation method of step 4 is:
[0032] The discriminator is built under the framework of gated neural network and includes two layers of gated recurrent unit layers, which receives the temperature sequence T(t), the strain sequence ε(t) and the logarithmic life prediction value lg(N t ) as input, passes through the gated recurrent unit for feature extraction, and then passes through two layers of fully connected layers to obtain the judgment of real samples or virtual samples.
[0033] Furthermore, the specific steps of each training iteration of step 5 are:
[0034] Step 5.1: The generator generates corresponding life prediction values according to the temperature sequence T(t) and the strain sequence ε(t);
[0035] The sample and the corresponding life expectancy prediction value are defined as virtual samples; the sample and the corresponding life expectancy true value are defined as real samples;
[0036] Step 5.2: The discriminator uses the input temperature sequence T(t) and strain sequence ε(t) as well as the logarithmic life value lg(N t ), and obtain the judgment of real samples and virtual samples;
[0037] Step 5.3: Calculate the loss function of the generator based on the classification results; use the back-propagation algorithm to update the parameters of the generator;
[0038] Step 5.4: Calculate the loss function of the discriminator; use the back-propagation algorithm to update the parameters of the discriminator.
[0039] Furthermore, the implementation method of step 7 is:
[0040] If the network model meets the prediction accuracy requirements, save the trained network model; if it does not meet the accuracy requirements, return to step 3 to adjust the model parameters and retrain the model.
[0041] Compared with the prior art, the method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network described in the present invention has the following advantages:
[0042] (1) The present invention discloses a method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network. The present invention discloses a method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network. It proposes to use a limited number of cylinder head material thermomechanical fatigue tests to obtain representative small sample data. By building a generator and a discriminator under the framework of a generative adversarial network and using the loss function of the generative adversarial network to provide feedback on the discrimination error, the spontaneity of the network model is stimulated, and the shortcomings of the supervised learning algorithm that rely on quantity and are prone to problems such as low prediction accuracy and overfitting are compensated.
[0043] (2) The method for predicting the thermal mechanical fatigue life of cylinder head materials based on a self-supervised generative adversarial network disclosed in the present invention shows good prediction results under small sample conditions by using a generative adversarial neural network model trained with adversarial training. The present invention provides a practical and effective method for predicting the thermal mechanical fatigue life of cylinder head materials, which has a strong guiding value for cylinder head design and fatigue analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] The accompanying drawings constituting a part of the present invention are used to provide a further understanding 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 accompanying drawings:
[0045] Figure 1 A schematic diagram of the method flow of the embodiment of the present invention;
[0046] Figure 2 A schematic diagram of a detailed network architecture of a generative adversarial network model used in an embodiment of the present invention;
[0047] Figure 3 This is a schematic diagram of the dimensions of a uniaxial specimen of a cylinder head grey cast iron material according to an embodiment of the present invention;
[0048] Figure 4 It is a strain load schematic diagram of a thermomechanical fatigue positive phase / reverse phase test with a strain amplitude of 0.8% under a temperature cycle of 100° C. to 550° C. as described in an embodiment of the present invention;
[0049] Figure 5 A schematic diagram of the root mean square error (RMSE) prediction performance of the self-supervised generative adversarial network model in a test set and its comparison with different machine learning models according to an embodiment of the present invention;
[0050] Figure 6 A diagram showing the effect of predicting the thermal engine fatigue life of gray cast iron material of a cylinder head based on the self-supervised generative adversarial network model according to an embodiment of the present invention and a schematic diagram showing the comparison with different machine learning models;
[0051] Figure 7 The figure is a schematic diagram of the principle of the method described in the embodiment of the present invention. DETAILED DESCRIPTION
[0052] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.
[0053] like Figures 1 to 7 As shown, a method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network is proposed. First, the temperature series, strain series and logarithmic fatigue life data under the half-life cycle obtained from the thermal engine fatigue test of the cylinder head material are normalized; then a generative adversarial network is built, and a generator and a discriminator based on a gated neural network are constructed to perform generative adversarial training to optimize the problems of low prediction accuracy or overfitting caused by insufficient training samples under small sample conditions; finally, the trained generator is used to realize fast and efficient prediction of the thermal engine fatigue life of the cylinder head material, thereby reducing costs and increasing efficiency in the research and development of cylinder head materials.
