Prediction method of fretting wear on milling material surface

Abstract

The present invention discloses a method for predicting micro-motion wear on the surface of milled materials, and relates to the technical field of micro-motion wear. The method comprises the following steps: obtaining micro-motion wear samples on the surface of unprocessed materials and micro-motion wear samples on the surface of milled materials; establishing a functional relationship between the friction coefficient and the wear coefficient of the surface of unprocessed materials and the micro-motion parameters based on nonlinear fitting and the micro-motion wear samples on the surface of unprocessed materials; establishing a finite element model to obtain multiple groups of simulated micro-motion wear samples; bringing the simulated micro-motion wear samples into the BPNN neural network model for training, and after the training, fine-tuning the micro-motion wear samples on the surface of unprocessed materials and the micro-motion wear samples on the surface of milled materials in turn based on the transfer learning method, and obtaining the model BPNN-TLM; based on the model BPNN-TLM, the micro-motion wear can be predicted. The method of the present invention has relatively low cost and high prediction accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of fretting wear, and in particular to a method for predicting fretting wear on the surface of a milling material. Background Art

[0002] Machining, especially milling, plays a key role in the manufacture of high-performance components such as GH4169 high-temperature alloy turbine disks. It not only determines the geometric accuracy and surface characteristics of the workpiece, but also profoundly changes the surface integrity of the material at the microscopic level. The high strain, high strain rate and temperature changes during milling cause significant changes in the microstructure of the surface and subsurface layers of the material. These changes have a significant impact on the fretting wear behavior of the material through the formation of metamorphic layers such as white layers, residual stresses, work hardening and recrystallization. Therefore, it is crucial to clarify the relationship between machining parameters and the fretting wear properties of the workpiece surface, which is related to the performance and durability of high-temperature alloy components in high-temperature and harsh environments such as gas turbines.

[0003] For the fretting wear on the surface of milled materials, we can test it with existing fretting wear testing equipment. However, the testing process will damage the material, making it difficult to reuse it, especially for some high-performance materials, such as GH4169 high-temperature alloy, which is expensive. In addition, due to the large number of milling process parameters, it is expensive to obtain more detailed data. Therefore, existing technicians try to use some methods to predict the fretting wear on the surface of materials, such as phenomenological method, finite element method and machine learning.

[0004] Phenomenological methods provide a modeling approach based on experimental data through multivariate mathematical and statistical analysis, but their high requirements for mathematical skills may limit their application in industrial environments with rapid iterations and real-time decision-making. The finite element method has become an important tool for predicting the mechanical behavior of materials due to its excellent physical simulation capabilities. However, this method has limitations and low accuracy when simulating micro-wear behavior, especially in capturing the changes in material microstructure and property gradients caused by machining. Machine learning has a high accuracy rate, but machine learning models strongly rely on the quality and quantity of data. For example, for machine learning training sets, it usually requires a large number of individuals, and the acquisition of these high-quality data sets is often costly and extremely time-consuming. Summary of the invention

[0005] In order to solve at least one of the above problems, the present invention proposes a method for predicting micro-wear on the surface of a milled material. The method has relatively low cost and high prediction accuracy.

[0006] The technical solution of the present invention is: a method for predicting fretting wear on the surface of a milled material, comprising the following steps:

[0007] S1, obtain 15 to 50 sets of fretting wear samples of unprocessed material surfaces and 10 to 30 sets of fretting wear samples of milled material surfaces;

[0008] S2. Based on the fretting wear samples of the unprocessed material surface, the functional relationship between the friction coefficient and the wear coefficient of the unprocessed material surface and the fretting parameters is obtained by nonlinear fitting;

[0009] S3, based on the functional relationship of S2 and combined with the finite element model, simulate at least 80 groups of simulated fretting wear samples;

[0010] S4, bringing the simulated fretting wear samples into the BPNN neural network model for training. After the training, based on transfer learning, the trained BPNN neural network model is fine-tuned using the fretting wear samples on the surface of the unprocessed material, and then further fine-tuned using the fretting wear samples on the surface of the milled material to obtain the model BPNN-TLM;

[0011] S5. Based on the BPNN-TLM model, the fretting wear parameters of the milling surface are input, which can predict the fretting wear of the milling material surface.

