A Machine Learning-Based Model for Identifying Bearing Electrical Breakdown Damage
The bearing electrical breakdown damage identification model established by machine learning methods solves the problems of low efficiency and difficulty in quantification in traditional methods, realizes efficient identification and analysis of bearing damage morphology, and provides data-driven analysis support.
Patent Information
- Application Number
- CN202611518610.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-29
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies are insufficient for effectively identifying and analyzing bearing electrical breakdown damage, especially in complex friction and wear processes. Traditional methods are inefficient, subjective, difficult to quantify, and struggle to characterize nonlinear relationships.
A bearing electrical breakdown damage identification model was established using machine learning methods. By combining support vector machine, BP neural network and random forest models with data preprocessing and multi-decision tree ensemble learning, the bearing damage morphology was identified and the correlation between operating parameters and damage was explored.
It enables efficient identification and analysis of bearing electrical breakdown damage, accurately classifies damage morphology under different working conditions, provides data-driven analysis methods, and improves the accuracy and stability of identification and evaluation.
Smart Images

Figure CN123068802A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bearing electrical breakdown damage identification technology, specifically a bearing electrical breakdown damage identification model and method based on machine learning. Background Technology
[0002] With the rapid development of new energy vehicles, high-speed railways, aerospace, and intelligent manufacturing, electric drive equipment, represented by permanent magnet synchronous motors and variable frequency drive systems, is constantly evolving towards higher power density, higher speed, and greater intelligence. Rolling bearings, as one of the most critical basic components in rotating machinery, directly affect the service life and operational reliability of the entire equipment. Traditional bearing damage morphology assessment mainly relies on characterization techniques such as optical microscopy and scanning electron microscopy (SEM), combined with researchers' experience to analyze wear mechanisms, damage types, and failure degrees. While this method can accurately reflect the microscopic damage characteristics of material surfaces, it suffers from low analytical efficiency, strong subjectivity, difficulty in quantitative evaluation, and insufficient repeatability. Subsequently, with the development of digital image processing technology, researchers began to use methods such as texture analysis, edge detection, grayscale statistics, and morphological analysis to extract damage image features, achieving semi-automatic identification of damaged areas. However, these methods typically rely on manually designed features, limiting their adaptability to multi-scale, multi-type coupled damage morphologies and making it difficult to fully characterize the nonlinear characteristic relationships formed during complex friction and wear processes. Therefore, it is extremely urgent to design a machine learning-based bearing electrical breakdown damage identification model. Summary of the Invention
[0003] The purpose of this invention is to provide a machine learning-based model for identifying bearing electrical breakdown damage. This model can fully characterize the nonlinear characteristic relationships formed during complex friction and wear processes, and further explore the correlation between operating parameters and damage morphology, providing data-driven methodological support for the identification and evolution analysis of bearing electrical breakdown damage.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: A machine learning-based bearing electrical breakdown damage identification model method includes the following: Establish database During the data processing, the data obtained in the experiment were first screened, organized and unified. Data from different sources were classified according to the same variable definition and data format, and some abnormal data were checked to ensure the validity and consistency of the data used in subsequent modeling. Experimental data preprocessing To eliminate the influence of differences in units and value ranges between different input variables, the input features are normalized before model training. The min-max normalization method is used to linearly map each input variable to the [0,1] interval; the calculation formula is as follows:
[0005] In the formula The original input data, , To be the minimum value, This is the normalized data.
[0006] Support Vector Machine Classification Model Establishment Support Vector Machines (SVMs) can map raw input data to a high-dimensional feature space through kernel functions, thereby establishing a non-linear mapping relationship between input features and different damage morphology categories; the kernel function formula is:
[0007] in, and For the standardized sample, For kernel function parameters, This represents the Euclidean distance. Compared with traditional classification methods, support vector machines obtain the optimal decision boundary by maximizing the classification margin, and have better adaptability to small samples, nonlinear and high-dimensional data. They are suitable for bearing damage morphology classification and recognition under limited sample conditions. Establishment of BP neural network classification model Backpropagation (BP) neural networks can learn complex correspondences between input features and output categories from training samples, exhibiting strong nonlinear mapping capabilities. Among the structural parameters of a BP neural network, the number of hidden layer neurons is a crucial parameter affecting the model's classification performance. The formula is: P = a + b + c; where P is the number of hidden layer neurons, a and b are the number of neurons in the input and output layers, respectively, and c is a positive integer adjustment constant from 1 to 10. Random Forest Model Establishment Multiple sub-training sets were randomly generated from the original training set using the Bootstrap resampling method, and decision tree models were built for each sub-set. The classification results were finally determined by majority voting. The load P, rotational speed R, current I, and voltage amplitude V during bearing operation were used as input variables for the random forest model, and different damage morphology levels were used as output labels. The input features were then divided level by level using multiple decision trees.
