A method for predicting the fatigue pitting life of gears by integrating fatigue criteria and data-driven
By integrating fatigue criteria and data-driven methods, using literature meta-analysis and GAN models to generate data, and combining graph convolutional networks (GCN) to predict gear fatigue pitting life, the problem of insufficient small sample data is solved, and efficient and accurate gear life prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2024-12-26
- Publication Date
- 2026-08-04
AI Technical Summary
Existing technologies lack accurate small sample data for predicting gear fatigue pitting life, resulting in insufficient prediction accuracy, and the cost of acquiring high-quality data is high.
We adopted an integrated fatigue criterion and data-driven approach. We constructed the original dataset using literature meta-analysis and generated new data by combining the multi-axis fatigue criterion and the GAN adversarial network model to expand the training set. We then used the GCN model for prediction.
Under small sample data conditions, accurate prediction of gear fatigue pitting life was achieved, reducing data acquisition costs and improving prediction accuracy and reliability.
Smart Images
Figure CN120257768B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of gear fatigue life prediction, and in particular to a method for predicting gear fatigue pitting life that integrates fatigue criteria and data-driven approaches. Background Technology
[0002] Gear transmission systems are widely used in various industrial fields such as aerospace, automotive, and wind power generation. With the continuous increase in requirements for power density, load capacity, and reliability of equipment, gear failure problems are becoming increasingly prominent. Studies have shown that contact fatigue is one of the main modes of gear failure. Gears are prone to fatigue pitting under high loads and long-term operation, which in turn affects the performance and safety of equipment. Therefore, accurately predicting the fatigue pitting life of gears is crucial for ensuring the normal operation of mechanical systems.
[0003] Developing accurate and reliable fatigue pitting life prediction models for gears has become a current research hotspot. Effective prediction methods can identify potential failure risks in advance, providing a scientific basis for equipment maintenance and management. This not only helps extend gear life and improve system reliability but also reduces maintenance costs and downtime, optimizing resource allocation.
[0004] As the core of artificial intelligence, machine learning's main function is to simulate human learning behavior, continuously training models by acquiring new information to improve their generalization ability. Thanks to its powerful data processing capabilities, machine learning methods have been widely applied in fields such as data mining, speech recognition, computer vision, fault diagnosis, and life prediction. With the advent of Industry 4.0, deep learning, as an important branch of machine learning, has developed rapidly, opening up new directions for this field. In recent years, many machine learning methods have been applied to life prediction, including artificial neural networks (ANN), random forests (RF), support vector machines (SVM), Gaussian processes (GPS), and deep learning.
[0005] While machine learning methods have shown great potential in fatigue life prediction, their accuracy often depends on a large amount of high-quality fatigue data. However, obtaining a large amount of accurate experimental data is usually very time-consuming and expensive, as gear contact fatigue testing is both time-consuming and costly.
[0006] Therefore, it is of great significance to establish an accurate gear fatigue pitting life prediction model with a small sample set. Summary of the Invention
[0007] The purpose of this invention is to provide a gear fatigue criterion that integrates fatigue criteria and data-driven methods.
[0008] A method for predicting pitting lifetime is proposed to address the problems existing in current technologies.
[0009] The technical solution adopted to achieve the purpose of this invention is as follows: a method for predicting gear fatigue pitting life by integrating fatigue criteria and data-driven methods, comprising the following steps:
[0010] 1) Using the meta-analysis method, 100 sets of grouped data were obtained from published literature to construct the original dataset. The data characteristics of the grouped data include gear contact stress, elastic modulus, Poisson's ratio, contact fatigue limit, lubricating oil kinematic viscosity, rotational speed, pitch circle radius, pressure angle, helix angle, tooth width, contact width, surface roughness, surface hardness, tooth core hardness, and fatigue pitting life. The original dataset was divided into the original training set and the validation set in an 8:2 ratio.
[0011] 2) The original training set was expanded using 50 sets of data obtained from the multi-axis fatigue criterion.
[0012] 3) Train the adversarial network GAN model using the expanded training set; generate 500 new sets of data using the trained adversarial network GAN model; mix the new data with the expanded training set to obtain the enhanced training set.