[0012] An implementation manner of the present invention is that the fretting wear sample includes fretting parameters, friction coefficient, wear coefficient and wear rate, and the fretting parameters include displacement amplitude, number of cycles and normal load.

[0013] One implementation of the present invention is that the finite element model includes a micro-motion wear part and a test part. By applying a normal load, a lateral displacement amplitude and a number of cycles to the micro-motion wear part, the wear depth of each node on the surface of the test part is calculated according to the energy wear theory and in combination with the UMESHMOTION subroutine, and the wear condition of the surface of the test part is updated after each incremental step using adaptive mesh technology.

[0014] One implementation of the present invention is to normalize the micro-motion parameters and the wear rate during the training and fine-tuning of the BPNN neural network model.

[0015] One embodiment of the present invention is that, during the transfer learning process, the weights and biases of the input layer and at least one hidden layer of the BPNN neural network model are frozen, and at least one hidden layer and the output layer of the neural network model are fine-tuned based on micro-wear samples of the surface of unprocessed materials or micro-wear samples of the surface of milled materials, so as to finally obtain an optimized model.

[0016] An implementation of the present invention is that, in S1, the base materials of the milling materials are the same and the milling processing parameters are the same.

[0017] Furthermore, the following steps are included: S6, obtaining 10 to 30 groups of micro-motion wear samples of the milling material surface under different milling processing parameters, repeating S4 and S5, that is, being able to predict the micro-motion wear under a variety of different milling parameter conditions.

[0018] Beneficial effect: Based on the BPNN neural network model and transfer learning, the present invention constructs a prediction model for the fretting wear on the surface of milling materials, and predicts the fretting wear on the surface of milling materials through the model. When only a small number of real samples are used, the prediction method of the present invention has a high accuracy rate, and the average relative error is only 6.657%. This shows that the method of the present invention is not only low in cost, but also still has a high accuracy rate in the end.

[0019] At the same time, the intermediate model obtained by the method of the present invention (a model that performs transfer learning only once) can also predict the micro-wear of the surface of unprocessed materials; at the same time, the method of the present invention is not only limited to the prediction of the wear rate of the machined surface, but is also applicable to other fields such as surface modification technology, and has a wider range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A finite element model diagram of an embodiment of the present invention;

[0021] Figure 2 It is the loading history diagram of the finite element model;

[0022] Figure 3 It is a comparison chart of the prediction results and measured values ​​of the model BPNN-TLM;

[0023] Figure 4 It is a comparison chart of the prediction results and measured values ​​of model BPNN-TL and model BPNN. DETAILED DESCRIPTION

[0024] The specific implementation modes of the present invention will be described clearly and completely below with reference to examples and drawings. Obviously, the described examples are only some embodiments of the present invention, rather than all embodiments.

[0025] In the present invention, the unprocessed material refers to the material that has not been milled.

[0026] A method for predicting fretting wear on a milling material surface comprises the following steps:

[0027] S1, obtain 15 to 50 sets of fretting wear samples of unprocessed material surfaces and 10 to 30 sets of fretting wear samples of milled material surfaces;

[0028] In this step, the fretting wear samples are obtained by testing with conventional fretting wear equipment. The fretting wear equipment can adopt existing equipment, such as the equipment developed by the inventor and disclosed in patent number CN201610041396.3. Those skilled in the art can also adopt other equipment.

[0029] At the same time, in this process, in order to avoid errors caused by different materials, the unprocessed material and the milled material belong to the same material. The difference between the two is that the milled material is milled.

[0030] Meanwhile, in this step, the fretting wear sample includes fretting parameters, friction coefficient, wear coefficient and wear rate, and the fretting parameters include displacement amplitude, number of cycles and normal load.

[0031] As for the fretting parameters, they are the three main control factors commonly found in the fretting wear test process. As for the measured values ​​of the friction coefficient and the wear coefficient, they can be calculated by the following formula: , , where COF represents the friction coefficient, E d represents the dissipated energy of a single cycle, F N represents the normal load, D represents the displacement amplitude, represents the energy wear coefficient, and V represents the wear volume.

[0032] The dissipation Ed of a single cycle can be calculated by the following formula: , where represents the tangential friction force, Represents the tangential displacement.

