Bearing fault classification method based on white whale optimization algorithm and random forest
By combining the white whale optimization algorithm and random forest, the feature selection is optimized, and the problem of the rolling bearings of the wind turbine transmission system are affected by redundant characteristics in the fault diagnosis is solved, achieving higher diagnostic accuracy and faster processing speed.
Patent Information
- Application Number
- CN202411972005.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-11-04
- Filing Date
- 2024-12-30
- Publication Date
- 2025-04-11
AI Technical Summary
The rolling bearings of the wind turbine transmission system are susceptible to redundant features in the diagnosis of faults, resulting in a decrease in diagnostic accuracy and an increase in model complexity.
The bearing failure classification method based on beluga whale optimization algorithm and random forest is adopted to optimize feature selection and reduce the impact of redundant features through wavelet transformation, beluga whale optimization algorithm and random forest.
It improves the accuracy and processing speed of bearing fault diagnosis, reduces the data dimension, and enhances the diagnostic accuracy and interpretability of the model.
Smart Images
Figure CN119807915A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a bearing fault classification method based on a Beluga optimization algorithm and a random forest. Background Art
[0002] As an indispensable core component in the transmission system of wind turbines, the performance of rolling bearings directly determines the stability and reliability of the entire power generation system. However, wind turbines are easily affected by adverse factors such as high temperature, high humidity, and frequent load changes during operation, which leads to frequent failures of rolling bearings, one of its core components. Therefore, it is of great significance to perform efficient and accurate fault diagnosis of rolling bearings.
[0003] Feature extraction is crucial for rolling bearing fault diagnosis, but with the explosive growth of data volume, the number of features has increased dramatically, resulting in many redundant features. Redundant features not only increase the computational cost of model training, but may also interfere with the identification of key information, thereby affecting the accuracy of diagnosis. In addition, redundant features may also increase model complexity, reduce model interpretability, increase storage and computing resource consumption, affect model stability and adaptability, and increase maintenance costs.
[0004] Therefore, a classification method that can reduce the influence of redundant features is needed to more accurately diagnose bearing faults. Summary of the invention
[0005] In view of this, the present invention provides a bearing fault classification method based on Beluga Whale Optimization (BWO) and Random Forest (RF). By integrating wavelet transform, Beluga Whale Optimization algorithm and Random Forest, the problem of rolling bearings in wind turbine transmission system being easily affected by redundant features in fault diagnosis is solved, and the diagnosis accuracy is improved.
[0006] To achieve the above purpose, the technical solution adopted by the present invention is: a bearing fault classification method based on the Beluga optimization algorithm and random forest, the method comprising the following steps: S1, data collection, obtaining the fault data of the rolling bearing of the motor group transmission system; S2, preprocessing the acquired data; S3, extracting fault features through time domain and frequency domain analysis to obtain a rolling bearing vibration signal feature set; S4. Build a bearing fault classification model based on the White Whale optimization algorithm and random forest, and perform fault classification: RF is used to preliminarily classify the data in the feature set and obtain the classification accuracy. The classification accuracy and feature selection rate are two key indicators that constitute the fitness. The BWO algorithm is used to find the optimal feature subset. S5. Divide the selected optimal feature subset into a training set and a test set and input them into the RF classifier to obtain the diagnosis result.
[0007] Furthermore, the preprocessing in S2 includes any one or more of the following processing methods: duplicate value integration, outlier removal, missing value supplementation and noise reduction.
[0008] Furthermore, the S3 extracting fault features specifically includes the following steps: S31, every 500 data points are a sample, forming a feature matrix, and the feature matrix is labeled; S32, extracting time domain related features and frequency domain related features from each sample; S33. Through linear transformation, the value range of the data is adjusted to a unified standardized interval to normalize the features.
[0009] Specifically, the label marking in S31 is to use numbers to score different working condition data in the feature matrix, using 1-3 integer markings.
[0010] Furthermore, in S4, the two key indicators of classification accuracy and feature selection rate constitute the fitness parameter fit, which conforms to the following formula:
[0011] Among them, fit represents fitness, α and β are weight coefficients used to balance the influence of two factors, ACC represents classification accuracy, and sum(x) / length(x) represents feature selection rate.
