High-Speed Train Bearing Fault Diagnosis Method Based on Ensemble Learning

Abstract

The present invention discloses a high-speed train bearing fault diagnosis method based on ensemble learning, which is specifically implemented according to the following steps: After obtaining the noisy original signal, fault marking division is carried out, and then CEEMDAN decomposition is performed. The obtained IMF components are denoised; then IMF component reconstruction is carried out, and then feature extraction is carried out; the extracted features are respectively input into the single models of the first layer of the ensemble learning model to obtain classification results; different weights are assigned to the single models of the first layer of the ensemble learning model according to the classification results, and then they are integrated into a training set. The generated training set is input into the random forest model of the second layer of the ensemble learning model for training to obtain the final bearing fault diagnosis result. The fault signal extracted by the present invention has high accuracy, improves the classification accuracy rate, and solves the problem of low accuracy rate of the existing high-speed train bearing fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of high-speed train bearing fault diagnosis, and relates to a high-speed train bearing fault diagnosis method based on integrated learning. Background Art

[0002] As an important public transportation mode, rail transit is characterized by large transportation capacity and high speed. Its operating environment is complex and the passenger volume is large. Once a failure occurs, it is directly related to the safety of passengers. The rolling bearings of high-speed trains are not only important components of various mechanical systems in the train running gear, but also one of the components prone to failure. It supports the axle and bears the load between the wheel set and the car body. Its operating conditions have an important impact on the running safety of high-speed trains. Ensuring its good operating condition is the key to the safe running of the train. Since high-speed trains need to experience complex operating conditions such as curves, high speeds, severe cold, and high temperatures, the wheel set bearings, as key load-bearing components of the train, will bear various impact loads during long-term service, and are prone to fatigue damage and various performance degradations. If the fault information of the wheel set bearings cannot be detected timely and correctly, it may lead to hot axles, burning axles, or even axle cutting, thus inducing major safety accidents such as train derailment. With the increase of rail transit operation lines and the daily passenger volume, the traditional regular maintenance cannot meet the requirements of timely and effective fault monitoring and diagnosis of bearings in actual engineering due to its low diagnostic efficiency, and it is necessary to improve the bearing diagnosis method. Therefore, carrying out research on the wheel set bearing fault detection method is of particularly important significance for the development of high-speed trains with high safety and high reliability in service.

[0003] The problem of high-speed train bearing fault diagnosis is essentially a fault classification problem. The high-speed train bearing fault diagnosis method adopted in this study is: installing a vibration signal acquisition device on the train and performing time-domain analysis on the fault vibration signal to judge the bearing fault type. The bearing fault diagnosis technology based on vibration signals is adopted. However, the operating conditions of high-speed trains are complex, and a large amount of noise will be mixed in the vibration signals of the rolling bearings of the train running gear collected in actual engineering. The noise signals will affect the effect of bearing fault diagnosis. Therefore, before fault diagnosis, it is necessary to perform noise reduction processing on the signals. The work of this study aims to design a bearing fault diagnosis algorithm based on vibration signals to timely detect the faults existing in the train axles and ensure the running safety of the train.

[0004] In the prior art, when training bearing fault data with a single model, the effect is often not ideal, and there is great instability in the diagnosis of different fault signals. Using advanced machine learning techniques for intelligent fault diagnosis of high-speed train bearings, a highly reliable intelligent fault diagnosis method has extremely high practical value. The theoretical significance of this research lies in the ability to apply the ensemble learning method to bearing fault diagnosis. The method of ensemble learning for bearing faults has high diagnostic accuracy and good classification effect, and will be one of the main development directions of future fault diagnosis technologies. Summary of the Invention

[0005] The object of the present invention is to provide a high-speed train bearing fault diagnosis method based on ensemble learning, which solves the problem of low accuracy of existing high-speed train bearing fault diagnosis.