[0033] In this embodiment, the material used is GH4169 high temperature alloy. 37 sets of data are obtained through micro-wear testing on the unprocessed surface of the material, and 28 sets of data are obtained on the milled surface. As for these data, those skilled in the art know that the sample size is much lower than the sample size in the conventional method, so the cost of obtaining the sample is low; at the same time, combined with the other steps of this embodiment, it can be seen that these 37 sets of data and 28 sets of data are only partially used for training the model, and some data are used to verify the trained model, that is, the sample size required for training is smaller.

[0034] S2. Based on the fretting wear samples of the unprocessed material surface, the functional relationship between the friction coefficient and the wear coefficient of the unprocessed material surface and the fretting parameters is obtained by nonlinear fitting;

[0035] In this embodiment, the inventor first obtains the relationship between the friction coefficient and the fretting parameters, and the relationship between the wear coefficient and the fretting parameters of the unprocessed material surface through nonlinear fitting using the fretting wear samples of the unprocessed material surface.

[0036] In this embodiment, the fitted functional relationship is as follows:

[0037] ;

[0038] ;

[0039] Where COF is the coefficient of friction, α represents the wear coefficient, E is the elastic modulus, and the R of the fitting function of the friction coefficient 2 The R of the fitting function of the wear coefficient is 0.851. 2 It is 0.885, indicating that it has a good correlation.

[0040] S3, based on the functional relationship of S2 and combined with the finite element model, simulate at least 80 groups of simulated fretting wear samples;

[0041] As mentioned above, since the actual micro-motion wear experiment has high cost and long experimental time, in order to save cost and experimental time, in this embodiment, the simulated micro-motion samples are obtained based on the finite element model. For the samples, those skilled in the art know that if a higher precision BPNN model is to be obtained, the larger the sample size, the better. Based on the actual situation of the present invention, the inventor found that when the sample size is at least 80 groups, the accuracy of the model finally trained is low. In this embodiment, 116 groups of micro-motion wear samples are obtained through the finite element model. Of course, those skilled in the art can also obtain more micro-motion wear samples according to the actual situation.

[0042] The finite element model includes micro-motion wear parts and unprocessed materials. By applying normal load, lateral displacement amplitude and number of cycles to the micro-motion wear parts, the wear depth of each node on the surface of the test piece is calculated according to the energy wear theory and combined with the UMESHMOTION subroutine. Using the adaptive mesh technology ALE, the wear condition of the test piece surface is updated after each incremental step.

[0043] In this embodiment, if Figure 1As shown, this embodiment is mainly obtained by using ABUQUS software. The micro-wear part is an aluminum oxide ball, which is hemispherical with a radius of 4 mm, an elastic modulus of 350 GPa, and a Poisson's ratio of 0.23. The test piece is a GH4169 plate with a size of 6×6×3 mm, an elastic modulus of 220 GPa, and a Poisson's ratio of 0.3. The mesh uses C3D8 units, and the contact area between the micro-motion wear part and the contact part is refined to 20×20×10μm. The contact interface adopts a master-slave setting, the friction part is defined as the master surface, and the test piece is defined as the slave surface. Multi-point constraints are applied to the top of the hemisphere, the normal load and displacement amplitude are applied to the top center of the micro-motion wear part, the bottom test piece is fixed, and the center area of ​​the specimen (0.8×0.6×0.22mm) is set as an arbitrary Lagrangian-Euler domain to ensure the convergence of the calculation. During the cycle, the wear depth under single cycle conditions is estimated based on the energy wear theory and combined with the UMESHMOTION subroutine, and then the wear area is updated for the next cycle.

[0044] In this process, the loading process is as follows Figure 2 shown.

[0045] The above-mentioned finite element model is the operating method and operating parameters of this embodiment. Those skilled in the art can obtain multiple groups of fretting wear samples by using different operating methods and operating parameters according to actual conditions.