[0012] The beneficial effects of the present invention are: The method of the present invention is based on the BWO algorithm and the RF algorithm, and solves the problem that the rolling bearing of the wind turbine transmission system is easily affected by redundant features in fault diagnosis; it reduces the dimension of the data, and improves the diagnostic accuracy and processing speed of the model by optimizing feature selection, providing strong experimental evidence and theoretical support for research and practice in the field of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0014] Figure 1 is a flow chart of the bearing fault classification method of the present invention; Figure 2 is a comparative optimization flow chart of the present method and the existing optimization method in an embodiment of the present invention; Figure 3 It is an iterative evolution curve diagram of the optimization algorithm of the present invention. DETAILED DESCRIPTION
[0015] The following specific embodiments are given to further clearly, completely and in detail illustrate the technical solution of the present invention. The embodiments described in the present invention are only some preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention.
[0016] Example 1 A bearing fault classification method based on the Beluga optimization algorithm and random forest, the method comprises the following steps: Step 1: Data collection, obtaining fault data of rolling bearings of the motor group transmission system; In this embodiment, the collected data is the fault data of the wind turbine transmission system simulation test bench of Southeast University. The test bench can reproduce the transmission system of the wind turbine to ensure the reliability of the collected data.
[0017] Among them, the Southeast University experimental bench is composed of several key components, including: the main shaft as the power transmission center, the variable frequency drive motor for simulating the driving force of the wind wheel, the load motor representing the actual wind power system load, the torque speed sensor for real-time monitoring of the system torque and speed, the coupling for connecting different rotating parts, and the rolling bearings used in the simulation of actual wind turbines. In addition, the experimental bench is also equipped with a planetary gearbox and a cylindrical gearbox to reproduce the speed increase mechanism of the wind turbine. The signal acquisition and analysis system of the experimental bench is used to capture and analyze the vibration and acoustic emission signals during the experiment.
[0018] Step 2: Preprocess the acquired data, including any one or more of the following processing methods: Duplicate value integration, outlier removal, missing value filling and noise reduction processing. Furthermore, in the field of feature selection and fault diagnosis of rolling bearings, data preprocessing is an extremely critical step, which is essential to improving data quality and optimizing data availability, thereby enabling the model to learn more efficiently and improve prediction performance; the denoising method used is mainly wavelet transform.
[0019] Step 3: Extract fault features and obtain the rolling bearing vibration signal feature set. For rolling bearing vibration signals, there are mainly two types of methods: stationary signal and non-stationary signal. Stationary signal analysis includes time domain and frequency domain feature extraction, while non-stationary signal analysis involves a variety of envelope analysis techniques. Specifically, the following steps are included: S31. After data preprocessing, every 500 data points are taken as a sample to form a feature matrix, and the feature matrix is labeled; Specifically, the label marking in S31 is to use numbers to score different working condition data in the feature matrix, using 1-3 integer markings to prepare for subsequent feature selection and data division.
[0020] S32, extracting time domain related features and frequency domain related features from each sample; In this embodiment, 16 time-domain related features are extracted, and the features are as follows: mean, root mean square, root mean square amplitude, variance, standard deviation, maximum value, minimum value, peak-to-peak value, absolute mean, kurtosis, skewness, peak index, waveform index, pulse factor, margin factor and kurtosis coefficient; Four frequency domain related features are extracted, the features are as follows: mean frequency, frequency standard deviation, spectrum root mean square and frequency center.
[0021] S33, normalize the features through linear transformation. In this embodiment, the minimum-maximum normalization technique is applied to adjust the value range of the data to a unified standardized interval, such as [0,1]; this standardization process not only balances the potential impact of different dimensional features on model training, but also promotes the convergence efficiency of the algorithm.
[0022] Step 4: Build a bearing fault classification model based on the White Whale optimization algorithm and random forest, and perform fault classification: RF is used to perform preliminary classification of the data in the feature set to obtain a benchmark classification accuracy; the two key indicators of classification accuracy and feature selection rate constitute the fitness, and then the fitness of the BWO algorithm is used as the evaluation parameter, and the path selection ability of the BWO algorithm is used to find the optimal feature subset among the above 20 features.
