A fatigue life prediction method based on a generative adversarial network and small sample data
By generating data similar to fatigue test data using generative adversarial networks and fitting the life distribution using the least squares method, the problem of insufficient data samples in the plotting of fatigue life curves of mechanical parts is solved, and effective enhancement and accurate plotting of small sample data are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2022-10-13
- Publication Date
- 2026-04-28
AI Technical Summary
In the existing technology, the drawing of fatigue life curves of mechanical parts relies on a large amount of fatigue test data, which leads to long test cycles and high costs, and makes it difficult to effectively draw accurate curves using small sample data.
A fatigue data augmentation model is established using generative adversarial networks. The generator and discriminator generate data similar to fatigue test data, and the least squares method is used to fit the life distribution to realize the plotting of fatigue life curves for small sample data.
It effectively enhances fatigue data of small-sample mechanical parts, enables accurate plotting of fatigue life curves of mechanical parts, and solves the problem of insufficient data samples.
Smart Images
Figure CN115630450B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fatigue life prediction, specifically a fatigue life prediction method based on generative adversarial networks and small sample data. Background Technology
[0002] Fatigue life curves of mechanical components are crucial foundational data supporting the high-reliability, long-life design of advanced equipment, and their accurate plotting relies on a large amount of fatigue test data. However, fatigue testing is characterized by long testing cycles and high costs, and fatigue test data for mechanical components is typically small-sample data. Therefore, it is essential to develop curves that accurately reflect the fatigue life relationship of gears based on small-sample fatigue data of mechanical components. Summary of the Invention
[0003] The purpose of this invention is to provide a fatigue life prediction method based on generative adversarial networks and small sample data, comprising the following steps:
[0004] 1) Obtain fatigue test data for mechanical parts;
[0005] Furthermore, the fatigue test data of the mechanical components are obtained by monitoring stress tests at test points of the mechanical components under different reliability conditions.
[0006] Furthermore, the fatigue test data of the mechanical parts is small sample data, and the number of test points of the monitored mechanical parts is less than or equal to 5.
[0007] Furthermore, the fatigue test data for the mechanical components includes stress-life data under different reliability levels, and the relationship between the two is as follows:
[0008] σ m ·N L,R =C (1)
[0009] In the formula, σ is the stress, and N is the stress. L,R Let R be the fatigue life under reliability R, and m and C be material parameters.
[0010] 2) Establish a fatigue data augmentation model for mechanical parts based on generative adversarial networks;
[0011] 3) Establish a fatigue generation dataset using fatigue data augmentation models for mechanical parts;
[0012] Furthermore, the mechanical component fatigue data augmentation model based on generative adversarial networks includes a generator and a discriminator;
[0013] Both the generator and discriminator are fully connected neural network models. The LeakyReLU function is used for parameter passing between layers, and the RMSprop optimization algorithm is used to update the fatigue data of mechanical parts to enhance the model weights.
[0014] Furthermore, the steps for establishing a fatigue generation dataset using a fatigue data augmentation model for mechanical parts include:
[0015] 3.1) Input the Gaussian random noise data z into the generator, and the generator produces fatigue generation data;
[0016] 3.2) Input the fatigue generation data and mechanical component fatigue test data into the discriminator. The discriminator calculates the similarity between the fatigue generation data and the mechanical component fatigue test data based on the Wasserstein distance algorithm. If the similarity is greater than the threshold, the fatigue generation dataset is output; otherwise, the generator parameters are updated, and the process returns to step 3.1).
[0017] 4) Clean the fatigue generation dataset to obtain optimized fatigue generation data, and mix the optimized fatigue generation data with the fatigue test data of mechanical parts to obtain the fatigue curve plotting dataset;
[0018] Furthermore, the steps for cleaning the fatigue-generated dataset include: retaining generated fatigue data with stress fluctuations within ±0.2% of the test stress level of small sample data of mechanical parts.
[0019] 5) Use the least squares method to fit the life distribution and perform goodness-of-failure tests on the fatigue curve plotting dataset to obtain the relationship between reliability and life under constant stress level.
[0020] Furthermore, the relationship between reliability and lifespan at the constant stress level is as follows:
[0021]
[0022] In the formula, P(N) L ) represents the empirical distribution function; μ ln N It is the mean of the log-normal distribution; σ ln N Φ is the variance of the log-normal distribution; -1 (P(N L )) represents lifetime distribution; N L This refers to fatigue life.
[0023] Wherein, the empirical distribution function P(N) L As shown below:
[0024]
[0025] In the formula, i F The number of data sequences, n, is used to plot the overall fatigue curve for a given experimental stress level. F The amount of data used to plot the total fatigue curve for a given test stress level.
