A method for measuring material wear resistance based on machine learning

Abstract

The application discloses a kind of based on machine learning accurate measurement material wear resistance's method, comprising the following steps: ① data collection: collection material in friction and wear working condition Friction data, measure the wear volume of material, with Friction data as input value, wear volume is output value and is made into data set;② model construction: using the Keras module under TensorFlow framework builds the convolutional neural network model of actual wear volume prediction;③ model verification: including the data set is divided into training set and test set to train neural network, obtains the material wear volume prediction model with high precision.The application is according to the friction and wear characteristics, and the wear volume of material is predicted using convolutional neural network.Compared with the prior art, the application makes full use of the wear state information contained in the friction data, has the characteristics of low data volume requirement and high precision.

Description

Technical Field

[0001] This invention relates to a testing method, and more particularly to a method for measuring the wear resistance of materials based on machine learning. Background Technology

[0002] Where there is industry, there is wear and tear. Losses due to wear-induced workpiece failure reach 950 billion yuan annually, accounting for 4.5% of GDP. To prevent equipment malfunctions or safety risks caused by workpiece wear failure, most workpieces have a predetermined service life. After reaching their service life, these workpieces are usually replaced by new ones, and the discarded workpieces lose their value. However, due to different service processes, the severity of wear on these discarded workpieces varies. This means that if the most severe wear condition is taken as the safe service life of a workpiece, a large number of still-usable workpieces will be discarded, resulting in waste. Therefore, to save costs, developing a method to determine the actual wear condition of worn parts based on the service process is of great significance.

[0003] The service life of a material can be reflected by changes in frictional force during wear. These changes involve variations in the surface condition, lubrication status, and energy dissipation of the material. By monitoring the frictional force, information about the material's wear state can be obtained. However, due to the complexity of friction and wear problems, no theoretical or empirical model currently exists that can establish a quantitative relationship between frictional force data and actual wear volume. This is mainly because the service life generates a large amount of frictional force data, making it difficult to establish such a quantitative relationship. Convolutional neural network (CNN) technology has significant advantages in processing large amounts of time-series data, and CNN models make it possible to establish this quantitative relationship.

[0004] Because the application of neural network models is limited by the dataset used for training, this model has high accuracy in predicting wear volume for the same type of material, but its generalization ability is limited. Application to dissimilar materials requires retraining the network with a new dataset. Summary of the Invention

[0005] Purpose of the invention: The present invention aims to provide a method for measuring the wear resistance of materials based on machine learning, which has high prediction accuracy and low data requirements.

[0006] Technical solution: The method for measuring the wear resistance of materials based on machine learning according to the present invention includes the following steps:

[0007] (1) Data set establishment: Collect frictional force data of materials under different temperature, load and friction speed conditions, and measure the wear volume of the materials; use the frictional force data as input value and the wear volume as output value to create a dataset;

[0008] (2) Data processing: The dataset is randomly divided into training and test sets;

[0009] (3) Model building: A convolutional neural network model was built based on the Keras module. The fully connected layer used a linear activation function. The accuracy of the model was measured by the mean square error between the actual value and the predicted value. The neural network was trained and the prediction accuracy of the network was improved by optimizing the hyperparameters to avoid overfitting of the network.

[0010] (4) Model verification: Input the friction data of the target worn part into the trained neural network model to obtain the predicted value of the wear volume of the worn material.

[0011] Preferably, the temperature control in step (1) is controlled by a feedback control system of actual temperature - set temperature, with the temperature probe within 3 cm of the sample; the friction speed is controlled by a continuously variable motor, and the friction speed is changed by changing the friction rotation speed and friction radius. The formula for calculating the friction speed is:

[0012]

[0013] Where n is the motor speed in revolutions per second, and r is the friction radius in meters;

[0014] The working load is controlled by applying weights of appropriate mass. During the wear process, the weights are steadily increased or decreased to complete the variable load experiment. The load application rod is mounted on the lever arm of the torque sensor, allowing the frictional force to be transmitted to the lever arm via the interaction force, thereby generating torque on the position of the torque sensor. The frictional force is calculated using this torque-lever arm system, making the frictional force test simple and accurate. The calculation formula is:

[0015]

[0016] Where P is torque, the unit is N·m. The value is the lever arm length, expressed in meters (m).