[0023] Furthermore, in S4, the two key indicators of classification accuracy and feature selection rate constitute the fitness parameter fit, which is intended to standardize the performance evaluation of the algorithm and guide it to make the best decision in feature selection, in accordance with the following formula: Among them, fit represents fitness, α and β are weight coefficients used to balance the influence of two factors, ACC represents classification accuracy (ACC = number of accurately predicted samples / total number of samples), and sum(x) / length(x) represents feature selection rate. By changing the weight coefficient, the algorithm can adjust the emphasis on accuracy and feature selection rate to optimize the feature selection process. The introduction of this fitness formula can make a thorough and fair evaluation of the performance of various optimization algorithms.
[0024] The fitness parameter combines two key indicators, classification accuracy and feature selection rate, to ensure that the algorithm improves diagnostic accuracy while reducing the number of features.
[0025] Step 5: Divide the selected optimal feature subset into training set and test set and input them into the RF classifier to obtain the diagnosis result. In order to verify the effectiveness of the selected optimization algorithm, it is compared with various optimization algorithms. The comparison flow chart is as follows Figure 2 As shown; the specific experiment is as follows: In order to ensure a fair comparison of the performance of different optimization algorithms, a unified experimental setting is adopted. Specifically, the number of iterations of all optimization algorithms is set to 50 times, and the population size is set to 60 to ensure the consistency of experimental conditions. In each round of iteration, the algorithm will output its current optimal fitness parameters. Based on the fitness data output by these iterations, the iterative evolution curve of the optimization algorithm can be obtained, such as Figure 3 shown.
[0026] Figure 3 The middle curve clearly shows the excellent performance of the BWO algorithm in the iteration process. By comparing the experimental results, it is found that the BWO algorithm performs well in efficiency and accuracy compared with the other five optimization algorithms. In particular, at the 32nd iteration, the BWO algorithm reached a fitness value of 0.1008, which is the lowest among all the optimization algorithms involved in the comparison. This result not only highlights the advantage of the BWO algorithm in iteration efficiency, but also shows its excellence in solution quality.
[0027] In summary, the results of the control experiment fully prove the correctness of the BWO optimization algorithm selected by the present invention, and confirm the reliability and effectiveness of the BWO algorithm in solving related optimization problems.
[0028] The present invention further utilizes the optimization ability of the BWO algorithm and uses the fitness of the BWO as an evaluation index to identify the optimal feature subset. In order to ensure the robustness and reliability of the results, three independent experiments were conducted, each of which was dedicated to finding the global optimal feature subset with the lowest fitness value. By analyzing and comparing the results of these three experiments, a set of comprehensive optimal feature subsets was finally obtained, as shown in Table 1. Table 1 Optimal feature selection results
[0029] In the feature selection results, 0 means that the feature is not selected, and 1 means that the feature is selected. The experimental results show that the four features of absolute mean, root mean square value, margin factor and kurtosis coefficient constitute the optimal feature set.
[0030] A simple random sampling method is used to divide the data set, and random indexes are generated to ensure the distribution consistency of data in the training set and the test set. 70% of the most characteristic subset is selected as the training set for model training; the remaining 30% is used as the test set to evaluate the generalization ability of the model.
[0031] Furthermore, in order to demonstrate the effectiveness of the method of the present invention, the original rolling bearing vibration data without any feature extraction processing, the complete data set containing all 20 features, and the four feature data selected by the BWO optimization algorithm, specifically the absolute mean, the root mean square value, the margin factor, and the kurtosis coefficient, are respectively input into the classifier, in order to calculate the classification accuracy after fault classification.
[0032] In order to ensure the reliability of the experimental results and reduce the impact of random errors, each set of experiments was repeated 6 times. At the same time, considering the impact of different classifiers on the experimental results, this paper sets KNN, RF and SVM classifiers. Through these experiments, the data of classification accuracy were collected and their average values were calculated. These experimental results are recorded in detail in Table 2, Table 3 and Table 4, which intuitively show the impact of different data on the fault diagnosis accuracy of the classifier. Table 2 KNN classification accuracy results under different data Table 3 RF classification accuracy results under different data Table 4 SVM classification accuracy results under different data
[0033] The experimental results show that the optimal feature subset selected by the BWO algorithm has a good effect in improving the accuracy of the classifier. The experimental results of three different classifiers are compared with each other to fully verify the effectiveness of the proposed method. In the initial experiment, the average accuracy of the KNN classifier was only 55.91%, the average accuracy of the RF classifier was only 62.08%, and the average accuracy of the SVM classifier was only 51.30% when the unprocessed raw vibration data was directly used. This reflects the lack of diagnostic accuracy of the model without feature extraction. However, when all 20 extracted features were introduced, the average accuracy increased to 76.15%, 90.36% and 85.84% respectively. This jump highlights the important role of feature extraction in improving the diagnostic ability of the model. More strikingly, when the four key features selected by the BWO optimization algorithm were used, the average accuracy of the three classifiers further climbed to 83.33%, 95.06% and 90.47%, which highlights the significant advantages of feature selection in improving fault diagnosis accuracy.