[0008] Furthermore, in the Support Vector Machine (SVM) classification model building module, another key parameter in the SVM model is the penalty factor C. The penalty factor C controls the degree of punishment the model inflicts on misclassified samples. When C is small, the model has a higher tolerance for classification errors, resulting in a relatively large classification margin, which reduces model complexity and improves generalization ability. When C is large, the model can reduce training error, but the classification margin decreases accordingly, increasing model complexity. When the penalty factor is 10... -1 At that time, the classification accuracy reached its highest level, at 78.26%.
[0009] Furthermore, in the BP neural network classification model building module, when the number of hidden layer neurons is too small, the network's ability to express bearing damage features is insufficient, which easily leads to underfitting; when the number of hidden layer neurons is too large, the network structure complexity increases, which easily leads to overlearning of local features in the training samples, thereby causing overfitting and reducing the model's ability to predict unknown samples.
[0010] Furthermore, the number of hidden layer neurons needs to be initially estimated based on the size of the input and output layers, and further determined through validation sets or cross-validation. Since the input layer has 4 nodes and the output layer has 3 nodes, the search range for hidden layer neurons is set to 3-14, and the network structure corresponding to the highest classification accuracy is used as the final model.
[0011] Furthermore, to obtain the best model prediction results, a candidate range for the number of hidden layer neurons was determined based on empirical formulas, and the number of hidden neurons was changed sequentially within this range. To reduce the impact of random initial weights on the training results, each network structure was trained multiple times. Five-fold cross-validation was used to calculate the average classification accuracy and average loss. Finally, the optimal number of hidden neurons was selected as the one with the highest classification accuracy on the validation set, the smallest fluctuation in the results, and the relatively simple network structure.
[0012] Furthermore, in the random forest model building module, the load P, rotational speed R, current I, and voltage amplitude V during the bearing operation process are used as input variables of the random forest model, and different damage morphology levels are used as output labels. The input features are then divided level by level through multiple decision trees.
[0013] Furthermore, in the random forest model, the number of decision trees, Ntree, is a key parameter affecting classification performance. Random forests complete classification through multiple decision trees, and the number of trees directly affects the stability and generalization ability of the model. When Ntree is small, the forest contains insufficient information, and the model is easily affected by the randomness of a single decision tree, resulting in large fluctuations in classification results. When Ntree increases, the model performance usually gradually improves, but an excessively large number of trees will increase computational complexity, while the improvement in classification accuracy tends to saturate.
[0014] The beneficial effects of this invention are as follows: This invention provides a machine learning-based bearing electrical breakdown damage identification model method. Taking the surface damage morphology of the bearing inner raceway as the main research object, it constructs a controllable experimental environment under different loads P, speeds R, currents I, and voltage amplitudes V to obtain bearing surface damage morphology data under different service conditions. The damage of different degrees is then classified into levels, thereby obtaining good prediction results. Simultaneously, considering that the bearing damage morphology under alternating current is affected by multiple factors such as load, speed, and current, the relationship between different operating conditions and damage characteristics is not a simple linear one, and it is difficult to accurately describe its variation law using a single parameter. This method can classify, identify, and model the bearing damage morphology under different operating conditions, and further explore the correlation between operating parameters and damage morphology, providing data-driven methodological support for the identification and evolution analysis of bearing electrical breakdown damage. Attached Figure Description
[0015] Figure 1 This is a table showing the bearing damage morphology levels under alternating current in an embodiment of the present invention; Figure 2 This is a diagram of the SVM neural network topology in the support vector machine classification model building module of this invention; Figure 3 This diagram illustrates the accuracy under different penalty factors in the support vector machine classification model building module of this invention. Figure 4 This is a diagram of the BP neural network topology in the BP neural network classification model building module of the present invention; Figure 5 The accuracy of the BP neural network classification model building module of this invention under different numbers of hidden neurons; Figure 6 This is a diagram of the random forest network topology in the random forest model building module of this invention; Figure 7 The accuracy is measured by the number of decision trees in the random forest model building module of this invention. Detailed Implementation