[0017] Preferably, the temperature in step (1) is -50 to 1000℃, the working load is 0.1 to 10000N, and the friction speed is 0.01 to 50m / s. The friction force data of the dataset used to train the neural network has the same acquisition time and frequency, the acquisition time is 2 to 300000s, the acquisition frequency is 0.01 to 100000Hz, and the acquired data is 50 to 10000 records used to train the neural network.

[0018] Preferably, the training set and test set in step (2) are divided in a ratio of 1:1 to 6:1. The division of the training set and test set should ensure that the accuracy of the neural network can increase during the iteration process without overfitting. Too few training sets will increase the training cost of the neural network, and too few test sets will reduce the reliability of the neural network model.

[0019] Preferably, the convolutional neural network model structure in step (3) is input value - convolutional layer - convolutional layer - convolutional layer - pooling layer - fully connected layer - output value. The hyperparameters that need to be optimized are: in the convolutional layers: the number of convolutional kernels is 4 to 1024, the kernel size is 3 to 1800, and the stride is 0 to 90; in the pooling layers: the filter size is 3 to 30. The hyperparameter optimization is related to the actual application conditions. For cases where the frictional force changes significantly, there are many data features, requiring a larger number of convolutional kernels to extract the data features. The corresponding convolutional kernel size needs to be larger to ensure complete feature extraction. For cases where the frictional force changes slightly, the number and size of convolutional kernels can be reduced, and a larger stride can be set, thereby reducing the cost of neural network training.

[0020] Preferably, the formula for calculating the mean square error, which measures the accuracy of the model, in step (3) is:

[0021]

[0022] Among them, y p y is the neural network prediction value. e To obtain the true value in the experiment.

[0023] Preferably, the neural network model for the frictional force input of the worn parts in step (4) is trained using a dataset collected from the same material.

[0024] Preferably, the goal of the convolutional layer is to extract features from time-series data. Different convolutional layers extract features at different scales. The first convolutional layer extracts features from a long range of data, the second convolutional layer extracts features from a smaller range of data, and the third convolutional layer extracts features from a very small range of data.

[0025] Preferably, different hyperparameters are used in the neural network model to improve prediction accuracy based on different material properties. Pooling layers further reduce data complexity and lower neural network training costs.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) It has high prediction accuracy. The present invention collects friction data of the worn material and trains a neural network to predict the actual wear volume of the material. The minimum mean square error between the actual value and the predicted value is only 3.6×10 -7(2) The neural network structure used in this invention is simple and effective, and a neural network model with high prediction accuracy can be obtained by training with a small dataset. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the implementation of the method described in this invention;

[0028] Figure 2 This is a diagram of a friction and wear test setup;

[0029] Figure 3 This is a schematic diagram of the neural network model structure in Example 1;

[0030] Figure 4 This is a schematic diagram of the prediction results in Example 1. Detailed Implementation

[0031] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0032] Example 1

[0033] The flowchart of the machine learning-based method for measuring material wear resistance is as follows: Figure 1 As shown, it includes the following steps:

[0034] (1) Collect frictional force data during the material wear process. The working temperature is 25℃, the working load is constant at 10N, the frictional speed is 0.04m / s, the sampling frequency is 1Hz, and the sampling time is 1799s. A total of 1799 frictional force data points were collected and used as input data for the neural network. The wear volume of the material was measured as the output prediction value. A total of 80 sets of data were collected to complete the dataset construction. Among them, the temperature control adopts the feedback control system of actual temperature-set temperature. The temperature probe is within 3 cm of the sample to ensure the accuracy and stability of the temperature test. The frictional speed is controlled by a continuously variable motor. The frictional speed is changed by changing the frictional speed and frictional radius. The formula for calculating the frictional speed is:

