Bearing heat treatment parameter influence degree analysis method based on machine learning

Abstract

The invention discloses a bearing heat treatment parameter influence degree analysis method based on machine learning. Firstly, according to a bearing heat treatment detection device used in the actual process production process, parameter data detected in real time are selected as sample features, then importance of the sample features is judged through LightGBM, the influence degree of each feature on the performance of a bearing after heat treatment is recorded, and the performance of the bearing after heat treatment is determined. And finally, training a decision tree regression device according to the influence degree of each feature. Different decision tree regression devices can be trained according to different bearing types and different heat treatment modes, so that the prediction accuracy of the model is improved. In the detection process, if the deviation between a predicted value and an actual value is large, early warning is carried out, so that the loss of a detection sample is reduced, or further damage to equipment is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of bearing heat treatment, and specifically relates to a method for analyzing the influence degree of bearing heat treatment parameters based on machine learning. Background Art

[0002] As a key rotating component in a mechanical system, the performance and service life of a bearing largely depend on the heat treatment process during manufacturing. Heat treatment is an important means to improve the material properties of bearings. It enhances their hardness, toughness, and wear resistance by changing the microstructure of the material. However, the heat treatment of bearings is a complex process involving multiple parameters such as heating temperature, holding time, cooling rate, and atmosphere composition. Minor changes in these parameters can have a significant impact on the performance of the bearings. Traditional bearing heat treatment processes mainly rely on experience and trial-and-error methods, which are inefficient and costly. With the rise of Industry 4.0 and intelligent manufacturing, higher requirements are put forward for the precise control and optimization of the bearing heat treatment process. To improve production efficiency and product quality, there is an urgent need for a method that can accurately analyze and predict the influence of heat treatment parameters on bearing performance.

[0003] CN202210493233.4 proposes an air quality prediction method that couples a GA-BP neural network and a decision tree, opening up a new way for regional air quality prediction and air quality index level classification prediction. At the same time, it can provide guidance and reference for regions to take air pollution prevention measures. CN202111382959.2 discloses a software defect prediction method based on clustering analysis and decision tree algorithm, which reduces the software repair time while maintaining its automation level as much as possible; can classify a large number of software samples more accurately; and determine the distribution of defects according to the pedigree from the software evolution process. CN202111382959.2 discloses a method for evaluating the reliability of the location of modern marine pastures based on a decision tree model, improving the recognition accuracy, rationality, and reliability of the subsequent decision tree model for the location of marine pastures. The present invention is applied to the field of geographic information data processing technology.

[0004] In the existing technical solutions, although there are many cases applying decision tree algorithms, there are few applications in the aspect of bearing heat treatment processes, especially the analysis of the influence degree of heat treatment parameters. This results in the fact that in the bearing forging process, researchers cannot rely well on accurate data to forge bearings. Summary of the Invention

[0005] The present invention proposes a method for analyzing the influence degree of bearing heat treatment parameters based on machine learning. The purpose is to screen out unqualified bearings during the actual heat treatment process if the predicted value deviates greatly from the actual detection result, so as to reduce the unqualified rate of bearing detection or reduce further damage to the equipment. This method uses LightGBM to evaluate the importance degree of sample features and records the features that affect the bearing heat treatment process as training data. At the same time, different decision tree regression models are trained according to different types of bearings to ensure the detection results of different types of bearings.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for analyzing the influence degree of bearing heat treatment parameters based on machine learning, comprising the following steps:

[0008] S1. Conduct heat treatment operations on the bearings. After the heat treatment of the bearings, select the preset parameter values of the bearing heat treatment device with qualified detection results as sample features; and select sample targets.

[0009] S2. Remove outliers and standardize the sample data.

[0010] S3. Select appropriate LightGBM model hyperparameters.

[0011] S4. Record the features with a higher influence degree on the heat treatment process in the LightGBM model.

[0012] S5. Divide the training set and the test set in the sample dataset.

[0013] S6. Select appropriate decision tree model hyperparameters.

[0014] S7. Train a decision tree regressor on the standard bearings using the retrieval heat treatment method according to the recorded features.

[0015] S8. Train a decision tree regressor on the modified bearings using the hybrid heat treatment method according to the recorded features.

[0016] S9. Train a decision tree regressor on the special bearings according to the recorded features and the data in the sample dataset.

[0017] S10. Analyze the influence degree of different heat treatment methods on the heat treatment parameters for different types of bearings respectively.

[0018] Further, in step S1,

[0019] The preset parameters of the bearing heat treatment device are: carbon fitness value, nitrogen value, methanol value, propane value, cleaning area pH value, cleaning area temperature, manual placement layer number, quenching temperature, quenching mesh belt running frequency, tempering temperature, and tempering mesh belt running frequency. After the heat treatment operation on the bearing, record each parameter value, and use a detection device to detect the heat-treated bearing. Select the heat treatment parameter values of the bearings with qualified detection results as sample features.

[0020] Among them, when detecting the heat-treated bearing, it is necessary to detect the deformation detection value, hardness detection value, tissue detection value, retained austenite detection value, ferrite detection value, and martensite detection value of the bearing, and use these detection results as sample targets.