[0046] S4, the simulated fretting wear samples are brought into the BPNN neural network model for training. After the training, based on transfer learning, the trained BPNN neural network model is fine-tuned using the fretting wear samples on the surface of the unprocessed material, and then further fine-tuned using the fretting wear samples on the surface of the milled material to obtain the model BPNN-TLM;

[0047] In this embodiment, the BPNN neural network model used is similar to the conventional model, and is composed of an input layer, a hidden layer, and an output layer, wherein the input layer contains 3 neurons, and for the displacement amplitude, number of cycles, and normal load in the micro-motion parameters, the output layer consists of 1 neuron for predicting the wear rate. For the hidden layer, we use the trial and error method to determine, and finally obtain a BPNN neural network structure of 3×8×16×16×8×1. At the same time, for the output layer, a linear activation function is used, and a ReLU activation function is used in the hidden layer. These are all conventional operations in this field, and those skilled in the art can also choose BPNN neural network models with different structures according to the situation.

[0048] In the process of model training, in order to reduce the influence of abnormal samples on it and accelerate the convergence speed of the model, this embodiment uses normalization processing for the parameters in the sample, and the normalization processing is performed using the following formula: , where is the sample data After normalization, and are the maximum and minimum values ​​of the sample data respectively.

[0049] In the process of model training, mean square error is used as the loss function. In order to prevent overfitting linearity of small sample problems, L2 regularization method is introduced. At the same time, Adam optimization algorithm is used to update the weights and biases of the model.

[0050] Since the simulation data has a large deviation compared to the real data, after training the BPNN neural network with 116 sets of simulated micro-motion samples, we migrated the model framework to the BPNN neural network model through transfer learning, and then performed transfer learning on the above model through micro-motion wear samples of the unprocessed material surface and micro-motion wear samples of the milled material surface. By fine-tuning the above model, we obtained a model suitable for predicting micro-motion wear on the milled material surface.

[0051] During the transfer learning process, the weights and biases of the input layer and at least one hidden layer of the BPNN neural network model are frozen, and at least one hidden layer and the output layer of the neural network model are fine-tuned in turn through the micro-wear samples of the unprocessed material surface and the micro-wear samples of the milled material surface, and finally an optimized model, namely the BPNN-TLM model, is obtained.

[0052] For materials, there is a big difference in performance before and after milling. According to the inventor's research, for GH4169 high-temperature alloy, before milling, the hardness of the material along the depth direction did not change significantly, about 425HV; while for the material after milling, its hardness along the depth direction showed a trend of first increasing, and its hardness reached the maximum at a depth of 40μm, which was 494HV, and its hardness was not lower than that of the material before milling. There is also a big difference in residual tensile stress. Therefore, in the actual research process, we cannot equate the material before milling with the material after milling.

[0053] To this end, the inventors conducted further research on the two, and observed the changing relationship between the friction coefficient and the wear rate under different micro-motion parameter conditions by changing one of the micro-motion parameters. The results show that when the normal load is changed, the friction coefficient decreases with the increase of the normal load, and the wear rate shows an initial increase and then decreases, regardless of whether it is the material before or after milling; when the displacement amplitude is changed, the friction coefficient of both increases with the increase of the displacement amplitude, and the wear rate shows a nonlinear growth trend; when the number of cycles is changed, the friction coefficient of both first increases and then decreases with the number of cycles, and finally stabilizes, and the wear rate of both increases first and then decreases.

[0054] The inventors have found from the above research results that the changes in friction coefficient and wear rate of the material before and after milling under different micro-motion parameter conditions have strong similarities, and the change trends are basically the same.

[0055] Under the condition of the same change trend, the prediction model obtained by transfer learning finally has better accuracy.

[0056] In this embodiment, since the structure of the BPNN neural network model is 3×8×16×16×8×1, the weights and biases of the first 4 layers of the model are frozen during transfer learning. During fine-tuning, only the last 2 layers of the model are changed. By using the actual samples obtained and fine-tuning the weights and biases of the last 2 layers, the final model BPNN-TL model has better accuracy.

[0057] In this embodiment, during transfer learning, there are 25 groups of fretting wear samples on the surface of unprocessed materials and 16 groups of fretting wear samples on the surface of milled materials.

[0058] From the above, we can see that although the final model is the BPNN-TLM model, after transfer learning through the micro-wear samples of the unprocessed material surface, an intermediate model BPNN-TL can be obtained. Through this model, the micro-wear of the unprocessed material surface can be better predicted.