[0035]

[0036] Where n is the motor speed in revolutions per second (rpm) and r is the friction radius in meters (m). The working load is controlled by applying weights of appropriate mass; the variable load experiment is completed by steadily increasing or decreasing the weights during the wear process. The load application rod is mounted on the lever arm of the torque sensor, allowing the frictional force to be transmitted to the lever arm via the interaction force, thereby generating torque on the position of the torque sensor. The frictional force is calculated using this torque-lever arm system, making the frictional force test simple and accurate. The calculation formula is:

[0037]

[0038] Where P is torque, in N·m, and l is lever arm length, in m.

[0039] (2) Randomly divide the dataset into training set and test set in a ratio of 4:1.

[0040] (3) Convolutional Neural Network Model: The construction order is Input layer - convolutional layer - convolutional layer - convolutional layer - pooling layer (max pooling layer) - fully connected layer - output value. A linear activation function is used between the fully connected layer and the output value. The accuracy of the model prediction is judged by the mean squared error. Based on the magnitude of the mean squared error, the hyperparameters are adjusted and optimized. The optimized hyperparameters are as follows: Convolutional layer 1: 8 kernels, kernel size 60, stride 0; Convolutional layer 2: 16 kernels, kernel size 10, stride 0; Convolutional layer 3: 32 kernels, kernel size 3, stride 0; Pooling layer: filter size 3, stride 0. The neural network structure is as follows. Figure 3 As shown. After 50 iterations, the neural network training is complete.

[0041] (4) Input the frictional force data of the target worn part into the trained neural network model to obtain the predicted value of the wear volume of the worn material. The mean square error between the actual value and the predicted value is calculated to be 3.6 × 10⁻⁶. -7 , Figure 4 This is a schematic diagram of the prediction results. The model has high prediction accuracy.

[0042] Example 2

[0043] The method for measuring the wear resistance of materials based on machine learning includes the following steps:

[0044] (1) Collect frictional force data during the material wear process. The working temperature is 300℃, the working load is constant at 10N, the frictional speed is 0.04m / s, the sampling frequency is 2Hz, and the sampling time is 3600s. A total of 7200 frictional force data points were collected and used as input data for the neural network. The wear volume of the material was measured and used as the output prediction value. A total of 200 sets of data were collected to complete the dataset construction.

[0045] (2) Randomly divide the dataset into training set and test set in a 3:1 ratio.

[0046] (3) Convolutional neural network model was constructed in the following order: Input layer - convolutional layer - convolutional layer - convolutional layer - pooling layer (max pooling layer) - fully connected layer - output value. A linear activation function was used between the fully connected layer and the output value. The accuracy of the model prediction was judged by the mean squared error. Based on the magnitude of the mean squared error, the hyperparameters were adjusted and optimized. The optimized hyperparameters were as follows: Convolutional layer 1: 16 kernels, kernel size 30, stride 10; Convolutional layer 2: 32 kernels, kernel size 10, stride 3; Convolutional layer 3: 32 kernels, kernel size 5, stride 0; Pooling layer: filter size 3, stride 0. The neural network training was completed after 130 iterations.

[0047] (4) Input the frictional force data of the target worn part into the trained neural network model to obtain the predicted value of the wear volume of the worn material. The mean square error between the actual value and the predicted value is calculated to be 5.3 × 10⁻⁶. -6 This model can accurately predict the actual wear volume of materials at high temperatures.

[0048] Example 3

[0049] The method of measuring the wear resistance of materials based on machine learning includes the following steps:

[0050] (1) Collect frictional force data during the material wear process. The working temperature is 25℃, the working load is 10N and 20N alternating at a frequency of 0.1Hz, the frictional speed is 0.1m / s, the sampling frequency is 20Hz, the sampling time is 1800s, and a total of 36,000 frictional force data points are collected as input data for the neural network. Measure the material wear volume as the output prediction value. Collect a total of 200 sets of data to complete the dataset construction.

