Material surface fretting wear prediction method and system based on transfer learning

Abstract

The invention discloses a material surface fretting wear prediction method and system based on transfer learning, and the method comprises the steps: constructing a fretting wear simulation model, and achieving the precise simulation of the motion form of a workpiece; meanwhile, aiming at a wear coefficient in the fretting wear simulation model, an attention mechanism is used for solving an influence weight of multi-source data on the wear coefficient to serve as an initial value, and the prediction of the fretting wear performance under a new working condition is quickly adapted by setting the weight and an offset, so that the high robustness of the model is realized; besides, a fretting wear simulation model is used for physical simulation to obtain a data set, a simulation fretting wear prediction model is trained in advance, and after general knowledge of fretting wear behaviors is learned by the model, a hidden layer in the model is divided into a freezing layer and a self-adaptive fine tuning layer according to the universality of wear knowledge, so that the fretting wear behavior is predicted. And a real data set is used to carry out fine tuning operation on the adaptive fine tuning layer, so that the model is closer to a real prediction result.

Description

Technical Field

[0001] This invention belongs to the field of digital intelligence empowerment technology, specifically relating to a method and system for predicting micro-motion wear on material surfaces based on transfer learning. Background Technology

[0002] The turbine disk tenon and mortise structure achieves a reliable connection between the blades and the turbine disk. It not only completes the precision assembly of the blades and turbine disk but also effectively reduces the impact of thermal deformation caused by thermal expansion during high-temperature operation, playing a significant role in high-end equipment such as aero-engines. Under the complex working environment of high temperature, high pressure, and high speed, the contact surface of the turbine disk tenon and mortise structure is continuously affected by fretting wear. Irreversible plastic deformation occurs in the contact area, leading to a continuous decrease in the fretting wear performance of the material surface. When the wear reaches a certain level, the tenon and mortise connection may loosen, potentially causing serious problems such as blade detachment. In engineering applications, most fault detection methods rely on manual inspection, which is insufficient to meet the needs of rapid maintenance in modern equipment. Therefore, establishing a surface performance fault prediction model is crucial for improving the operational efficiency of the turbine disk tenon and mortise structure.

[0003] Currently, deep learning methods are often used to build prediction models for surface properties. For example, patent application number CN202110658416.2 discloses a deep learning-based method for predicting the electromagnetic performance of nano-antenna arrays. This method comprehensively considers the interaction of three factors: the shape of the nanorod, the wavelength, and the scattering cross-section. By simulating the structure of the nano-antenna array and performing finite element numerical simulation to form a training set, a prediction model for the electromagnetic performance of the nano-antenna array is obtained by relying on pure data-driven deep learning.

[0004] However, building a prediction model for the fretting wear performance of turbine disk tenon and mortise structures using deep learning requires a large amount of reliable data for training. However, due to the large wear area of ​​the tenon and mortise structure contact surface, its non-reusability, and the limited material, only a small sample of real data can be obtained. Furthermore, wear is a gradual process, and simulation experiments of turbine disk tenon and mortise structures cannot guarantee the effectiveness of the wear process, resulting in low accuracy of the simulation model. In addition, prediction models trained using datasets obtained under specific working conditions can only predict fretting wear performance under the same conditions, making it difficult to predict fretting wear performance under different environments, thus indicating insufficient generalization ability of the model. Summary of the Invention

[0005] This invention addresses the problems of limited high-quality data, low model accuracy, and poor generalizability in predicting the fretting wear performance of turbine disk tenon and mortise structures. It proposes a material surface fretting wear prediction method and system based on transfer learning, achieving low defect detection rate, high robustness, and wide scalability in fretting wear performance.

[0006] In a first aspect, the present invention provides a method for predicting fretting wear on material surfaces based on transfer learning, the method comprising the following steps:

[0007] Obtain real and simulation datasets containing key operating condition parameters and corresponding fretting wear data; construct a simulation fretting wear prediction model, with key operating conditions as input data and fretting wear data as output data; train the simulation fretting wear prediction model using the simulation dataset;

[0008] A realistic fretting wear prediction model is constructed. The realistic fretting wear prediction model has the same structure as the simulated fretting wear prediction model. The weights and biases of the trained simulated fretting wear prediction model are used as the initial weights and initial biases in the realistic fretting wear prediction model.