[0059] S5. Based on the BPNN-TLM model, the fretting wear parameters are input to predict the fretting wear.

[0060] In order to illustrate the prediction accuracy of the model obtained by the embodiment of the present invention, 12 groups of micro-wear samples of milling material surfaces were used as test sets to test the accuracy of the model BPNN-TLM. At the same time, in order to illustrate the accuracy of the present invention, the model BPNN trained only with simulated micro-wear samples was directly used as a comparison. The final results are shown in Figure 3 As shown, from Figure 3 It can be seen that the prediction results obtained by the method of the embodiment of the present invention are highly correlated with the measured values, with an average relative error of 6.657%, R 2 It is 0.9465.

[0061] In order to illustrate that the model that only performs transfer learning once in S4 of this embodiment can be used for the prediction of fretting wear on the surface of unprocessed materials, 12 groups of fretting wear samples on the surface of unprocessed materials are used as test sets to test the accuracy of the model BPNN-TL. In order to illustrate the accuracy of the present invention, the model BPNN trained only with simulated fretting wear samples is directly used as a comparison. The final results are shown in Figure 4 As shown. Figure 4It can be seen that the prediction results obtained by the method of the embodiment of the present invention are highly correlated with the measured values, with an average error of 6.23% and R 2 is 0.9599. In the comparison model, the average relative error is 15.9%, R 2 It is 0.76.

[0062] After completing the above operations, the prediction of the micro-wear of the milling material surface under the same milling process conditions is completed. However, since there are many control parameters in the milling process, the following steps are also included: S6, obtain 10 to 30 groups of micro-wear samples of the milling material surface under different milling processing parameters, repeat S4 and S5, and the micro-wear under a variety of different milling parameter conditions can be predicted.

[0063] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A method for predicting fretting wear on the surface of a milled material, characterized in that: The following steps are involved: S1, obtaining 15 to 50 groups of fretting wear samples of unprocessed material surfaces and 10 to 30 groups of fretting wear samples of milled material surfaces; the fretting wear samples include fretting parameters, friction coefficient, wear coefficient and wear rate, and the fretting parameters include displacement amplitude, number of cycles and normal load; S2. Based on the fretting wear samples of the unprocessed material surface, the functional relationship between the friction coefficient and the wear coefficient of the unprocessed material surface and the fretting parameters is obtained by nonlinear fitting; S3, based on the functional relationship of S2 and combined with the finite element model, simulate at least 80 groups of simulated fretting wear samples; S4, bringing the simulated fretting wear samples into the BPNN neural network model for training. After the training, based on transfer learning, the trained BPNN neural network model is fine-tuned using the fretting wear samples on the surface of the unprocessed material, and then further fine-tuned using the fretting wear samples on the surface of the milled material to obtain the model BPNN-TLM; S5. Based on the BPNN-TLM model, the fretting wear parameters of the milling surface are input, which can predict the fretting wear of the milling material surface; In S4, the transfer learning includes the following sub-steps: freezing the weights and biases of the input layer and at least one hidden layer of the BPNN neural network model, fine-tuning at least one hidden layer and the output layer of the neural network model based on the micro-wear samples of the surface of the unprocessed material or the micro-wear samples of the surface of the milled material, and finally obtaining the fine-tuned model.

2. The method according to claim 1, characterized in that The finite element model includes a micro-motion wear part and a test part. By applying a normal load, a lateral displacement amplitude and a number of cycles to the micro-motion wear part, the wear depth of each node on the surface of the test part is calculated according to the energy wear theory and combined with the UMESHMOTION subroutine. The wear condition of the surface of the test part is updated after each incremental step using the adaptive grid technology.

3. The method according to claim 1, characterized in that In the process of training and fine-tuning the BPNN neural network model, the micro-motion parameters and wear rate are normalized.

4. The method according to claim 1, characterized in that: In S1, the base materials of the milling materials are the same, and the milling processing parameters are the same.

5. The method according to claim 4, characterized in that The following steps are also included: S6. Obtain 10 to 30 groups of fretting wear samples of the milled material surface under different milling processing parameters, and repeat S4 and S5, so that the fretting wear under different milling parameters can be predicted.

Images