[0021] Furthermore, step S2 is specifically as follows:

[0022] Remove the outliers in the sample, and use the Z-Score standardization method to standardize the data in the sample. The processing formula is: where μ is the mean of the sample data and σ is the standard deviation of the sample data. After the data is standardized, it is beneficial to improve the model convergence speed, improve the model accuracy, and simplify the calculation;

[0023] Furthermore, step S3 is specifically as follows:

[0024] Select hyperparameters, set the maximum depth of the tree (max_depth) to 5 - 10, the number of nodes (num_leaves) to less than 2 max_depth , and set the learning rate to 0.01. According to the mean squared error: to fine-tune and evaluate the model after each hyperparameter adjustment, and determine the appropriate hyperparameters.

[0025] Furthermore, step S4 is specifically as follows:

[0026] After step S3, use all the sample data to fit LightGBM;

[0027] LightGBM uses decision trees for modeling, and its core is to minimize the loss function. For each tree, the model is updated through the following formula: First is the residual calculation: where, is the residual at the t-th iteration, y i is the true value, is the model prediction value at the (t - 1)-th round. Secondly is the split point selection. By traversing all features, find the best split point to minimize the loss of the objective function. Assume the partitioning feature is x j , the split point is s, and the dataset is divided into L and R after splitting: where and is the mean of the left and right branches. After that, feature importance analysis is carried out. LightGBM calculates the importance of features through the frequency and gain of feature splitting. Whenever a feature is used for splitting, the gain it contributes is accumulated. Feature x j The importance formula of is: Importance(x j ) = ∑splitonx j Gain(x j ), where Gain represents the information gain brought by each split of this feature.

[0028] According to the feature importance scores output by LightGBM, determine the influence order of heat treatment parameters on bearing performance, record each feature importance score, and use these features as the sample features for training the decision tree regressor next.

[0029] Furthermore, step S5 is specifically as follows:

[0030] Adopt a random division method to ensure uniform data distribution, and divide the data set into a training set and a test set. The training set accounts for 70%, and the test set accounts for 30%.

[0031] Furthermore, step S6 is specifically as follows:

[0032] Select hyperparameters, start trying from the 3rd layer of the shallower tree, gradually increase the depth, and observe the performance of the model. The default value of the minimum number of samples for node splitting (min_samples_split) is set to 2. Add a regularization term to the loss function to control the model complexity and prevent overfitting, where L is the loss function, λ is the regularization strength, set to 0.01, and θ i is the model parameter. According to R 2 score: to fine-tune and evaluate the model after each hyperparameter adjustment, and determine the appropriate hyperparameters.

[0033] Furthermore, step S7 is specifically as follows:

[0034] S71. Separate the features and the target variable. Use the detection results after bearing heat treatment as the target variable, and the recorded features as the sample features;

[0035] S72. Fit the decision tree regressor model using the training set data;

[0036] S73. Make predictions on the test set;

[0037] S74. Adjust the model hyperparameters according to the evaluation results.

[0038] Use the K-fold cross-validation method to optimize the hyperparameters. The average performance metric for cross-validation is: Among them is the i-th validation set is the remaining K - 1 training sets after removing D i ; represents the performance metric calculated on denotes the hyperparameter θ obtained by training with and calculated on .

[0039] S75. The accuracy of the model is evaluated using the MSE method. The evaluation formula is: where y i is the predicted value of the model (M cv (θ) in S74), is the actual target value, and n is the total number of samples;

[0040] Compared with the prior art, the beneficial technical effects of the present invention are:

[0041] 1. The method for analyzing the influence degree of bearing heat treatment parameters based on machine learning proposed by the present invention successfully applies the machine learning method to the field of bearing heat treatment. It realizes that during the actual detection process of the equipment, if the deviation between the predicted value and the actual detection result is large, early warnings for equipment failures or unqualified bearing quality can be achieved, thereby reducing the loss of detection samples or further damage to the equipment, etc.

[0042] 2. The present invention introduces a feature importance extraction method. After fitting with LightGBM, the importance scores of the features are output, and the importance of each feature is recorded, which is convenient for subsequent training using decision tree regression, making the training results more accurate and providing an effective basis for relevant researchers to analyze the influence degree of bearing heat treatment parameters BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is the flowchart of the method of the embodiment of the present invention

[0044] Figure 2 is the schematic diagram of the construction of the decision tree regressor model of the present invention DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] As Figure 1-2 shown, a method for analyzing the influence degree of bearing heat treatment parameters based on machine learning includes the following steps:

[0046] S1. Perform heat treatment operations on the bearings. After the bearing heat treatment, select the preset parameter values of the bearing heat treatment device with qualified detection results as sample features; and select sample targets;

[0047] The preset parameters of the bearing heat treatment device are: carbon fitness value, nitrogen value, methanol value, propane value, cleaning area pH value, cleaning area temperature, manual placement layer number, quenching temperature, quenching mesh belt running frequency, tempering temperature, and tempering mesh belt running frequency. After the heat treatment operation on the bearing, record each parameter value, and use a detection device to detect the heat-treated bearing. Select the heat treatment parameter values of the bearings with qualified detection results as sample features.