[0009] The real fretting wear prediction model includes a primary frozen layer and an adaptive fine-tuning layer; the real fretting wear prediction model is trained using a real dataset; during training, only the parameters in the adaptive fine-tuning layer are updated, while the parameters in the primary frozen layer remain unchanged.

[0010] The trained real fretting wear prediction model is used to predict the fretting wear data of the tested material.

[0011] Preferably, the simulation dataset is obtained using the following method:

[0012] A fretting wear simulation model including a first wear component and a second wear component was constructed, and the model was meshed. The actual operation of the turbine disk tenon and mortise structure was simulated by the relative motion of the first and second wear components, and fretting wear data under different working conditions were extracted as simulation results. The fretting wear simulation model was adjusted and optimized by comparing the simulation results with the actual experimental results. By repeating the above adjustment and optimization process, the optimal fretting wear simulation model was finally constructed. The fretting wear data obtained by the optimal fretting wear simulation model under different key working conditions were obtained, and a simulation dataset was constructed by using the key working conditions and the fretting wear data.

[0013] Preferably, the first wear component is a cube; the second wear component is a hemisphere; the material coefficients of the first wear component and the second wear component are the same; a surface on the first wear component is selected as the contact slave surface, and the semicircular surface of the second wear component is selected as the contact master surface; the first wear component and the second wear component are simulated for fretting wear through the contact slave surface and the contact master surface.

[0014] Preferably, the method for obtaining the fretting wear data in the simulation dataset is as follows:

[0015] The contact stress and slippage at the contact points between the first and second wear parts are extracted, and fretting wear data are obtained through the contact stress and slippage. Its expression is:

[0016]

[0017] Where μ is the wear coefficient; This represents the average value of the contact stress. H represents the average slippage; H represents the hardness of the first wear component.

[0018] Preferably, the method for obtaining the wear coefficient is as follows:

[0019] The material coefficients of the tested material are vectorized, and the vectorized material coefficients are mapped to query matrix, key matrix and value matrix through linear projection. Attention scores are obtained for different material coefficients based on the query matrix, key matrix and value matrix corresponding to different material coefficients. The initial wear coefficient K is obtained based on the attention score and material coefficient. The initial wear coefficient K is updated by introducing fusion weight and adjustment bias to obtain the updated wear coefficient μ.

[0020] Preferably, the key operating conditions include load, offset, and number of friction cycles.

[0021] Preferably, the loss function of the simulated micro-motion wear prediction model during the training process includes a mean square error loss function and a physical loss function.

[0022] Preferably, the design method for the primary freezing layer and the adaptive fine-tuning layer is as follows:

[0023] The physical loss function values ​​corresponding to each hidden layer of the simulation fretting wear prediction model are obtained during the training process. The hidden layer with a function value less than a preset threshold is selected as the primary frozen layer, and the remaining hidden layers are selected as adaptive fine-tuning layers.

[0024] Preferably, the simulated fretting wear prediction model employs a multilayer perceptron.

[0025] Secondly, this invention provides a material surface fretting wear prediction system based on transfer learning, which is used to execute the aforementioned material surface fretting wear prediction method. This material surface fretting wear prediction system includes a simulation model construction module, a data acquisition module, and a fretting wear data prediction module. The simulation model construction module simulates the actual operation of the turbine disk tenon and mortise structure by constructing a fretting wear simulation model; the data acquisition module is used to extract simulation results and real experimental results to construct simulation datasets and real datasets; the fretting wear data prediction module is used to predict the fretting wear data of the tested material.

[0026] The beneficial effects of this invention are:

[0027] 1. This invention addresses the differences in material processing conditions by using an attention mechanism to weightedly fuse features from multiple data sources to obtain the initial wear coefficient, thereby improving the generalization ability under small sample sizes. Simultaneously, it performs weight and bias allocation on the initial wear coefficient to achieve transfer learning, enabling rapid adaptation to accurate prediction of fretting wear performance under different working conditions, thus improving the model's learning efficiency and robustness.