[0013] 4) Train the fatigue pitting life prediction model using an enhanced training set until the model converges. The fatigue pitting life prediction model is a graph convolutional network (GCN) model. The input features of the fatigue pitting life prediction model are the gear's contact stress, elastic modulus, Poisson's ratio, contact fatigue limit, lubricating oil kinematic viscosity, rotational speed, pitch circle radius, pressure angle, helix angle, tooth width, contact width, surface roughness, surface hardness, and core hardness. The output feature is the fatigue pitting life.
[0014] 5) The fatigue pitting life prediction results of the gear pair under test are obtained by using the trained fatigue pitting life prediction model.
[0015] Furthermore, in step 1), surface hardness is used to evaluate the material fatigue characteristic parameter σ. f ′ and ε f This yields two new data features.
[0016] Furthermore, step 2) specifically includes the following sub-steps:
[0017] 2-1) Based on the gear geometry parameters in the original dataset, draw the model of the gear pair using drawing software and import it into the ABAQUS simulation software.
[0018] 2-2) Complete the pre-processing settings for the gear pair model using ABAQUS simulation software. These settings include defining material properties, establishing assembly and contact relationships, creating analysis steps and required output variables, mesh generation, and setting loads and constraints.
[0019] 2-3) Submit the job to view the contact torque history and maximum contact stress of the gear tooth surface to verify the accuracy of the model.
[0020] 2-4) Extract the stress and strain of all nodes from the pitch line of the intermediate gear pair to the tooth root.
[0021] 2-5) Use the FS method to calculate the fatigue life of each node.
[0022] 2-6) Use cumulative frequency distribution maps to predict the fatigue pitting life of gears.
[0023] 2-7) The FS method was validated using experimental data from the literature.
[0024] 2-8) Change the load within the range of experimental contact stress, use the FS method to calculate the lifetime and obtain new data to obtain an expanded training set.
[0025] Step 3) specifically includes the following sub-steps:
[0026] 3-1) Import the expanded training set data and perform preprocessing.
[0027] 3-2) Create the generator and discriminator.
[0028] 3-3) Build and train the GAN model.
[0029] 3-4) Use a generator to generate new data.
[0030] 3-5) The new training set is mixed with the expanded training set to obtain the enhanced training set.
[0031] Furthermore, in step 2-2), a set needs to be created for the intermediate gear pair and the gear pair contact area, and the output variables are selected as contact pressure, stress, and strain. In step 2-2), the contact area is finely meshed, with the mesh size ratio in the tooth diameter and tooth width directions being approximately 1:6, and the remaining areas are globally seeded.
[0032] Furthermore, in steps 2-6), a 3x error band (10) is used. 0.5 The cumulative frequency distribution map is obtained by fitting the fatigue life of all nodes at intervals, and the average value of all nodes in the interval with the highest probability is selected as the fatigue pitting life of the gear.
[0033] Furthermore, after steps 2-8), there is a step of inputting the data obtained with altered loads and the original training set into typical machine learning methods for training, and observing the model's predictive performance. After steps 3-4), there is a step of inputting the preprocessed GAN-generated data and the training set into typical machine learning methods for training, and observing the model's predictive performance. Typical machine learning methods include Gaussian Process (GPS), Support Vector Machine (SVM), Convolutional Network (CNN), and Artificial Neural Network (ANN).
[0034] Furthermore, step 4) specifically includes the following sub-steps:
[0035] 4-1) Create a graph convolutional layer.
[0036] 4-2) Create an adjacency matrix.
[0037] 4-3) Define the hyperparameters of the model.
[0038] 4-4) Train the GCN model.
[0039] 4-5) Predict fatigue life and compare the mean absolute error of the validation set.
[0040] Furthermore, the generation of the adjacency matrix in step 4-2) is selected according to different graph types.
[0041] Furthermore, in step 4-3), hyperparameters are defined before training the model. Hyperparameters include the number of nodes, graph type, number of units, Dropout rate, batch size, and number of training epochs.
[0042] Further, after steps 4-5), determine whether the model's prediction performance meets expectations. If not, return to step 4-3) to adjust the model structure or hyperparameters.