[0051] (2) Randomly divide the dataset into training set and test set in a 3:1 ratio.

[0052] (3) Convolutional Neural Network (CNN) model was constructed in the following order: Input layer - convolutional layer - convolutional layer - convolutional layer - pooling layer (max pooling layer) - fully connected layer - output value. A linear activation function was used between the fully connected layer and the output value. The accuracy of the model prediction was judged by the mean squared error. Based on the magnitude of the mean squared error, the hyperparameters were adjusted and optimized. The optimized hyperparameters were as follows: Convolutional layer 1: 32 kernels, kernel size 120, stride 30; Convolutional layer 2: 64 kernels, kernel size 30, stride 10; Convolutional layer 3: 128 kernels, kernel size 10, stride 0; Pooling layer: filter size 5, stride 0. The neural network training was completed after 300 iterations.

[0053] (4) Input the frictional force data of the target worn part into the trained neural network model to obtain the predicted value of the wear volume of the worn material. The mean square error between the actual value and the predicted value is calculated to be 7.2 × 10⁻⁶. -6 This model can predict the actual wear volume of materials under varying loads.

Claims

1. A method for measuring the wear resistance of materials based on machine learning, characterized in that: Includes the following steps: (1) Data set establishment: Collect friction force data of the same wear material under different temperature, load and friction speed conditions, and measure the wear volume of the wear material; use the friction force data as input value and the wear volume as output value to make a dataset; The temperature is controlled by a feedback control system that modulates the actual temperature versus the set temperature, with the temperature probe within 3 cm of the sample. The friction speed is controlled by a continuously variable motor; the friction speed is changed by altering the friction rotation speed and the friction radius. The formula for calculating the friction speed is: ； in, The motor speed is given in revolutions per second (rpm), and r is the friction radius, measured in meters (m). The load is controlled by applying weights of appropriate mass. During the wear process, the weights are steadily increased or decreased to complete the variable load experiment. The load application rod is mounted on the lever arm of the torque sensor. The frictional force is calculated using this torque-lever arm system, and the calculation formula is as follows: ； Where P is torque, and the unit is N·m. The length of the lever arm is in meters (m). (2) Data processing: The dataset is randomly divided into training and test sets; (3) Model building: A convolutional neural network model was built based on the Keras module. The fully connected layer used a linear activation function. The accuracy of the model was measured by the mean square error between the actual value and the predicted value. The neural network was trained and the prediction accuracy of the network was improved by optimizing the hyperparameters to avoid overfitting of the network. The structure of the convolutional neural network model is: input value - convolutional layer - convolutional layer - convolutional layer - pooling layer - fully connected layer - output value. The hyperparameters that need to be optimized are: in the convolutional layer: the number of convolutional kernels is 4 to 1024, the kernel size is 3 to 1800, and the stride is 0 to 90; in the pooling layer: the filter size is 3 to 30. (4) Model verification: Input the friction data of the target material to be worn into the trained neural network model to obtain the predicted value of the wear volume of the material.

2. The method for measuring the wear resistance of materials based on machine learning according to claim 1, characterized in that: The temperature mentioned in step (1) is -50 to 1000℃, the working load is 0.1 to 10000N, and the friction speed is 0.01 to 50m / s. The friction force data of the dataset used to train the neural network has the same acquisition time and frequency. The acquisition time is 2 to 300000s, the acquisition frequency is 0.01 to 100000Hz, and the acquired data is 50 to 10000 records used to train the neural network.

3. The method for measuring the wear resistance of materials based on machine learning according to claim 1, characterized in that: The training set and test set mentioned in step (2) are divided into a ratio of 1:1 to 6:

1.

4. The method for measuring the wear resistance of materials based on machine learning according to claim 1, characterized in that: Depending on the properties of the material being worn, different hyperparameters are used in the neural network model to improve prediction accuracy.

Images