[0028] 2. This invention trains a fretting wear prediction model using a hierarchical learning method. The model incorporates an initial frozen layer and an adaptive fine-tuning layer, and then introduces small experimental sample data to fine-tune the model, thereby improving its accuracy. This reduces the model's need for small sample data, decreases computational load, and accelerates training, solving the problem of overfitting caused by insufficient sample data. Furthermore, this invention utilizes physical simulation data to enable the model to grasp the fundamental laws of the prediction domain, improving its transferability.

[0029] 3. This invention proposes a method for simulating the fretting wear test of the mortise and tenon structure of a turbine disk, namely, a method of dynamic mesh adjustment and refinement and embedding of material physical wear equations. Through real-time feedback optimization subroutine, it achieves low deviation between the simulation environment and the actual working condition experiment, and improves the reliability of simulation data. Attached Figure Description

[0030] Figure 1 This is the overall flowchart of the present invention.

[0031] Figure 2 This is a schematic diagram of the fretting wear simulation model in this invention.

[0032] Figure 3 This is a schematic diagram of the wear coefficient acquisition method in this invention.

[0033] Figure 4 This is a partial top view of the grid in this invention before adaptive adjustment.

[0034] Figure 5 This is a partial top view of the mesh in this invention after adaptive adjustment.

[0035] Figure 6 This is a schematic diagram of the fretting wear prediction model in this invention. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings.

[0037] A method for predicting fretting wear on material surfaces based on transfer learning is presented. The fretting wear prediction system includes a simulation model building module, a data acquisition module, and a fretting wear data prediction module. The simulation model building module constructs a fretting wear simulation model to simulate the actual operation of a turbine disk tenon and mortise structure. The data acquisition module extracts simulation results and real experimental results to construct simulation datasets and real datasets. The fretting wear data prediction module includes a simulation fretting wear prediction model and a real fretting wear prediction model. The simulation fretting wear prediction model is trained and used to predict on the simulation dataset. The real fretting wear prediction model inherits data from the simulation fretting wear prediction model and is trained and used to predict on the real dataset.

[0038] like Figure 1 As shown, the method for predicting fretting wear on the material surface includes the following steps:

[0039] Step 1: Construct a fretting wear simulation model

[0040] like Figure 2 As shown, the fretting wear simulation model includes a first wear component 1 and a second wear component 2. The first wear component 1 is a cube with a length of 20 mm, a width of 10 mm, and a height of 5 mm. The second wear component 2 is a hemisphere with a radius of 6.35 mm. The material of both the first wear part 1 and the second wear part 2 is set to GH4169 high-temperature alloy, and their material properties are set.

[0041] In this embodiment, the first wear part 1 and the second wear part 2 have the same material coefficient, with a Poisson's ratio of 0.3 and a density of [missing information]. Its elastic modulus is 206 GPa, and its coefficient of thermal expansion is... It has a thermal conductivity of 14.7 W / (m•k) (20℃), a specific heat capacity of approximately 0.435 J / (g•℃), and a coefficient of friction of 0.6.

[0042] Step 2: Changes in wear data due to multiple data sources lead to variations. For example, changes in various material coefficients, such as Poisson's ratio, elastic modulus, and coefficient of thermal expansion, result in different wear conditions, posing a challenge to the transfer learning of wear prediction models. To address this issue, a multi-fidelity training strategy is proposed. This strategy uses an attention mechanism to weightedly fuse features from multiple data sources, improving generalization ability under small sample sizes. The specific process is as follows:

[0043] 2-1. Adaptive modification of wear coefficient based on the material coefficient of the tested material

[0044] like Figure 3As shown, since the wear coefficient K changes significantly with the material factor, before training the fretting wear prediction model for different materials, the wear coefficient K was modified to maximize the universality of the pre-trained model, making it more consistent with actual working conditions and applicable to different materials. The relationship between the wear coefficient K and the material factor of the tested material is expressed as:

[0045]

[0046] in, The mapping function relationship between material coefficient and wear coefficient; P is Poisson's ratio of the tested material; f is the surface friction coefficient of the tested material; E is the elastic modulus of the tested material. C is the density of the material being tested; C is the specific heat capacity of the material being tested. The coefficient of thermal expansion of the material being tested; The value is the thermal conductivity of the material being tested.