[0043] The technical effect of this invention is undeniable: it can expand the data by combining multi-axis fatigue criteria and generative models when gear fatigue pitting data is scarce, and establish an accurate small sample set gear fatigue pitting life prediction model, thereby providing support for failure analysis of gear transmission systems. Attached Figure Description
[0044] Figure 1 This is the overall flowchart of the method;
[0045] Figure 2 A comparison chart of predicted and experimental data using the FS method;
[0046] Figure 3 A comparison chart of the errors of each model before and after using the FS method to expand the original training set;
[0047] Figure 4 Here is a diagram of the GAN structure;
[0048] Figure 5 A comparison chart of the errors of each model before and after using GAN to expand the original training set;
[0049] Figure 6 GCN model diagram;
[0050] Figure 7 A comparison chart of the prediction results of each model;
[0051] Figure 8 A comparison chart of the prediction results of each model;
[0052] Figure 9 This is a comparison chart of the predicted data and experimental data from the GCN-FS-GAN model. Detailed Implementation
[0053] 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.
[0054] Example 1:
[0055] This embodiment addresses the issue of insufficient data on gear fatigue pitting by providing a method for predicting gear fatigue pitting life that integrates fatigue criteria and data-driven approaches, comprising the following steps:
[0056] 1) A meta-analysis method was used to obtain grouped data from published literature to construct the original dataset. The data characteristics of the grouped data included gear contact stress, elastic modulus, Poisson's ratio, contact fatigue limit, lubricating oil kinematic viscosity, rotational speed, pitch circle radius, pressure angle, helix angle, tooth width, contact width, surface roughness, surface hardness, core hardness, and fatigue pitting life. The original dataset was divided into an original training set and a validation set.
[0057] 2) The original training set is augmented using the multi-axis fatigue criterion to obtain the augmented training set. Step 2) specifically includes the following sub-steps:
[0058] 2-1) Based on the gear geometry parameters in the original dataset, draw the model of the gear pair using drawing software and import it into the ABAQUS simulation software.
[0059] 2-2) Complete the pre-processing settings for the gear pair model using ABAQUS simulation software. These settings include defining material properties, establishing assembly and contact relationships, creating analysis steps and required output variables, mesh generation, and setting loads and constraints.
[0060] 2-3) Submit the job to view the contact torque history and maximum contact stress of the gear tooth surface to verify the accuracy of the model.
[0061] 2-4) Extract the stress and strain of all nodes from the pitch line of the intermediate gear pair to the tooth root.
[0062] 2-5) Use the FS method to calculate the fatigue life of each node.
[0063] 2-6) Use cumulative frequency distribution maps to predict the fatigue pitting life of gears.
[0064] 2-7) The FS method was validated using experimental data from the literature.
[0065] 2-8) Change the load within the range of experimental contact stress, use the FS method to calculate the lifetime and obtain new data to obtain an expanded training set.
[0066] 3) Train the GAN (Generative Adversarial Network) model using the augmented training set. Generate new data using the trained GAN model. Mix the new data with the augmented training set to obtain the enhanced training set. Step 3) specifically includes the following sub-steps:
[0067] 3-1) Import the expanded training set data and perform preprocessing.
[0068] 3-2) Create the generator and discriminator.
[0069] 3-3) Build and train the GAN model.
[0070] 3-4) Use a generator to generate new data.
[0071] 3-5) The new training set is mixed with the expanded training set to obtain the enhanced training set.
[0072] 4) Train the fatigue pitting life prediction model using an enhanced training set until the model converges. The fatigue pitting life prediction model is a graph convolutional network (GCN) model. The input features of the fatigue pitting life prediction model are the gear's contact stress, elastic modulus, Poisson's ratio, contact fatigue limit, lubricating oil kinematic viscosity, rotational speed, pitch circle radius, pressure angle, helix angle, tooth width, contact width, surface roughness, surface hardness, and core hardness. The output feature is the fatigue pitting life. Step 4) specifically includes the following sub-steps:
[0073] 4-1) Create a graph convolutional layer.
[0074] 4-2) Create an adjacency matrix.
[0075] 4-3) Define the hyperparameters of the model.
[0076] 4-4) Train the GCN model.
[0077] 4-5) Predict fatigue life and compare the mean absolute error of the validation set.
[0078] 5) The fatigue pitting life prediction results of the gear pair under test are obtained by using the trained fatigue pitting life prediction model.