[0026] 6) Based on the relationship between reliability and life under a constant stress level, plot the fatigue life curve of mechanical parts;
[0027] 7) Monitor the current stress of mechanical parts and determine the fatigue life prediction result corresponding to the current stress of mechanical parts based on the fatigue life curve of mechanical parts.
[0028] The technical effects of this invention are undeniable. By generating data similar to fatigue test data, this invention solves the problem of insufficient data samples in the process of drawing fatigue curves for mechanical parts, thereby realizing the enhancement of fatigue data and the drawing of fatigue life curves for small sample mechanical parts. Attached Figure Description
[0029] Figure 1 A fatigue data augmentation model for mechanical parts based on generative adversarial networks;
[0030] Figure 2 Generate a dataset for gear bending fatigue;
[0031] Figure 3 The fatigue life curves of mechanical parts under different reliability levels are shown. Detailed Implementation
[0032] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.
[0033] Example 1:
[0034] See Figures 1 to 3 A fatigue life prediction method based on generative adversarial networks and small sample data includes the following steps:
[0035] 1) Obtain fatigue test data for mechanical parts;
[0036] The fatigue test data of the mechanical components were obtained by monitoring stress tests at test points of the mechanical components under different reliability conditions.
[0037] The fatigue test data of the mechanical parts is small sample data, and the number of test points of the monitored mechanical parts is less than or equal to 5.
[0038] The fatigue test data for the mechanical components includes stress-life data under different reliability levels, and the relationship between the two is as follows:
[0039] σ m ·N L，R =C (1)
[0040] In the formula, σ is the stress, and N is the stress. L，R Let R be the fatigue life under a reliability level of R, where R is the reliability level and m and C are material parameters.
[0041] 2) Establish a fatigue data augmentation model for mechanical parts based on generative adversarial networks;
[0042] 3) Establish a fatigue generation dataset using fatigue data augmentation models for mechanical parts;
[0043] The mechanical component fatigue data augmentation model based on generative adversarial networks includes a generator and a discriminator;
[0044] Both the generator and discriminator are fully connected neural network models. The LeakyReLU function is used for parameter passing between layers, and the RMSprop optimization algorithm is used to update the fatigue data of mechanical parts to enhance the model weights.
[0045] The steps for building a fatigue generation dataset using fatigue data augmentation models of mechanical parts include:
[0046] 3.1) Input the Gaussian random noise data z into the generator, and the generator produces fatigue generation data;
[0047] 3.2) Input the fatigue generation data and mechanical component fatigue test data into the discriminator. The discriminator calculates the similarity between the fatigue generation data and the mechanical component fatigue test data based on the Wasserstein distance algorithm. If the similarity is greater than the threshold, the fatigue generation dataset is output; otherwise, the generator parameters are updated, and the process returns to step 3.1).
[0048] 4) Clean the fatigue generation dataset to obtain optimized fatigue generation data, and mix the optimized fatigue generation data with the fatigue test data of mechanical parts to obtain the fatigue curve plotting dataset;
[0049] The steps for cleaning the fatigue-generated dataset include: retaining small sample data of mechanical parts and generating fatigue data with stress fluctuations within ±0.2% of the test stress level.
[0050] 5) Use the least squares method to fit the life distribution and perform goodness-of-failure tests on the fatigue curve plotting dataset to obtain the relationship between reliability and life under constant stress level.
[0051] The relationship between reliability and lifespan at the specified stress level is shown below:
[0052]
[0053] In the formula, P(N) L ) represents the empirical distribution function; μ ln N It is the mean of the log-normal distribution; σln N Φ is the variance of the log-normal distribution; -1 (P(N L )) represents lifetime distribution; N L This refers to fatigue life.
[0054] Wherein, the empirical distribution function P(N) L As shown below:
[0055]
[0056] In the formula, i F The number of data sequences, n, is used to plot the overall fatigue curve for a given experimental stress level. F The amount of data used to plot the total fatigue curve for a given test stress level.
[0057] 6) Based on the relationship between reliability and life under a constant stress level, plot the fatigue life curve of mechanical parts;
[0058] 7) Monitor the current stress of mechanical parts and determine the fatigue life prediction result corresponding to the current stress of mechanical parts based on the fatigue life curve of mechanical parts.
[0059] Example 1:
[0060] A fatigue life prediction method based on generative adversarial networks and small sample data includes the following:
[0061] 1) Obtain gear bending fatigue test data, as shown in Table 1. The stress-life curve relationship under different reliability levels can be represented by the Basquin model:
[0062] σ m ·N L，R =C (1)
[0063] In the formula, σ is the stress, and N is the stress. L R represents fatigue life, R represents reliability, and m and C represent material parameters.