[0047] The material coefficients are vectorized, and then mapped to a query matrix through linear projection. Key matrix Sum matrix ;in, M represents the number of material coefficients in the tested material that affect the wear coefficient. Attention scores for different material coefficients are obtained based on the query matrix, key matrix, and value matrix corresponding to different material coefficients. Its expression is:

[0048]

[0049] in, Let be the dimension of the key matrix; This represents the activation function.

[0050] The initial wear coefficient K is obtained based on the attention score and material coefficient, and its expression is as follows:

[0051]

[0052] in, is the m-th material coefficient of the material being tested.

[0053] 2-2. To further optimize the model's adaptability to new materials and working conditions, an offset is added to update the initial wear coefficient K. This allows the model to learn patterns from the data, continuously updating and iterating the weights and offsets, making the prediction results closer to reality and more accurate. The expression for the updated wear coefficient μ is:

[0054]

[0055] in, Represents the fitted function; B represents the fusion weight; B represents the adjustment bias.

[0056] Step 3: Set boundary conditions

[0057] The C3D8 (8-node hexahedron) mesh generation method was used to mesh the fretting wear simulation model. A face on the first wear component 1 was selected as the contact slave face, and the semi-circular face of the second wear component 2 was selected as the contact master face. Surface condition processing was performed on the contact area between the contact slave and contact master faces, and the roughness and oxidation degree of the contact surfaces were set to simulate real working conditions. To obtain a dataset of turbine disk fretting wear performance under different working conditions, different motion parameters were set: loads of 10N, 15N, and 20N; offsets of 10μm, 20μm, and 30μm; and friction cycles of 500, 1000, and 1500.

[0058] Step 4: Wear Simulation

[0059] A fretting wear simulation model was used to simulate the actual operation of the turbine disk tenon and mortise structure. A subroutine, DISP, was developed for motion control to simulate both symmetrical and asymmetrical motions. The DISP program was used to control the motion of the fretting wear simulation model. After each wear event, the contact stress F and slippage S of some nodes in the contact area of ​​the first wear component 1 and the second wear component 2 were extracted. To reduce the impact of data fluctuations caused by errors in edge regions and to increase data accuracy, the extracted data were averaged to obtain the contact stress generated on the contact surface after each wear event. With slip It is represented as:

[0060]

[0061]

[0062] in, This represents the contact stress at the j-th node in the contact area after the i-th wear event; This represents the slippage of the j-th node in the contact area after the i-th wear event; N represents the number of nodes in the contact area.

[0063] Fretting wear data is calculated using the contact stress and slippage generated on the contact surface after each wear cycle. Its expression is:

[0064]

[0065] in, H represents the wear amount obtained from the i-th wear behavior; H represents the hardness of the first wear component.

[0066] Step 5: Extract the displacement amplitude, stress, strain, and friction force of the primary contact surface, as well as the friction depth, contact pressure, stress, strain, slip, and wear morphology of the secondary contact surface obtained from the simulation. Since fretting wear is a subtle, gradual process, such as... Figure 4 As shown, each wear event disrupts the mesh morphology of the cube's upper surface, leading to a decrease in computational accuracy and simulation deviations. Therefore, an automatic mesh update subroutine, UMESHMOTION, was written to automatically update the mesh after each wear event. The results are as follows. Figure 5 As shown, by comparing the simulation results with the actual experimental results, adjustments and optimizations were made to the mesh update and motion control.

[0067] Step 6: Repeat steps 2 to 5, and continue to simulate and iterate until the termination condition is met. Finally, construct the optimal fretting wear simulation model, which can accurately simulate the wear of the turbine disk tenon and mortise structure under specific working conditions in actual operation.

[0068] Step 7: Obtain the optimal fretting wear simulation model by performing FEM (Finite Element Method) simulations under different key operating conditions. Key operating conditions include load, offset, and number of friction cycles. A simulation dataset is constructed based on the key operating conditions and the fretting wear data. To reduce the influence of outlier data and improve model convergence, a general predictive model is obtained. The fretting wear data is then standardized. Represented as:

[0069]

[0070] in, This represents the b-th fretting wear data under the a-th critical operating condition before standardization. and These represent the maximum and minimum fretting wear data under the a-th critical operating condition, respectively; A represents the number of different key operating conditions in the dataset; ; This refers to the number of fretting wear data under the same critical operating conditions.