[0034] That is to say, the feature selection strategy of the BWO algorithm has increased the accuracy rate instead of decreasing it when the number of features is reduced. The more complete feature data set achieved an accuracy increase of 7.18%, 4.7% and 4.63%, reaching 83.33%, 95.06% and 90.47% respectively. At the same time, the processing speed of the three classifiers has also been improved to a certain extent when performing feature classification under this optimal feature subset.
[0035] In summary, the present invention successfully verifies the excellent performance of the best feature subset obtained by the BWO algorithm in fault diagnosis. This strategy not only reduces the dimension of the data, but also improves the diagnostic accuracy and processing speed of the model by optimizing feature selection, providing strong experimental evidence and theoretical support for research and practice in the field of fault diagnosis.
[0036] The above shows and describes the main features, basic principles and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements according to actual conditions, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A bearing fault classification method based on the Beluga optimization algorithm and random forest, characterized in that: The method comprises the following steps: S1, data collection, obtaining the fault data of the rolling bearing of the motor group transmission system; S2, preprocessing the acquired data; S3, extracting fault features through time domain and frequency domain analysis to obtain a rolling bearing vibration signal feature set; S4. Build a bearing fault classification model based on the White Whale optimization algorithm and random forest, and perform fault classification: RF is used to preliminarily classify the data in the feature set and obtain the classification accuracy. The classification accuracy and feature selection rate are two key indicators that constitute the fitness. The BWO algorithm is used to find the optimal feature subset. S5. Divide the selected optimal feature subset into a training set and a test set and input them into the RF classifier to obtain the diagnosis result.
2. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 1 is characterized in that: The preprocessing in S2 includes any one or several processing methods, including duplicate value integration, outlier removal, missing value supplementation and noise reduction.
3. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 2 is characterized in that: The noise reduction process adopts wavelet transform.
4. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 1 is characterized in that: The S3 extracting fault features specifically includes the following steps: S31, every 500 data points are a sample, forming a feature matrix, and the feature matrix is labeled; S32, extracting time domain related features and frequency domain related features from each sample; S33. Through linear transformation, the value range of the data is adjusted to a unified standardized interval to normalize the features.
5. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 4 is characterized in that: The label marking in S31 is specifically to score different working condition data using numbers for the feature matrix, using 1-3 integer markings.
6. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 1 is characterized in that: In S4, the two key indicators of classification accuracy and feature selection rate constitute the fitness parameter fit, which conforms to the following formula: Among them, fit represents fitness, α and β are weight coefficients used to balance the influence of two factors, ACC represents classification accuracy, and sum(x) / length(x) represents feature selection rate.
7. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 1 is characterized in that: In S5, a simple random sampling method is used to divide the data set, and random indexes are generated to ensure the distribution consistency of data in the training set and the test set.
8. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 1 is characterized in that: In S5, the original rolling bearing vibration data, the rolling bearing vibration signal feature set and the optimal feature subset are respectively input into the RF classifier, and the respective classification accuracy rates are calculated.
9. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 4, characterized in that: The time domain related characteristics are mean, root mean square, root mean square amplitude, variance, standard deviation, maximum value, minimum value, peak-to-peak value, absolute mean, kurtosis, skewness, peak index, waveform index, pulse factor, margin factor and kurtosis coefficient.
10. The bearing fault classification method based on the Beluga optimization algorithm and random forest according to claim 4, characterized in that: The frequency domain related features are mean frequency, frequency standard deviation, spectrum root mean square and frequency center.