[0016] To make the technical problems, technical solutions, and beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. It should be noted that when an element is referred to as being "fixed to" or "disposed on" another element, it can be directly on or indirectly on that other element. When an element is referred to as being "connected to" another element, it can be directly connected to or indirectly connected to that other element. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the present invention, "a plurality of" means two or more unless otherwise explicitly specified. "Several" means one or more unless otherwise explicitly specified. In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0017] Specific Embodiment 1: As per the appendix to the specification of this invention. Figure 1 To the instruction manual Figure 7 As shown, the bearing electrical breakdown damage identification model method based on machine learning provided by the present invention includes the following: Database establishment: During the data processing, the data obtained from independent experiments and literature were first screened, organized, and unified, as shown in the appendix to the instruction manual. Figure 1As shown, data from different sources were categorized according to the same variable definitions and data formats, and some outliers were checked to ensure the validity and consistency of the data used in subsequent modeling. Based on this, and considering data completeness and operating condition coverage, 114 representative samples were selected to construct a bearing damage morphology recognition dataset. Compared to relying solely on self-experimental data, introducing literature data can broaden the range of operating conditions for the samples to a certain extent, increasing the coverage of bearing damage morphology features under different alternating current conditions, thus providing a more sufficient data foundation for machine learning models. Considering the relatively limited number of samples, if a single random partitioning method is directly used for model training and evaluation, the model performance may be affected by the sample partitioning results, leading to fluctuations in the evaluation results. Therefore, a 5-fold cross-validation method is used during model training to improve the stability of the model evaluation results. Simultaneously, the training and validation sets are divided in the overall dataset at a ratio of 80% and 20%. This approach can fully utilize the limited samples while reducing the possibility of overfitting due to insufficient training samples, and further examine the model's ability to classify and recognize different bearing damage morphologies.
[0018] Experimental data preprocessing: To eliminate the influence of differences in units and value ranges among different input variables, the input features are normalized before model training. This paper uses the min-max normalization method to linearly map each input variable to the [0,1] interval, and the calculation formula is as follows:
[0019] In the formula The original input data, , To be the minimum value, This is the normalized data. For the test samples, the normalized parameters calculated from the training set are used for same-scale transformation to avoid data leakage and improve the reliability of model prediction results.
[0020] Support Vector Machine (SVM) Classification Model Establishment: For the bearing damage morphology classification problem, the SVM can map the original input data to a high-dimensional feature space through a kernel function, thereby establishing a nonlinear mapping relationship between the input features and different damage morphology categories. The kernel function formula is as follows:
[0021] in, and For the standardized sample, For kernel function parameters, This represents the Euclidean distance. Support Vector Machines (SVMs) obtain the optimal decision boundary by maximizing the classification margin, exhibiting good adaptability to small samples, nonlinearity, and high-dimensional data. Therefore, they are suitable for bearing damage morphology classification and identification under finite sample conditions. Another key parameter in the SVM model is the penalty factor C. This parameter mainly controls the degree of penalty imposed on misclassified samples, and its value directly affects the size of the classification hyperplane margin and the model's generalization ability. When C is small, the model has a higher tolerance for classification errors, resulting in a relatively large classification margin, which can reduce model complexity and improve generalization ability. When C is large, the model places more emphasis on the correct classification of training samples, reducing training error, but the classification margin decreases accordingly, increasing model complexity. By comparison, when the penalty factor is 10... -1 At that time, the classification accuracy reached its highest level, at 78.26%.