[0054] The prediction method specifically includes the following steps:
[0055] Step 1: Based on the periodic thermal shock load characteristics of the cylinder head fire surface, with mechanical and thermal loads as boundary inputs, the strain amplitude and temperature are tested under the same / opposite phase pulse cyclic loading conditions to determine the thermal mechanical fatigue behavior of the cylinder head material, so as to establish a life prediction relationship;
[0056] The test results of the thermal mechanical fatigue behavior of the cylinder head material include sample data with temperature sequence and strain sequence under half-life cycle as input features and logarithmic fatigue life as output features;
[0057] Step 2: Normalize the sample data obtained in step 1;
[0058] The normalization formula is:
[0059] In the formula, x represents sample data, is the sample data variance, σ(x) is the sample data standard deviation, x nor is the normalized sample feature;
[0060] Step 3: Based on the gated neural network, a generator is constructed using the temperature sequence T(t) and strain sequence ε(t) under the half-life cycle normalized in step 2 to predict the fatigue life of the cylinder head material under different load conditions;
[0061] The generator loss function of the generative adversarial network is:
[0062]
[0063] In the formula, D represents the discriminator, G represents the generator, Represents the discriminator's expectation of the judgment result of the virtual sample;
[0064] Step 4: Based on the gated neural network, the temperature series T(t), strain series ε(t) and logarithmic fatigue life lg(N) under the half-life cycle normalized in step 2 are used. t ), build a discriminator to determine whether the input sample is a real test sample or a virtual sample;
[0065] The structure of the discriminator is two layers of gated recurrent unit layers connected to two layers of fully connected layers. The length of the gated recurrent unit layer is the same as the temperature sequence. The output layer has one neuron, which transforms the temperature sequence T(t) and the strain sequence ε(t) and the logarithmic life prediction value lg(N t ) to extract features and obtain the judgment of real samples or virtual samples;
[0066] The discriminator loss function of the generated adversarial network is defined as follows:
[0067]
[0068] In the formula, D represents the discriminator, G represents the generator, Represents the discriminator's expectation of the judgment result of the real sample, Represents the discriminator's expectation of the judgment result of the virtual sample;
[0069] Step 5: The training process of the generator and the discriminator is performed alternately under the framework of the generative adversarial network (GAN) model. The discriminator continuously judges the real samples and virtual samples and propagates the judgment results to the discriminator and the generator through error back propagation;
[0070] Step 6: The generator improves the generated results based on the error, with the goal of making it impossible for the discriminator to distinguish between real samples and virtual samples, thereby achieving accuracy optimization.
[0071] Step 7: The prediction accuracy is evaluated by the root mean square error (RMSE);
[0072] The evaluation formula is:
[0073] In the formula, y i,pre is the logarithmic fatigue life predicted by the model, y i,exp Logarithmic fatigue life obtained from the experiment.
[0074] Advantages of the present invention:
[0075] 1. The present invention discloses a method for predicting the thermal mechanical fatigue life of cylinder head materials based on a self-supervised generative adversarial network. It proposes to use a limited number of thermomechanical fatigue tests on cylinder head materials to obtain representative small sample data. By building a generator and a discriminator under the framework of a generative adversarial network and using the loss function of the generative adversarial network to provide feedback on the discrimination error, the spontaneity of the network model is stimulated, and the shortcomings of the supervised learning algorithm that rely on quantity and are prone to problems such as low prediction accuracy and overfitting are compensated.
[0076] 2. The invention discloses a method for predicting the thermal mechanical fatigue life of cylinder head materials based on a self-supervised generative adversarial network. The generative adversarial neural network model using adversarial training shows good prediction results under small sample conditions. The invention provides a practical and effective method for predicting the thermal mechanical fatigue life of cylinder head materials, which has a strong guiding value for cylinder head design and fatigue analysis.
[0077] In a preferred embodiment of the present invention, the method for implementing step 1 is:
[0078] Step 1.1: The strain amplitude range of the thermal mechanical fatigue test of the cylinder head material is generally 0.3%-0.8%, and the temperature cycle range is designed to be heated from 100°C to the test temperature value, and the strain and temperature are loaded in a triangular waveform in positive and negative phases respectively;
[0079] Step 1.2: Under each temperature cycle, 4-6 groups of tests were carried out with different strain amplitudes.