[0071] Step 8, as follows Figure 6 As shown, a simulation-based fretting wear prediction model is constructed. Its input data consists of key operating conditions, and its output data is the fretting wear data. After receiving the input data, the simulation-based fretting wear prediction model processes the input data using a built-in nonlinear activation function, assigning weights and biases to different neural network layers to ultimately obtain the predicted fretting wear data.

[0072] The simulation fretting wear prediction model was trained using the backpropagation (BP) algorithm on a simulation dataset. During training, the mean squared error (MSE) was combined with the physical equations to construct the loss function. It is represented as:

[0073]

[0074] Where ɑ is the weighting coefficient; For physical loss function; The mean squared error is expressed as follows:

[0075]

[0076] Where N represents the number of samples in the simulation dataset; This represents the i-th simulated fretting wear data; This represents the i-th predicted fretting wear data.

[0077] The resulting error is fed back from the output layer to the hidden layer. The weights in the hidden layer are continuously adjusted via gradient descent until the error is minimized, thus completing the training of the simulated fretting wear prediction model. During this process, because the loss function incorporates physical equations, the output closely matches both the original data and the physical equations, resolving the black-box problem of pure data models.

[0078] In this embodiment, the simulation fretting wear prediction model adopts a multilayer perceptron (MLP).

[0079] Step 9: Construct a realistic fretting wear prediction model

[0080] The real fretting wear prediction model has the same structure as the simulated fretting wear prediction model; the weights and biases of the trained simulated fretting wear prediction model are used as the initial weights and biases in the real fretting wear prediction model. The real fretting wear prediction model includes an initial frozen layer and an adaptive fine-tuning layer. The function values ​​of the physical loss function corresponding to each hidden layer of the simulated fretting wear prediction model are obtained during training, and the hidden layers with function values ​​less than a preset threshold are selected as the initial frozen layer, while the remaining hidden layers are used as adaptive fine-tuning layers.

[0081] Step 10: Train a realistic fretting wear prediction model

[0082] A real dataset was constructed based on fretting wear data obtained under different key operating conditions in actual experiments, and the real fretting wear prediction model was trained using this dataset. Since the entire wear process of the mortise and tenon structure varies considerably, with each process possessing different characteristics, while most wear processes exhibit some similarity, fine-tuning was performed on the small sample prediction model. The hidden layer containing the wear universality feature extraction module in the simulated fretting wear prediction model was "frozen," i.e., the weights, bias parameters, and relational functions in the initial frozen layer were fixed, and the learning rate was set to 0. The adaptive fine-tuning layer is significantly affected by specific operating conditions and has only limited universality; therefore, small sample data was used for fine-tuning.

[0083] To avoid overfitting caused by small sample data during training, the learning rate in the adaptive fine-tuning layer is set to an initial value of 1×e-5, which is linearly increased to 1×e-3 with the number of iterations. A larger learning rate is used in the output layer to enable it to quickly adapt to the new task. This can avoid destroying the previously trained general features and avoid getting stuck in local optima.

[0084] During training, the physical equations of the Archard model and energy wear theory are embedded into the loss function as feedback. The adaptive fine-tuning layer within the model iteratively fine-tunes the model, starting from the output wear value and using gradient descent to redistribute weights and biases along the direction of maximum gradient. Due to the small number of samples in the real dataset, only a few training iterations are performed to appropriately reduce training time and prevent overfitting.

[0085] Step 11: Use the trained real fretting wear prediction model to predict the fretting wear data of the tested material.