[0048] Among them, when detecting the heat-treated bearing, it is necessary to detect the deformation detection value, hardness detection value, tissue detection value, retained austenite detection value, ferrite detection value, and martensite detection value of the bearing, and use these detection results as sample targets.

[0049] S2. Remove outliers and standardize the sample data;

[0050] Remove the outliers in the sample, and use the Z-Score standardization method to standardize the data in the sample. The processing formula is: where μ is the mean of the sample data and σ is the standard deviation of the sample data. After the data is standardized, it is beneficial to improve the model convergence speed, improve the model accuracy, and simplify the calculation;

[0051] S3. Select appropriate LightGBM model hyperparameters;

[0052] Select hyperparameters, set the maximum depth of the tree (max_depth) to 5 - 10, the number of nodes (num_leaves) to less than 2 max_depth , and set the learning rate to 0.01. According to the mean squared error: to fine-tune and evaluate the model after each hyperparameter adjustment, and determine the appropriate hyperparameters.

[0053] S4. Record the features in the LightGBM model that have a higher impact on the heat treatment process;

[0054] After step S3, use all the sample data to fit LightGBM;

[0055] LightGBM uses decision trees for modeling, and its core is to minimize the loss function. For each tree, the model is updated through the following formula: First is the residual calculation: where, is the residual at the t-th iteration, y i is the true value, is the predicted value of the model at the (t - 1)-th round. Secondly, the split point selection is to find the best split point by traversing all features to minimize the loss of the objective function. Assume the partition feature is x j , the split point is s, and the data set is divided into L and R after splitting: Among them and are the means of the left and right branches. After that, feature importance analysis is carried out. LightGBM calculates the importance of features through the frequency and gain of feature splitting. Whenever a feature is used for splitting, the gain it contributes will be accumulated. Feature x j The importance formula of is: Importance(x j ) = ∑splitonx j Gain(x j ), where Gain represents the information gain brought by each split of this feature.

[0056] According to the feature importance scores output by LightGBM, determine the influence order of heat treatment parameters on bearing performance, record each feature importance score, and use these features as the sample features for training the decision tree regressor next.

[0057] S5. Divide the training set and test set in the sample dataset

[0058] Adopt the random division method to ensure uniform data distribution, divide the dataset into the training set and test set, with the training set accounting for 70% and the test set accounting for 30%.

[0059] S6. Select appropriate hyperparameters for the decision tree model;

[0060] Select hyperparameters, start trying from the 3rd layer of the shallower tree, gradually increase the depth, and observe the performance of the model. The default value of the minimum number of samples for node splitting (min_samples_split) is set to 2, add a regularization term to the loss function to control the model complexity and prevent overfitting, where L is the loss function, λ is the regularization strength, set to 0.01, and θ i is the model parameter. According to R 2 score: to fine-tune and evaluate the model after each hyperparameter adjustment, and determine the appropriate hyperparameters.

[0061] S7. Train a decision tree regressor on the standard bearing using the retrieval-based heat treatment method according to the recorded features;

[0062] S71. Separate the features and target variables. Take the detection results after bearing heat treatment as the target variable, and the recorded features as the sample features;

[0063] S72. Use the training set data to fit the decision tree regressor model;

[0064] S73. Make predictions on the test set;

[0065] S74. Adjust the model hyperparameters according to the evaluation results.

[0066] Use the K-fold cross-validation method to optimize the hyperparameters. The average performance metric for cross-validation is: where is the i-th validation set, is the remaining K - 1 training sets after removing D i , represents the hyperparameter θ trained with , and the performance metric calculated on .

[0067] S75. Use the MSE method to evaluate the accuracy of the model. The evaluation formula is: where y i is the predicted value of the model (M cv (θ) in S74), is the actual target value, and n is the total number of samples;

[0068] S8. Train a decision tree regressor on the modified bearings using the hybrid heat treatment method according to the recorded features;

[0069] S9. Train a decision tree regressor on the special bearings according to the recorded features and the data in the sample dataset;

[0070] S10. Analyze the influence degree of different heat treatment methods on the heat treatment parameters for different types of bearings respectively.

[0071] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention, and all of them are within the protection scope of this technology.

Claims

1. A method for analyzing the influence of bearing heat treatment parameters based on machine learning, characterized in that: The steps include: S1. Perform heat treatment on the bearing. After the heat treatment, select the preset parameter value of the bearing heat treatment device with qualified test results as the sample feature; and select the sample target; S2, remove outliers and standardize sample data; S3. Select appropriate LightGBM model hyperparameters; S4, record the features in the LightGBM model that have a high degree of influence on the heat treatment process; S5, dividing the sample data set into a training set and a test set; S6. Select appropriate decision tree model hyperparameters; S7, training a decision tree regressor on a standard bearing using a retrieval-based heat treatment method based on the recorded features; S8. training a decision tree regressor on a modified bearing using a hybrid heat treatment method based on the recorded features; S9, training a decision tree regressor on a dedicated bearing based on the recorded features and the data of the sample data set; S10. Analyze the influence of different heat treatment parameters on different types of bearings using different heat treatment methods.

Images