[0064] Table 1 Gear Bending Fatigue Life Data
[0065] Gear bending stress / MPa Logarithmic bending fatigue life 676.27 6.47712 676.27 5.71158 676.27 5.15246 750.75 5.90008 750.75 4.8023 750.75 5.33053 825.64 5.07593 825.64 4.72722 825.64 4.36696 951.35 4.70944 951.35 4.37585 951.35 4.00996
[0066] 2) Establish a fatigue data augmentation model for mechanical parts based on generative adversarial networks:
[0067] Using Gaussian random noise data z as input to the generator, the generator produces generated data. This generated data is then combined with small sample data from fatigue tests of mechanical parts and fed into a discriminator to determine their similarity. When the similarity is high, the final generated fatigue data is formed. Otherwise, the generator model parameters are updated.
[0068] Both the generator and discriminator are fully connected neural network models. The LeakyReLU function is used for parameter passing between layers. The fatigue test data has a dimension of 2, and the RMSprop optimization algorithm is used to update the network weights.
[0069] 3) Create a generated dataset: After each training session, take 3000 generated fatigue data points to create a fatigue generation dataset, such as... Figure 2 As shown.
[0070] 4) Dataset generation, cleaning, and mixing: Take small sample data of mechanical parts and test stress levels within ±1MPa to generate fatigue data, mix it with the small sample test data, and establish a fatigue curve plotting dataset;
[0071] 5) Plot the fatigue curves of the small sample data under the test stress levels in ascending order to determine the life distribution at each stress level. Use the least squares method to fit and test the goodness of fit of the life distribution. The fitting formula can be expressed as:
[0072]
[0073] In the formula, P(N) L ) is the empirical distribution function, μ ln N It is the mean of a log-normal distribution, where P(N) L The median rank formula is approximately expressed as:
[0074]
[0075] In the formula, i F The number of data sequences, n, is used to plot the overall fatigue curve for a given experimental stress level. F The amount of data required to plot the total fatigue curve for a given test stress level;
[0076] 6) Fatigue life curve plotting: Based on equation (2), the relationship between reliability and life at a given stress level is obtained. Furthermore, the fatigue life curves of mechanical components at different reliability levels can be plotted using the least squares method, such as... Figure 3 As shown.
[0077] Example 3:
[0078] A fatigue life prediction method based on generative adversarial networks and small sample data includes the following steps:
[0079] 1) Obtain small sample data from fatigue tests on mechanical parts;
[0080] 2) Establish a fatigue data augmentation model for mechanical parts based on generative adversarial networks;
[0081] 3) Establish a fatigue generation dataset using fatigue data augmentation models for mechanical parts;
[0082] 4) Clean the fatigue generation dataset to obtain optimized fatigue generation data, and mix the optimized fatigue generation data with the fatigue test data of mechanical parts to obtain the fatigue curve plotting dataset;
[0083] 5) Use the least squares method to fit the life distribution and perform goodness-of-failure tests on the fatigue curve plotting dataset to obtain the relationship between reliability and life under constant stress level.
[0084] 6) Based on the relationship between reliability and life under a constant stress level, plot the fatigue life curve of mechanical parts;
[0085] 7) Monitor the current stress of mechanical parts and determine the fatigue life prediction result corresponding to the current stress of mechanical parts based on the fatigue life curve of mechanical parts.
[0086] Example 4:
[0087] A fatigue life prediction method based on generative adversarial networks and small sample data is described in Example 3. The fatigue test data of the mechanical parts is obtained by monitoring the stress at test points of the mechanical parts under different reliability conditions.
[0088] Example 5:
[0089] A fatigue life prediction method based on generative adversarial networks and small sample data is described in Example 3. The fatigue test data of the mechanical parts is small sample data, and the number of test points of the monitored mechanical parts is less than or equal to 5.
[0090] Example 6:
[0091] A fatigue life prediction method based on generative adversarial networks and small sample data is described in Example 3. The fatigue test data of the mechanical parts includes stress-life data under different reliability levels, and the relationship between the two is as follows:
[0092] σ m ·N L,R =C (1)
[0093] In the formula, σ is the stress, and N is the stress. L,R Let R be the fatigue life under reliability R, and m and C be material parameters.
[0094] Example 7:
[0095] A fatigue life prediction method based on generative adversarial networks and small sample data is described in Example 3. The fatigue data enhancement model for mechanical parts based on generative adversarial networks includes a generator and a discriminator.
[0096] Both the generator and discriminator are fully connected neural network models. The LeakyReLU function is used for parameter passing between layers, and the RMSprop optimization algorithm is used to update the fatigue data of mechanical parts to enhance the model weights.