Claims

1. A method for predicting fretting wear on material surfaces based on transfer learning, characterized in that: The method includes the following steps: Obtain real and simulation datasets containing key operating condition parameters and corresponding fretting wear data; construct a simulation fretting wear prediction model, with key operating conditions as input data and fretting wear data as output data; train the simulation fretting wear prediction model using the simulation dataset; A realistic fretting wear prediction model is constructed. The realistic fretting wear prediction model has the same structure as the simulated fretting wear prediction model. The weights and biases of the trained simulated fretting wear prediction model are used as the initial weights and initial biases in the realistic fretting wear prediction model. The real fretting wear prediction model includes a primary frozen layer and an adaptive fine-tuning layer; the real fretting wear prediction model is trained using a real dataset; during training, only the parameters in the adaptive fine-tuning layer are updated, while the parameters in the primary frozen layer remain unchanged. The trained real fretting wear prediction model is used to predict the fretting wear data of the tested material.

2. The method for predicting fretting wear on material surfaces based on transfer learning according to claim 1, characterized in that: The simulation dataset is obtained as follows: A fretting wear simulation model including a first wear component and a second wear component was constructed, and the model was meshed. The actual operation of the turbine disk tenon and mortise structure was simulated by the relative motion of the first and second wear components, and fretting wear data under different working conditions were extracted as simulation results. The fretting wear simulation model was adjusted and optimized by comparing the simulation results with the actual experimental results. By repeating the above adjustment and optimization process, the optimal fretting wear simulation model was finally constructed. The fretting wear data obtained by the optimal fretting wear simulation model under different key working conditions were obtained, and a simulation dataset was constructed by using the key working conditions and the fretting wear data.

3. The method for predicting fretting wear on material surfaces based on transfer learning according to claim 2, characterized in that: The first wear component is a cube; the second wear component is a hemisphere; the material coefficients of the first wear component and the second wear component are the same; a surface on the first wear component is selected as the contact slave surface, and the semicircular surface of the second wear component is selected as the contact master surface; the first wear component and the second wear component are simulated for fretting wear through the contact slave surface and the contact master surface.

4. The method for predicting fretting wear on material surfaces based on transfer learning according to claim 2, characterized in that: The method for obtaining the fretting wear data in the simulation dataset is as follows: The contact stress and slippage at the contact points between the first and second wear parts are extracted, and fretting wear data are obtained through the contact stress and slippage. Its expression is: Where μ is the wear coefficient; This represents the average value of the contact stress. H represents the average slippage; H represents the hardness of the first wear component.

5. The method for predicting fretting wear on material surfaces based on transfer learning according to claim 4, characterized in that: The method for obtaining the wear coefficient is as follows: The material coefficients of the tested material are vectorized, and the vectorized material coefficients are mapped to query matrix, key matrix and value matrix through linear projection. Attention scores are obtained for different material coefficients based on the query matrix, key matrix and value matrix corresponding to different material coefficients. The initial wear coefficient K is obtained based on the attention score and material coefficient. The initial wear coefficient K is updated by introducing fusion weight and adjustment bias to obtain the updated wear coefficient μ.

6. The method for predicting fretting wear on material surfaces based on transfer learning according to claim 1, characterized in that: The loss function of the simulated micro-motion wear prediction model during training includes the mean square error loss function and the physical loss function.

7. The method for predicting fretting wear on material surfaces based on transfer learning according to claim 6, characterized in that: The design method for the primary freezing layer and the adaptive fine-tuning layer is as follows: The physical loss function values ​​corresponding to each hidden layer of the simulation fretting wear prediction model are obtained during the training process. The hidden layer with a function value less than a preset threshold is selected as the primary frozen layer, and the remaining hidden layers are selected as adaptive fine-tuning layers.

8. The method for predicting fretting wear on material surfaces based on transfer learning according to claim 1, characterized in that: The key operating conditions include load, offset, and number of friction cycles.

9. The method for predicting fretting wear on material surfaces based on transfer learning according to claim 1, characterized in that: The simulated fretting wear prediction model employs a multilayer perceptron.

10. A material surface fretting wear prediction system based on transfer learning, characterized in that: This system is used to execute the material surface fretting wear prediction method based on transfer learning as described in claim 1; the material surface fretting wear prediction system includes a simulation model construction module, a data acquisition module, and a fretting wear data prediction module. The simulation model building module simulates the actual operation of the turbine disk tenon and mortise structure by constructing a fretting wear simulation model; the data acquisition module is used to extract simulation results and real experimental results to construct simulation datasets and real datasets. The fretting wear data prediction module is used to predict fretting wear data of the material under test.

Images