[0006] The technical solution adopted by the present invention is that the high-speed train bearing fault diagnosis method based on ensemble learning is specifically implemented according to the following steps:

[0007] Step 1: After obtaining the original vibration signal of the high-speed train bearing, perform fault label division, and then perform CEEMDAN decomposition to obtain a series of intrinsic mode functions IMF;

[0008] Step 2: Denoise the IMF components obtained in Step 1;

[0009] Step 3: Reconstruct the IMF components processed in Step 2, and perform feature extraction after obtaining the reconstructed signal;

[0010] Step 4: Respectively input the features extracted in Step 3 into the single models in the first layer of the ensemble learning model to obtain classification results;

[0011] Step 5: Calculate the classification accuracy of the single models in the first layer of the ensemble learning model according to the classification results in Step 4, assign weights to the model accuracies, and integrate them into a training set;

[0012] Step 6: Input the training set generated in Step 5 into the random forest model in the second layer of the ensemble learning model for training to obtain the final bearing fault diagnosis result.

[0013] The characteristics of the present invention also lie in that

[0014] Step 1 is specifically implemented according to the following steps:

[0015] Step 1.1: Obtain the original noisy vibration signal of the high-speed train bearing. Extract the bearing data features from the noisy original signal to obtain a bearing data feature dataset. Mark the features of the normal bearing data, and then find the fault data from the full-life data. Divide the "fault" marks of the bearing fault data; when dividing the "fault" marks, according to the fault degree of the bearing from the outside to the inside, they are successively divided into: the features of the outer ring slightly faulty bearing data, the features of the outer ring moderately faulty bearing data, the features of the outer ring severely faulty bearing data, the features of the inner ring slightly faulty bearing data, the features of the inner ring moderately faulty bearing data, and the features of the inner ring severely faulty bearing data;

[0016] Step 1.2: Input the original signal marked with "fault" in Step 1.1 into the CEEMDAN algorithm model. After being processed by CEEMDAN, the signal is decomposed into several IMF components and a Res residue. Each IMF component corresponds to a different frequency component, and the several IMF components are distributed in the order of high frequency to low frequency of the frequency components.

[0017] Step 2 is specifically implemented according to the following steps:

[0018] Step 2.1: Calculate the RMSE values of 17 IMF components in the order of high frequency to low frequency using the root mean square error. When the RMSE values of the IMF components gradually increase and the IMF components before the increase are monotonically decreasing, the IMF components before the increase are high-frequency IMF components and contain interference signals, and the remaining IMF components after the increase are low-frequency IMF components. Select the high-frequency IMF components for wavelet denoising processing; when performing wavelet denoising processing, a noisy model is expressed as:

[0019]

[0020] where f(k) is the useful signal, s(k) is the noisy signal, e(k) is the noise, and ε is the standard deviation of the noise coefficient;

[0021] Step 2.2: Perform threshold selection on the high-frequency IMF components after the denoising processing in Step 2.1;

[0022] The threshold selection uses hard threshold quantization to retain the local features of the bearing vibration signal edge.

[0023] The specific process of Step 3 is: After quantization processing, the high-frequency IMF components processed in Step 2 are used to obtain high-frequency coefficients, and wavelet decomposition is used to process the low-frequency IMF components to obtain low-frequency coefficients. The low-frequency coefficients and high-frequency coefficients are linearly added for wavelet reconstruction of the signal, and the time-domain features of the vibration signal are extracted.

[0024] The specific process of step 4 is as follows: The ensemble learning model adopts the Stacking ensemble learning model. The single models in the first layer of the Stacking ensemble learning model include SVM, KNN, AdaBoost, XGBoost, and LightGBM. The features extracted in step 3 are respectively input into the five single models to obtain five different classification results.

[0025] The classification processes of the five single models in step 4 are as follows:

[0026] SVM constructs multiple classifiers by the indirect method to classify fault diagnosis. During training, the samples of a certain category are successively classified into one category, and the other remaining samples are classified into another category. During classification, the unknown samples are classified into the one with the maximum classification function value.

[0027] During KNN classification, by calculating the distances between the points to be classified and the points of known categories, sorting them in ascending order of distance, selecting the K points with the smallest distances to the points to be classified, determining the occurrence times of the categories where the first K points are located, and returning the category with the most occurrence times among the first K points as the classification of the points to be classified. KNN classification only determines the category of the samples to be classified based on the categories of one or several nearest samples in the class decision.

[0028] During AdaBoost classification, the idea of iteration is adopted. Only one weak classifier is trained in each iteration, and the trained weak classifier will participate in the next iteration.