[0022] Establishment of a BP Neural Network Classification Model: In the structural parameters of a BP neural network, the number of hidden layer neurons is a crucial parameter affecting the model's classification performance. The formula is: P = a + b + c; where P is the number of hidden layer neurons, a and b are the number of neurons in the input and output layers, respectively, and c is a positive integer adjustment constant from 1 to 10. If the number of hidden layer neurons is too small, the network's ability to express bearing damage features is insufficient, easily leading to underfitting; if the number of hidden layer neurons is too large, the network structure complexity increases, easily leading to overlearning of local features in the training samples, thus causing overfitting and reducing the model's predictive ability for unknown samples. Therefore, the number of hidden layer neurons needs to be initially estimated based on the size of the input and output layers, and further determined through validation sets or cross-validation. Since the input layer has 4 nodes and the output layer has 3 nodes, the search range for hidden layer neurons is set to 3-14, with the network structure corresponding to the highest classification accuracy being used as the final model. To obtain the best model prediction results, the candidate range for the number of hidden layer neurons is first determined based on empirical formulas, and the number of hidden neurons is changed sequentially within this range. To minimize the impact of random initial weights on training results, each network architecture was trained multiple times, and five-fold cross-validation was used to calculate the average classification accuracy and average loss. The optimal number of hidden neurons was selected based on the highest validation set classification accuracy, minimal fluctuation in results, and relatively simple network structure. (See attached manual.) Figure 1 As shown, the classification accuracy reaches its highest level of 95.65% when the number of hidden neurons is 8.
[0023] Random Forest Model Construction: Random forests use the Bootstrap resampling method to randomly generate multiple sub-training sets from the original training set, and build decision tree models for each. The classification result is then determined by majority voting. Compared to a single decision tree model, random forests reduce the risk of overfitting to the training data and improve the model's generalization ability and classification stability. For the bearing damage morphology classification problem, the load P, rotational speed R, current I, and voltage amplitude V during bearing operation can be used as input variables for the random forest model, with different damage morphology levels as output labels. Multiple decision trees are used to progressively divide the input features. In the random forest model, the number of decision trees, Ntree, is a key parameter affecting classification performance. Random forests complete classification through multiple decision trees, so the number of trees directly affects the model's stability and generalization ability. When Ntree is small, the forest contains insufficient information, and the model is easily affected by the randomness of a single decision tree, leading to large fluctuations in classification results. As Ntree increases, model performance usually gradually improves, but an excessively large number of trees increases computational complexity, while the improvement in classification accuracy tends to saturate. Therefore, it is necessary to determine the optimal number of decision trees through experiments. Figure 3 As shown, the search range for the number of decision trees is first set to 20-200, with a step size of 20; the minimum number of leaf node samples is also set to 5. The parameter combination corresponding to the highest classification accuracy is taken as the optimal parameters for the random forest model. Through comparison, the highest classification accuracy of 96.89% is achieved with 100 decision trees. Considering both classification performance and computational efficiency, this embodiment selects 100 decision trees and a minimum number of leaf nodes of 5 as the optimal parameters for the random forest model.
[0024] This invention presents a machine learning-based model for identifying bearing electrical breakdown damage. Addressing the issue of bearing damage morphology being influenced by multiple factors such as load, rotational speed, current, and voltage amplitude under alternating current, a damage morphology identification dataset incorporating four operating parameters (P, R, I, and V) is constructed. Normalization and 5-fold cross-validation are employed to improve the reliability of model training. Based on this, three classification models—Support Vector Machine (SVM), Backpropagation Neural Network (BP neural network), and Random Forest (RF)—are established to achieve intelligent identification of bearing electrical breakdown damage morphology levels under different operating conditions, providing a data-driven method for bearing damage analysis under complex electromechanical coupling. This model effectively reduces random errors introduced by a single classifier through a multi-decision tree ensemble learning mechanism, exhibiting better adaptability to bearing damage morphology classification under small sample sizes and multi-factor coupling conditions. It can serve as a preferred model for identifying and analyzing the evolution of bearing electrical breakdown damage morphology, providing a theoretical basis for subsequent research on bearing electro-erosion failure mechanisms and intelligent condition assessment. The basic principles, main features, and advantages of this invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.