[0080] In a preferred embodiment of the present invention, the method for implementing step 3 is:
[0081] The generator is built under the gated neural network framework and includes two layers of gated recurrent unit layers. It receives the temperature sequence T(t) and the strain sequence ε(t) data, extracts features through the gated recurrent unit, and then obtains the logarithmic life prediction value lg(N t ).
[0082] In a preferred embodiment of the present invention, the method for implementing step 4 is:
[0083] The discriminator is built under the framework of gated neural network and includes two layers of gated recurrent unit layers, which receives the temperature sequence T(t), the strain sequence ε(t) and the logarithmic life prediction value lg(N t ) as input, passes through the gated recurrent unit for feature extraction, and then passes through two layers of fully connected layers to obtain the judgment of real samples or virtual samples.
[0084] In a preferred embodiment of the present invention, in step 5, the steps of each training iteration are specifically as follows:
[0085] Step 5.1: The generator generates corresponding life prediction values according to the temperature sequence T(t) and the strain sequence ε(t);
[0086] The sample and the corresponding life expectancy prediction value are defined as virtual samples; the sample and the corresponding life expectancy true value are defined as real samples;
[0087] Step 5.2: The discriminator uses the input temperature sequence T(t) and strain sequence ε(t) as well as the logarithmic life value lg(N t ), and obtain the judgment of real samples and virtual samples;
[0088] Step 5.3: Calculate the loss function of the generator based on the classification results; use the back-propagation algorithm to update the parameters of the generator to reduce the loss of the generator;
[0089] Step 5.4: Calculate the loss function of the discriminator; use the back-propagation algorithm to update the parameters of the discriminator to reduce the classification error rate of the discriminator for real samples and fake samples.
[0090] In a preferred embodiment of the present invention, the method for implementing step 7 is:
[0091] If the network model meets the prediction accuracy requirements, save the trained network model; if it does not meet the accuracy requirements, return to step 3 to adjust the model parameters and retrain the model.
[0092] Example 1
[0093] This embodiment discloses a method for predicting the thermal mechanical fatigue life of cylinder head materials based on a self-supervised generative adversarial network, which predicts the thermal mechanical fatigue life of cylinder head gray cast iron RuT400 material. The specific implementation steps are as follows:
[0094] Step 1: Thermomechanical fatigue test of gray cast iron material of cylinder head to obtain fatigue life under different temperature loads and strain loads; the specific steps are:
[0095] Step 1.1: Use wire cutting to remove the original gray cast iron material from the engine cylinder head fire surface, and then process it into Figure 3 The rod-shaped uniaxial fatigue specimen shown has its surface polished with sandpaper to ensure the surface roughness;
[0096] Step 1.2: Refer to GJB6213-2008 "Methods for Thermomechanical Fatigue Test of Metallic Materials" to conduct fatigue tests with positive and negative loading phases on the MTS-810 fatigue testing machine, and cyclic loading is performed in a pulsed manner to simulate the influence of cyclic burst pressure during engine operation; the temperature cycle range of the designed test is 100℃-450℃, 100℃-500℃ and 100℃-550℃, and 4, 6, and 5 groups of tests are carried out under each temperature cycle; taking the test with a mechanical strain amplitude of 0.8% under the temperature cycle of 100℃-550℃ as an example, the loading waveforms of positive and negative phases are shown respectively, as shown in the figure below: Figure 4 The specific test parameters and life results are shown in Table 1.
[0097] Table 1
[0098]
[0099] Step 2: Normalize the sample data obtained in step 1, including the temperature series, strain series and logarithmic fatigue life under the half-life cycle;
[0100] The normalization formula is:
[0101] In the formula, x represents sample data, is the sample data variance, σ(x) is the sample data standard deviation, x nor is the normalized sample feature;
[0102] Step 3: Based on the gated neural network, a generator is constructed using the temperature sequence T(t) and strain sequence ε(t) under the half-life cycle normalized in step 2 to predict the fatigue life of the cylinder head material under different load conditions;
[0103] The structure of the generator is two layers of gated recurrent unit layers connected to one fully connected layer. The length of the gated recurrent unit layer is the same as the temperature sequence. The output layer has one neuron, which extracts features of the temperature sequence T(t) and the strain sequence ε(t) to obtain the fatigue life prediction value.