[0079] Example 2:
[0080] The main content of this embodiment is the same as that of embodiment 1, except that the specific implementation method is as follows: Figure 1 The process shown includes the following steps:
[0081] 1) Obtain 100 sets of data from published literature to construct the original dataset. Data features include gear contact stress, material elastic modulus, material Poisson's ratio, contact fatigue limit, lubricating oil kinematic viscosity, rotational speed, pitch circle radius, pressure angle, helix angle, tooth width, contact width, surface roughness, surface hardness, core hardness, and fatigue pitting life. The material fatigue characteristic parameter σ′ is evaluated using gear hardness. f and ε′ f Two new data features were obtained. Fatigue pitting lifetime was used as the output feature, and the remaining data features were used as input features. 20% of the original dataset was used as the validation set to verify the model's prediction performance, and 80% was used as the training set for model training. The expressions for evaluating the material fatigue property parameters are as follows:
[0082]
[0083] In the formula, σ′ f ε′ is the fatigue strength coefficient. f The fatigue ductility coefficient, b is the fatigue strength index, and c is the fatigue ductility index.
[0084] 2) The original dataset was expanded using 50 sets of data obtained from the multi-axis fatigue criterion:
[0085] 2-1) To create a two-dimensional model of the gear pair, you can use 3D modeling software such as KISSSOFT, ROMAX, MASTA, or SOLIDWORKS to accurately model the gear pair according to the gear design or machining parameters. To facilitate importing into finite element software, you can save the file in a format such as .sat.
[0086] 2-2) Pre-processing settings for the gear pair model were completed using ABAQUS simulation software. ① Define material properties: Material properties include Young's modulus and Poisson's ratio, with units of mm-N-MPa. All gear pairs were set to use the same material. ② Establish assembly and contact relationships: Based on the characteristic that the driving and driven gears are tangent at their pitch circles, an assembly drawing was created. The contact surface of the driving gear was designated as the driving surface, and the contact surface of the driven gear as the driven surface. The contact method and static friction coefficient were set. ③ Create analysis steps and required output variables: Three analysis steps were created, with large deformation enabled in all cases. The analysis step time and incremental step were set. ④ Mesh generation: A fine mesh was generated for the contact area of the gear pair. To ensure the continuity of the contact pressure image, the mesh size ratio in the tooth diameter and tooth width directions was set to approximately 1:6. Coarse meshing was used for the remaining areas. ⑤ Set loads and constraints: A coupling point was established on the inner ring surface of the driving gear, and a torque load was applied to this node to simulate the actual load transfer. A coupling point was also established on the inner ring surface of the driven gear, and displacement constraints were applied to this node to ensure the normal rotation of subsequent gears.
[0087] 2-3) Submit the job to check the contact torque history of the gear tooth surface and the maximum contact stress to verify the accuracy of the model.
[0088] 2-4) Extract the stress and strain of all nodes from the pitch line of the intermediate gear pair to the tooth root. Specifically, this involves using Python to extract the stress and strain of all nodes from the pitch line of the intermediate gear pair to the tooth root from the ABAQUS simulation software.
[0089] 2-5) Calculate the fatigue life of each node using the FS method. The expression for this method is as follows:
[0090]
[0091] In the formula, Δγ max For the maximum shear strain amplitude, σ max σ is the maximum normal stress on the critical plane, k is a material constant, usually taken as k = 1. Y For yield stress, v e and v p These represent the material's elastic and plastic Poisson's ratios, with a common value of v. e =0.3, v p =0.5.
[0092]
[0093] In the formula, σ res For residual stress, HV top The Vickers hardness of the tooth surface, HV core The value represents the Vickers hardness of the material's core.
[0094] Based on the stress-strain and expression extracted in steps 2-4) The fatigue life of all nodes was calculated using Newton's iterative method.
[0095] 2-6) Predicting gear fatigue pitting life using cumulative frequency distribution plots: with a 3x error band (10 0.5 The cumulative frequency distribution map is obtained by fitting the fatigue life of all nodes at intervals, and the average value of all nodes in the interval with the highest probability is selected as the fatigue pitting life of the gear.