[0029] During XGBoost classification, the base learner used is the CART regression tree. The ensemble model is constructed by gradually adding trees. Assuming that the model has a total of K trees integrated, the sum of the leaf node values corresponding to the K trees is the final classification result of the model. Newton's method is used to solve the extreme value of the loss function, the loss function is Taylor-expanded to the second order, and a regularization term is added to the loss function.

[0030] During LightGBM classification, a model with a higher diagnostic rate is established through gradual optimization. The Histogram-based decision tree algorithm is adopted, the Leaf-wise leaf growth strategy with depth limit is adopted, the histogram difference is used for acceleration, categorical features are directly supported, the Cache hit rate optimization is adopted, the Histogram-based sparse feature optimization is adopted, and the multi-thread optimization is adopted.

[0031] In step 5, different weights are assigned to the classification accuracies of the single models SVM, KNN, AdaBoost, XGBoost, and LightGBM in the first layer, denoted as w1, w2, w3, w4, and w5 respectively. The information entropy model is used to construct a function and calculate the weight values of each parameter, and they are integrated into a training set according to the weights.

[0032] The specific process of step 5 is as follows:

[0033] Step 5.1, establish the mathematical model of the system. Assume that X is a known matrix, where represents the j-th index of the i-th evaluation object, construct the data matrix, eliminate the dimension of the data matrix X and perform normalization processing to obtain the matrix Y.

[0034]

[0035] In formula (2), maxx*j and minx*j represent the maximum and minimum values of the j-th column of the data matrix X respectively. is the average value of the data matrix X, and any value in the matrix Y is within [0, 1].

[0036] Step 5.2, the information entropy model establishes the weight matrix P with the bearing fault diagnosis accuracy rate as the evaluation index, then P j represents the weight of the j-th evaluation index, and the sum of P j is 1 and P j ≥0. Use the entropy value to construct a function and calculate the weight value of each parameter:

[0037] Construct the function H for calculating the matrix Y. The function H has symmetry, so H(x1, x2) = H(x2, x1). When the order of the evaluation objects changes, the weight of the same evaluation index remains unchanged, that is, when any two rows of the matrix Y change, the value of the function remains unchanged; the function H requires monotonic increase, continuity, and additivity, so as to construct the function:

[0038]

[0039] Calculate the entropy value of each parameter. Among them, the entropy value of the j-th index is calculated as:

[0040]

[0041] To ensure that the entropy value is positive, a negative sign is taken. Information entropy is a quantity used in information theory to describe the degree of information redundancy. The larger the entropy value, the higher the degree of information disorder and the corresponding higher information efficiency.

[0042] The normalization coefficient is defined as:

[0043]

[0044] Use the entropy value to calculate the weight value of each parameter:

[0045]

[0046] The basic unit of the random forest model is a decision tree. Each decision tree is a classifier. For an input sample, N trees will have N classification results. By integrating all the classification voting results of the N trees, the category with the most votes is designated as the final output, completing the diagnosis of high-speed train bearing faults.

[0047] In step 6, the Bootstraping method is used to randomly sample m samples from the training set synthesized in step 5 with replacement, and the sampling is performed n_tree times to generate n_tree training sets. For the n_tree training sets, n_tree decision tree models are trained respectively. For a single decision tree model, assuming the number of training sample features is n, the best feature is selected for splitting according to the Gini coefficient each time of splitting. Each tree keeps splitting like this until all the training examples at this node belong to the same class. Pruning is not required during the splitting process of the decision tree. The multiple generated decision trees are combined into a random forest, and the final fault classification result is determined by the voting of multiple tree classifiers.

[0048] The beneficial effects of the present invention are as follows: The present invention uses the method of combining CEEMDAN wavelet threshold joint denoising to jointly process the denoising of bearing vibration signals, separates the high-frequency components and low-frequency components for processing, improves the denoising accuracy of the original signal, has a high accuracy of the extracted fault signals, obtains relatively pure bearing fault signals, uses the Stacking ensemble learning model to preprocess the initial data set and then trains several models, and then combines the output of each model in the first layer, that is, various different classification diagnosis results, as the input of the second layer and continues to train in the second layer, improving the classification accuracy rate, and diagnosing more accurately whether the bearing is faulty and the type of fault, solving the problem of low accuracy of the existing high-speed train bearing fault diagnosis. Description of the Drawings

[0049] Figure 1 is a schematic flow chart of the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention;