[0104] The generator loss function of the generated adversarial network is used as follows:
[0105]
[0106] Step 4: Based on the gated neural network, the temperature series T(t), strain series ε(t) and logarithmic fatigue life lg(N) under the half-life cycle normalized in step 2 are used. t ), build a discriminator to determine whether the input sample is a real test sample or a virtual sample;
[0107] The structure of the discriminator is two layers of gated recurrent unit layers connected to two layers of fully connected layers. The length of the gated recurrent unit layer is the same as the temperature sequence. The output layer has one neuron, which transforms the temperature sequence T(t) and the strain sequence ε(t) and the logarithmic life prediction value lg(N t ) to extract features and obtain the judgment of real samples or virtual samples;
[0108] The discriminator loss function of the generated adversarial network is used as follows:
[0109]
[0110] Step 5: The training process of the generator and the discriminator is carried out under the framework of the generative adversarial network (GAN) model. The discriminator continuously judges the real samples and virtual samples and propagates the judgment results to the discriminator and the generator through error back propagation;
[0111] Step 5.1: The generator generates corresponding life prediction values according to the temperature sequence T(t) and the strain sequence ε(t). The sample and the corresponding life prediction value are defined as virtual samples. The sample and the corresponding life true value are defined as real samples;
[0112] Step 5.2: The discriminator uses the input temperature sequence T(t) and strain sequence ε(t) as well as the logarithmic life value lg(N t ), and obtain the judgment of real samples and virtual samples;
[0113] Step 5.3: Calculate the loss function of the generator based on the classification results, and use the back-propagation algorithm to update the parameters of the generator to reduce the loss of the generator;
[0114] Step 5.4: Calculate the loss function of the discriminator and use the back-propagation algorithm to update the parameters of the discriminator to reduce the classification error rate of the discriminator for real samples and fake samples;
[0115] Step 6: The generator predicts fatigue life based on the input value, so that the discriminator cannot distinguish between real samples and virtual samples, thus achieving accuracy optimization;
[0116] Step 7: The prediction accuracy is evaluated by the root mean square error (RMSE);
[0117] The evaluation formula is:
[0118] In the formula, y i,pre is the logarithmic fatigue life predicted by the model, y i,exp Logarithmic fatigue life obtained from the experiment.
[0119] like Figure 5 The box plot shown shows the life prediction accuracy of the generative adversarial network model in repeated experiments, where the two ends of the line segment represent the maximum and minimum values in the repeated experimental results, and the two ends of the box represent the 75% and 25% values in the repeated experimental results; this embodiment compares the prediction effects of the present invention with other 6 supervised learning neural network models (artificial neural network ANN, K nearest neighbor algorithm KNN, random forest RF, support vector machine SVM, long short-term memory network LSTM, gated neural network GRU); from Figure 5 It can be seen that the prediction results after training the thermomechanical fatigue data used in this embodiment using the supervised learning model have a large dispersion, while the prediction results based on the self-supervised generative adversarial network model (GAN) show good repeatability and low prediction error;
[0120] Figure 6 The specific prediction effect of each model is demonstrated. The prediction effects of supervised machine learning models ANN, LSTM, and GRU under the neural network framework are poor in both the training set and the test set; non-neural network machine learning models KNN, RF, and SVM show obvious overfitting behavior; in contrast, the GAN model of the present invention shows better training effect and generalization performance.