[0097] Example 8:
[0098] A fatigue life prediction method based on generative adversarial networks and small sample data, the main contents of which are described in Example 3, includes the following steps for establishing a fatigue generation dataset using a mechanical parts fatigue data augmentation model:
[0099] 1) Input the Gaussian random noise data z into the generator, and the generator will generate fatigue generation data;
[0100] 2) Input the fatigue generation data and mechanical component fatigue test data into the discriminator. The discriminator calculates the similarity between the fatigue generation data and the mechanical component fatigue test data based on the Wasserstein distance algorithm. If the similarity is greater than the threshold, the fatigue generation dataset is output; otherwise, the generator parameters are updated, and the process returns to step 1).
[0101] Example 9:
[0102] A fatigue life prediction method based on generative adversarial networks and small sample data is described in Example 3. The steps for cleaning the fatigue generation dataset include: retaining the generated fatigue data with stress fluctuations within ±0.2% of the test stress level of the mechanical parts in the small sample data test.
[0103] Example 10:
[0104] A fatigue life prediction method based on generative adversarial networks and small sample data is described in Example 3. The relationship between reliability and life at a constant stress level is as follows:
[0105]
[0106] In the formula, P(N) L ) represents the empirical distribution function; μ ln N It is the mean of the log-normal distribution; σ ln N Φ is the variance of the log-normal distribution; -1 (P(N L )) represents lifetime distribution; N L This refers to fatigue life.
[0107] Wherein, the empirical distribution function P(N) L As shown below:
[0108]
[0109] In the formula, i FThe number of data sequences, n, is used to plot the overall fatigue curve for a given experimental stress level. F The amount of data used to plot the total fatigue curve for a given test stress level.
Claims
1. A fatigue life prediction method based on generative adversarial networks and small sample data, characterized in that, Includes the following steps: Step 1) Obtain small sample data from fatigue tests on mechanical parts; Step 2) Establish a fatigue data augmentation model for mechanical parts based on generative adversarial networks; Step 3) Establish a fatigue generation dataset using a fatigue data augmentation model for mechanical parts; Step 4) Clean the fatigue generation dataset to obtain optimized fatigue generation data, and mix the optimized fatigue generation data with the fatigue test data of mechanical parts to obtain the fatigue curve plotting dataset; Step 5) Use the least squares method to fit the life distribution and perform goodness-of-failure test on the fatigue curve plotting dataset to obtain the relationship between reliability and life under a constant stress level. Step 6) Based on the relationship between reliability and life under a constant stress level, plot the fatigue life curve of the mechanical parts; Step 7) Monitor the current stress of the mechanical parts and determine the fatigue life prediction result corresponding to the current stress of the mechanical parts based on the fatigue life curve of the mechanical parts; The fatigue test data for the mechanical components includes stress-life data under different reliability levels, and the relationship between the two is as follows: ；（1） In the formula, For stress, The fatigue life is given by reliability R, where m and C are material parameters. The relationship between reliability and lifespan at the specified stress level is shown below: ；（2） In the formula, It is the empirical distribution function; It is the mean of the log-normal distribution; The variance of the log-normal distribution; Indicates lifetime distribution; For fatigue life; Wherein, the empirical distribution function As shown below: ；（3） In the formula, To plot the data sequence number for the total fatigue curve under a given test stress level, The amount of data used to plot the total fatigue curve for a given test stress level.
2. The fatigue life prediction method based on generative adversarial networks and small sample data according to claim 1, characterized in that: The fatigue test data of the mechanical components were obtained by monitoring stress tests at test points of the mechanical components under different reliability conditions.
3. The fatigue life prediction method based on generative adversarial networks and small sample data according to claim 2, characterized in that: The fatigue test data of the mechanical parts is small sample data, and the number of test points of the monitored mechanical parts is less than or equal to 5.
4. The fatigue life prediction method based on generative adversarial networks and small sample data according to claim 1, characterized in that: The mechanical component fatigue data augmentation model based on generative adversarial networks includes a generator and a discriminator; Both the generator and discriminator are fully connected neural network models. The LeakyReLU function is used for parameter passing between layers, and the RMSprop optimization algorithm is used to update the fatigue data of mechanical parts to enhance the model weights.
5. The fatigue life prediction method based on generative adversarial networks and small sample data according to claim 4, characterized in that, The steps for building a fatigue generation dataset using fatigue data augmentation models of mechanical parts include: Step 1) Input the Gaussian random noise data z into the generator, and the generator produces fatigue generation data; Step 2) Input the fatigue generation data and mechanical component fatigue test data into the discriminator. The discriminator calculates the similarity between the fatigue generation data and the mechanical component fatigue test data based on the Wasserstein distance algorithm. If the similarity is greater than the threshold, the fatigue generation dataset is output; otherwise, the generator parameters are updated, and the process returns to step 1).
6. The fatigue life prediction method based on generative adversarial networks and small sample data according to claim 5, characterized in that, The steps for cleaning the fatigue-generated dataset include: retaining small sample data of mechanical parts and generating fatigue data with stress fluctuations within ±0.2% of the test stress level.