Claims
1. A machine learning-based bearing electrical breakdown damage identification model method, characterized in that, Including the following: Establish database During the data processing, the data obtained in the experiment were first screened, organized and unified. Data from different sources were classified according to the same variable definition and data format, and some abnormal data were checked to ensure the validity and consistency of the data used in subsequent modeling. Experimental data preprocessing To eliminate the influence of differences in units and ranges of values among different input variables, the input features are normalized before model training. The max-min normalization method is used to linearly map each input variable to the [0,1] interval. The calculation formula is as follows: In the formula The original input data, , To be the minimum value, The data is after normalization; Support Vector Machine Classification Model Establishment Support Vector Machines (SVMs) can map raw input data to a high-dimensional feature space through kernel functions, thereby establishing a non-linear mapping relationship between input features and different damage morphology categories; the kernel function formula is: ;in, and For the standardized sample, For kernel function parameters, The Euclidean distance is represented by the Euclidean distance. Support vector machines obtain the optimal decision boundary by maximizing the classification margin. They have good adaptability to small samples, nonlinear and high-dimensional data, and are suitable for bearing damage morphology classification and recognition under limited sample conditions. Establishment of BP neural network classification model Backpropagation (BP) neural networks can learn complex correspondences between input features and output categories from training samples and have strong nonlinear mapping capabilities. Among the structural parameters of BP neural networks, the number of hidden layer neurons is an important parameter affecting the classification performance of the model. The formula is: Where P is the number of neurons in the hidden layer, a and b are the number of neurons in the input layer and output layer, respectively, and c is a positive integer adjustment constant from 1 to 10; Random Forest Model Establishment Multiple sub-training sets were randomly generated from the original training set using the Bootstrap resampling method, and decision tree models were built for each sub-set. The classification results were finally determined by majority voting. The load P, rotational speed R, current I, and voltage amplitude V during bearing operation were used as input variables for the random forest model, and different damage morphology levels were used as output labels. The input features were then divided level by level using multiple decision trees.
2. The bearing electrical breakdown damage identification model method based on machine learning according to claim 1, characterized in that, In the Support Vector Machine (SVM) classification model building module, another key parameter is the penalty factor C. The penalty factor C controls the degree to which the model penalizes misclassified samples. When C is small, the model has a higher tolerance for classification errors, resulting in a relatively large classification margin, which reduces model complexity and improves generalization ability. When C is large, the model can reduce training error, but the classification margin decreases accordingly, increasing model complexity. When the penalty factor is 10... -1 At that time, the classification accuracy reached its highest level, at 78.26%.
3. The bearing electrical breakdown damage identification model method based on machine learning according to claim 2, characterized in that, In the BP neural network classification model building module, if the number of hidden layer neurons is too small, the network's ability to express bearing damage features is insufficient, which can easily lead to underfitting; if the number of hidden layer neurons is too large, the network structure complexity increases, which can easily lead to overlearning of local features in the training samples, resulting in overfitting and reducing the model's ability to predict unknown samples.
4. The bearing electrical breakdown damage identification model method based on machine learning according to claim 3, characterized in that, The number of hidden layer neurons needs to be initially estimated based on the size of the input and output layers, and further determined through validation sets or cross-validation. Since the input layer has 4 nodes and the output layer has 3 nodes, the search range for hidden layer neurons is set to 3-14, and the network structure with the highest classification accuracy is used as the final model.
5. The bearing electrical breakdown damage identification model method based on machine learning according to claim 4, characterized in that, To obtain the best model prediction results, a candidate range for the number of hidden layer neurons is determined based on empirical formulas, and the number of hidden neurons is changed sequentially within this range.
6. The bearing electrical breakdown damage identification model method based on machine learning according to claim 5, characterized in that, To reduce the impact of random initial weights on the training results, each network structure was trained multiple times. Five-fold cross-validation was used to calculate the average classification accuracy and average loss. Finally, the optimal number of hidden neurons was selected as the one with the highest classification accuracy on the validation set, the smallest fluctuation in the results, and the relatively simple network structure.
7. The bearing electrical breakdown damage identification model method based on machine learning according to claim 6, characterized in that, In the random forest model building module, the load P, rotational speed R, current I, and voltage amplitude V during the bearing operation process are used as input variables of the random forest model, and different damage morphology levels are used as output labels. The input features are divided level by level through multiple decision trees.
8. The bearing electrical breakdown damage identification model method based on machine learning according to claim 7, characterized in that, In the random forest model, the number of decision trees, Ntree, is a key parameter affecting classification performance. Random forests complete classification through multiple decision trees, and the number of trees directly affects the model's stability and generalization ability.
9. The bearing electrical breakdown damage identification model method based on machine learning according to claim 8, characterized in that, When Ntree is small, the forest contains insufficient information, and the model is easily affected by the randomness of a single decision tree, resulting in large fluctuations in classification results. When Ntree increases, the model performance usually gradually improves, but an excessive number of trees will increase computational complexity, while the improvement in classification accuracy tends to saturate.