[0096] 2-7) The FS method was validated using experimental data from the literature. Numerous studies have shown that fatigue life prediction errors within a 3x error band are acceptable. For example... Figure 2 As shown, the errors between the experimental lifetimes and the lifetimes predicted by the FS method for the five materials are all within the 3x error band, verifying the feasibility of the method.
[0097] 2-8) Change the load within the range of experimental contact stress and use the FS method to calculate the lifetime to obtain new data.
[0098] 2-9) Input the data obtained using the FS method and the original training set into a typical machine learning method for training, and the prediction results are as follows: Figure 3 As shown, the MAPE value of the validation set decreased significantly. The specific decrease in MAPE value is shown in Table 1.
[0099] Table 1. Prediction results of each model using data generated by FS. MAPE, or Mean Absolute Percentage Error, can be expressed as:
[0100]
[0101] In the formula, This is the predicted fatigue pitting lifetime value. These are the experimental values for fatigue pitting life.
[0102] 3) Expand the training set by building a GAN model: Train the adversarial network GAN model using the expanded training set. Generate 500 new datasets using the trained GAN model. Mix the new datasets with the expanded training set to obtain an enhanced training set.
[0103] 3-1) Import the original training set and perform normalization, as shown in the following expression:
[0104]
[0105] In the formula, x_scalar is the normalized value, x is the input data, and x max x represents the maximum value in the input data.min This is the minimum value in the input data.
[0106] 3-2) Create a generator G and a discriminator D. The goal of the generator is to generate data that is as realistic as possible, causing the discriminator D to incorrectly identify it as real data. The loss of the generator can be expressed as:
[0107]
[0108] The goal of discriminator D is to correctly distinguish between real and generated data. The loss of the discriminator can be expressed as:
[0109]
[0110] 3-3) Build and train the GAN model. The structure of the GAN model is as follows: Figure 4 As shown, the goal of GAN is to train the generator and discriminator by minimizing an adversarial loss function. This can be expressed as follows:
[0111]
[0112] In the formula, D(x) is the output of the discriminator on the real data x, G(z) is the output that maps the random vector z to the data space, and p data (x) represents the distribution of the real data, p z (z) is the distribution of the latent space, set as a standard normal distribution.
[0113] During training, the weights of the generator and discriminator are updated through the following steps: a) Update the discriminator (maximize the discrimination loss). b) Correctly classify the real data x and the generated data G(z). c) Update the generator (minimize the generation loss). d) Aim to make it harder for the discriminator to distinguish between real and generated data, i.e., minimize L. G .
[0114] 3-4) Use a generator to generate new data. After training, the generator G can extract samples from the latent space z, and finally perform inverse normalization to obtain the generated data, which can be represented as:
[0115] x genaerated =G(z) (9)
[0116] In the formula, x genaerated To generate the data, z is derived from the latent distribution p. z The input sampled in (z).
[0117] The inverse normalization process can be expressed as:
[0118] x GAN =x_scalar(x max -xmin )+x min (10)
[0119] In the formula, x GAN Generate data for the final GAN.
[0120] 3-5) Input the preprocessed GAN-generated data and training set into a typical machine learning method for training, and the prediction results are as follows: Figure 5 As shown, the MAPE value of the validation set exhibits a significant downward trend. The specific magnitude of the MAPE value decrease is shown in Table 2. The analysis demonstrates that using GAN-generated data for training effectively improves the model's prediction accuracy. An enhanced training set is obtained by further expanding the training set based on the expanded training set and the GAN model.
[0121] Table 2 shows the prediction results of each model using data generated by GAN.
[0122]
[0123] 4) Build a GCN model for fatigue life prediction:
[0124] 4-1) Create a graph convolutional layer. In GCN, the input data needs to pass through a graph G = (V, E), where V is the node geometry representing different devices or components, and E is the edge geometry representing the relationships between nodes. Each node v i There is an eigenvector x i , can represent the relevant characteristics of the equipment.
[0125] 4-2) Create an adjacency matrix A, where A represents the connection relationships between nodes. ij =1 indicates that node i and node j are connected, A ij =0 indicates that they are not connected.
[0126] 4-3) Define the model's hyperparameters: the number of features is 17 (the number of input features), the graph type is set to random, the graph convolutional layers are set to 2, the final Dense layer has 1 unit to output the predicted value, Dropout is set to 0.2, the batch size is set to 64, and the number of training epochs is set to 300. The final GCN model is as follows. Figure 6 As shown, the Input Layer requires the number of input features and nodes, the Graph Convolution Layer is a graph convolution layer, and the Lambda is a shape transformation layer.