[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting the thermal mechanical fatigue life of cylinder head materials based on self-supervised generative adversarial networks, Features: The steps include: Step 1: Based on the periodic thermal shock load characteristics of the cylinder head fire surface, with mechanical and thermal loads as boundary inputs, the strain amplitude and temperature are tested under the same / opposite phase pulse cyclic loading conditions to determine the thermal mechanical fatigue behavior of the cylinder head material, so as to establish a life prediction relationship; Based on the test results of the thermomechanical fatigue behavior of cylinder head materials, including sample data with temperature series and strain series under half-life cycles as input features and logarithmic fatigue life as output features; Step 2: Normalize the sample data obtained in step 1; The normalization formula is: In the formula, x represents sample data, is the sample data variance, σ(x) is the sample data standard deviation, x nor is the normalized sample feature; Step 3: Based on the gated neural network, a generator is constructed using the temperature sequence T(t) and strain sequence ε(t) under the half-life cycle normalized in step 2 to predict the fatigue life of the cylinder head material under different load conditions; The generator loss function of the generative adversarial network is: In the formula, D represents the discriminator, G represents the generator, Represents the discriminator's expectation of the judgment result of the virtual sample; Step 4: Based on the gated neural network, the temperature series T(t), strain series ε(t) and logarithmic fatigue life lg(N) under the half-life cycle normalized in step 2 are used. t ), build a discriminator to determine whether the input sample is a real test sample or a virtual sample; The discriminator loss function of the generated adversarial network is defined as follows: In the formula, D represents the discriminator, G represents the generator, Represents the discriminator's expectation of the judgment result of the real sample, Represents the discriminator's expectation of the judgment result of the virtual sample; Step 5: The training process of the generator and the discriminator is carried out alternately under the framework of the generative adversarial network model. The discriminator continuously judges the real samples and virtual samples and propagates the judgment results to the discriminator and the generator through error back propagation; Step 6: The generator improves the generated results based on the error; Step 7: The prediction accuracy is evaluated by the root mean square error; The evaluation formula is: In the formula, y i,pre is the logarithmic fatigue life predicted by the model, y i,exp Logarithmic fatigue life obtained from the experiment.
2. According to claim 1, a method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network, Features: Step 1 is implemented as follows: Step 1.1: The strain amplitude range of the thermal mechanical fatigue test of the cylinder head material is 0.3%-0.8%, and the temperature cycle range is designed to be heated from 100°C to the test temperature value, and the strain and temperature are loaded in a triangular waveform in positive and negative phases respectively; Step 1.2: Under each temperature cycle, control different strain amplitudes and conduct 4-6 groups of tests respectively.
3. According to claim 1, a method for predicting the thermal mechanical fatigue life of cylinder head materials based on a self-supervised generative adversarial network, Features: Step 3 is implemented as follows: The generator is built under the gated neural network framework and includes two layers of gated recurrent unit layers. It receives the temperature sequence T(t) and the strain sequence ε(t) data, extracts features through the gated recurrent unit, and then obtains the logarithmic life prediction value lg(N t ).
4. According to claim 1, a method for predicting the thermal mechanical fatigue life of cylinder head materials based on a self-supervised generative adversarial network, Features: Step 4 is implemented as follows: The discriminator is built under the gated neural network framework and includes two layers of gated recurrent units. It receives the temperature sequence T(t), the strain sequence ε(t) and the logarithmic life prediction value lg(N t ) as input, passes through the gated recurrent unit for feature extraction, and then passes through two layers of fully connected layers to obtain the judgment of real samples or virtual samples.
5. According to claim 1, a method for predicting the thermal mechanical fatigue life of cylinder head materials based on a self-supervised generative adversarial network, Features: The specific steps of each training iteration of step 5 are: Step 5.1: The generator generates corresponding life prediction values according to the temperature sequence T(t) and the strain sequence ε(t); The sample and the corresponding life expectancy prediction value are defined as virtual samples; the sample and the corresponding life expectancy true value are defined as real samples; Step 5.2: The discriminator uses the input temperature sequence T(t) and strain sequence ε(t) as well as the logarithmic life value lg(N t ), and obtain the judgment of real samples and virtual samples; Step 5.3: Calculate the loss function of the generator based on the classification results; use the back-propagation algorithm to update the parameters of the generator; Step 5.4: Calculate the loss function of the discriminator; use the back-propagation algorithm to update the parameters of the discriminator.
6. A method for predicting the thermal engine fatigue life of cylinder head materials based on a self-supervised generative adversarial network according to claim 1, Features: Step 7 is implemented as follows: If the network model meets the prediction accuracy requirements, save the trained network model; if it does not meet the accuracy requirements, return to step 3 to adjust the model parameters and retrain the model.