[0127] 4-4) Training the GCN model. GCN aggregates information from neighboring nodes through graph convolution operations. The update expression for each layer of GCN can be represented as:
[0128]
[0129] In the formula, H (l) It is the feature matrix of the nodes in layer l (which is the input feature matrix X in layer 0), and σ is the activation function. It is an adjacency matrix with self-loops. yes The degree matrix, W (l) It is the weight matrix that needs to be learned.
[0130] To predict fatigue life, features are updated through a GCN layer and then mapped to the output using a fully connected layer, as shown in the following expression:
[0131] Y = H (L) W+b (12)
[0132] In the formula, Y is the predicted value, and W and b are the weights and bias terms of the linear layer, respectively.
[0133] 4-5) Compare the mean absolute percentage error (MAPE) on the validation set to determine if the model's prediction performance meets expectations. If not, return to step 4-3) to adjust the model structure or hyperparameters. The final model prediction performance is as follows: Figure 7 As shown, training the augmented datasets obtained using the FS and GAN methods demonstrates better prediction accuracy in the GCN model, with a significant decrease in the model's MAPE and RMSE values. RMSE, or root mean square error, can be expressed as:
[0134]
[0135] like Figure 8 As shown, the method proposed in this invention significantly improves prediction accuracy, with all predicted values within a 3x error band and most within a 2x error band. Further validation of the proposed method is achieved using experimental data from literature outside the original dataset, such as… Figure 9 As shown, the gear fatigue pitting life predicted by this invention matches the experimental data in the literature, further demonstrating the effectiveness and accuracy of the method. In summary, the method described in this example can provide a certain reference for future research on gear fatigue pitting life and can be further extended to other small-sample failure analyses.
[0136] Example 3:
[0137] This embodiment is mainly the same as embodiment 1 or 2. However, in step 2-2), a set needs to be created for the intermediate gear pair and the gear pair contact area, and the output variables are selected as contact pressure, stress, and strain. In step 2-2), the contact area is finely meshed, with the mesh size ratio in the tooth diameter and tooth width directions approximately 1:6. The remaining areas are globally seeded. In step 2-6), a 3x error band (10...) is used... 0.5The cumulative frequency distribution map is obtained by fitting the fatigue life of all nodes at intervals, and the average value of all nodes in the interval with the highest probability is selected as the fatigue pitting life of the gear.
[0138] Example 4:
[0139] This embodiment mainly consists of the same content as any one of embodiments 1 to 3. However, after steps 2-8), it further includes a step of inputting the data obtained with altered loads and the original training set into a typical machine learning method for training, and observing the model's predictive performance. After steps 3-4), it further includes a step of inputting the preprocessed GAN-generated data and the original training set into a typical machine learning method for training, and observing the model's predictive performance. Typical machine learning methods include Gaussian Process (GPS), Support Vector Machine (SVM), Convolutional Network (CNN), and Artificial Neural Network (ANN).
[0140] Example 5:
[0141] This embodiment is largely the same as any one of embodiments 1-4, except that the generation of the adjacency matrix in step 4-2) is selected according to different graph types. In step 4-3), hyperparameters are defined before training the model. Hyperparameters include the number of nodes, graph type, number of units, Dropout rate, batch size, and number of training epochs. After step 4-5), it is determined whether the model's prediction performance meets expectations. If not, the process returns to step 4-3) to adjust the model structure or hyperparameters.
Claims
1. A method for predicting gear fatigue pitting life by integrating fatigue criteria and data-driven approaches, characterized in that, Includes the following steps: 1) A meta-analysis method was used to obtain grouped data from published literature to construct the original dataset. The data characteristics of the grouped data included gear contact stress, elastic modulus, Poisson's ratio, contact fatigue limit, lubricating oil kinematic viscosity, rotational speed, pitch circle radius, pressure angle, helix angle, tooth width, contact width, surface roughness, surface hardness, tooth core hardness, and fatigue pitting life. The original dataset was divided into an original training set and a validation set. 2) The original training set is expanded using the multi-axis fatigue criterion to obtain the expanded training set; Step 2) specifically includes the following sub-steps: 2-1) Based on the gear geometry parameters in the original dataset, use drawing software to draw the model of the gear pair and import it into the ABAQUS simulation software; 2-2) Complete the pre-processing settings of the gear pair model using ABAQUS simulation software; the pre-processing settings include defining material properties, establishing assembly and contact relationships, creating analysis steps and required output variables, mesh generation, and setting loads and constraints; 2-3) Submit the job to view the contact torque history and maximum contact stress of the gear tooth surface to verify the accuracy of the model; 2-4) Extract the stress and strain of all nodes from the pitch line of the intermediate gear pair to the tooth root; 2-5) Calculate the fatigue life of each node using the FS method; 2-6) Use cumulative frequency distribution plots to predict gear fatigue pitting life; 2-7) Validate the FS method using experimental data from the literature; 2-8) Change the load within the range of experimental contact stress, use the FS method to calculate the lifetime to obtain new data, and obtain an expanded training set; 3) Train the GAN model using the expanded training set; generate new data using the trained GAN model; mix the new data with the expanded training set to obtain the enhanced training set; 4) Train the fatigue pitting life prediction model using an enhanced training set until the model converges; wherein, the fatigue pitting life prediction model is a graph convolutional network (GCN) model; the input features of the fatigue pitting life prediction model are the contact stress, elastic modulus, Poisson's ratio, contact fatigue limit, kinematic viscosity of lubricating oil, rotational speed, pitch circle radius, pressure angle, helix angle, tooth width, contact width, surface roughness, surface hardness, and tooth core hardness of the gear, and the output feature is the fatigue pitting life; 5) The fatigue pitting life prediction results of the gear pair under test are obtained by using the trained fatigue pitting life prediction model.
2. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 1, characterized in that: Step 1) uses surface hardness to evaluate the fatigue characteristics of the material. and This yields two new data features; among them, The fatigue strength coefficient, Fatigue ductility coefficient.
3. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 1, characterized in that, Step 3) specifically includes the following sub-steps: 3-1) Import the expanded training set data and perform preprocessing; 3-2) Create the generator and discriminator; 3-3) Build and train the GAN model; 3-4) Use a generator to generate new data; 3-5) The new training set is mixed with the expanded training set to obtain the enhanced training set.
4. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 3, characterized in that: In step 2-2), a set needs to be created for the intermediate gear pair and the contact area of the gear pair, and the output variables are selected as contact pressure, stress and strain; in step 2-2), the contact area is divided into fine meshes, where the mesh size ratio in the tooth diameter and tooth width directions is 1:6, and the remaining areas are divided into global seed meshes.
5. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 3, characterized in that: In steps 2-6), the fatigue life of all nodes is fitted with a 3x error band interval to obtain the cumulative frequency distribution map. The average value of all nodes in the interval with the highest probability is selected as the fatigue pitting life of the gear.
6. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 3, characterized in that: Steps 2-8) are followed by steps to input the data obtained by changing the load and the training set into typical machine learning methods for training and to observe the prediction effect of the model; Steps 3-4) are followed by steps to input the preprocessed GAN-generated data and the original training set into typical machine learning methods for training and to observe the prediction effect of the model; among them, typical machine learning methods include Gaussian process GPS, support vector machine SVM, convolutional network CNN and artificial neural network ANN.
7. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 1, characterized in that, Step 4) specifically includes the following sub-steps: 4-1) Create a graph convolutional layer; 4-2) Create an adjacency matrix; 4-3) Define the hyperparameters of the model; 4-4) Train the GCN model; 4-5) Predict fatigue life and compare the mean absolute error of the validation set.
8. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 1, characterized in that: In step 4-2), the generation of the adjacency matrix is selected according to the different graph types.
9. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 1, characterized in that: In step 4-3), define hyperparameters before training the model; hyperparameters include the number of nodes, graph type, number of units, Dropout rate, batch size, and number of training rounds.
10. The gear fatigue pitting life prediction method integrating fatigue criteria and data-driven approach according to claim 7, characterized in that: After steps 4-5), determine whether the model's prediction performance meets expectations; if not, return to step 4-3) to adjust the model structure or hyperparameters.