[0050] Figure 2 is a wavelet threshold denoising flow chart of the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention;

[0051] Figure 3 is a time series diagram after CEEMDAN decomposition of the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention;

[0052] Figure 4 is an instantaneous frequency diagram after CEEMDAN decomposition of the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention;

[0053] Figure 5(a) is the curve graph of the original signal before wavelet threshold processing by the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention;

[0054] Figure 5(b) is the curve graph of the original signal after wavelet threshold processing by the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention;

[0055] Figure 6(a) is the curve graph of the IMF1 component before wavelet threshold processing by the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention;

[0056] Figure 6(b) is the curve graph of the IMF1 component after wavelet threshold processing by the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention;

[0057] Figure 7 is the comparison graph of the accuracy rates of the classification results of each model of the high-speed train bearing fault diagnosis method based on ensemble learning of the present invention. Detailed implementation manners

[0058] The present invention will be described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0059] The high-speed train bearing fault diagnosis method based on ensemble learning of the present invention has a process as Figure 1 shown, and is specifically implemented according to the following steps:

[0060] Step 1: After obtaining the original vibration signal of the high-speed train bearing, perform fault label division, and then perform CEEMDAN decomposition to obtain a series of intrinsic mode components IMF;

[0061] Step 1 is specifically implemented according to the following steps:

[0062] Step 1.1: Obtain the original noisy vibration signal of the high-speed train bearing, extract the bearing data features from the noisy original signal to obtain a bearing data feature dataset, label the features of the normal bearing data, and then find the fault data from the full-life data to divide the "fault" label of the bearing fault data;

[0063] When dividing the "fault" label, it is successively divided according to the fault degree of the bearing from the outside to the inside into: the features of the outer ring slightly faulty bearing data, the features of the outer ring moderately faulty bearing data, the features of the outer ring severely faulty bearing data, the features of the inner ring slightly faulty bearing data, the features of the inner ring moderately faulty bearing data, and the features of the inner ring severely faulty bearing data.

[0064] Step 1.2: Input the original signal marked as "fault" in Step 1.1 into the CEEMDAN algorithm model. After being processed by CEEMDAN, the signal is decomposed into several IMF components and a Res residue. Each IMF component corresponds to a different frequency component, and the several IMF components are distributed in the order of high frequency to low frequency of the frequency components.

[0065] Step 2: Denoise the IMF components obtained in Step 1;

[0066] Step 2 is specifically implemented according to the following steps:

[0067] Step 2.1: Calculate the RMSE values of several IMF components in the order of high frequency to low frequency using the root mean square error. When the RMSE values of the IMF components gradually increase and the IMF component before the increase is monotonically decreasing, the IMF component before the increase is the high-frequency IMF component and contains interference signals, and the remaining IMF components after the increase are low-frequency IMF components. Select the high-frequency IMF components for wavelet denoising processing;

[0068] The root mean square error reflects the precision of the measurement and is used as a standard for evaluating which IMF components to denoise. Wavelet denoising combines feature extraction and low-pass filtering. Since wavelet denoising retains the part of feature extraction, its performance is superior to traditional denoising methods.

[0069] When performing wavelet denoising processing, a noisy model is expressed as:

[0070] s(k) = f(k) + ε * e(k) k = 0, 1......n - 1 (1)

[0071] where f(k) is the useful signal, s(k) is the noisy signal, e(k) is the noise, and ε is the standard deviation of the noise coefficient.

[0072] Step 2.2: Perform threshold selection on the high-frequency IMF components after the denoising processing in Step 2.1;

[0073] Threshold selection uses hard threshold quantization to retain the local features of the bearing vibration signal edge and avoid edge blurring and distortion.

[0074] Step 3: Reconstruct the IMF components processed in Step 2, and extract features after obtaining the reconstructed signal;

[0075] The specific process of Step 3 is as follows: The high-frequency IMF components processed in Step 2 are quantized to obtain high-frequency coefficients. As Figure 2 shown, use wavelet decomposition to process the low-frequency IMF components to obtain low-frequency coefficients, linearly add the low-frequency coefficients and the high-frequency coefficients for wavelet reconstruction of the signal, and extract the time-domain features of the vibration signal.

[0076] Wavelet denoising is based on the characteristic that the wavelet decomposition coefficients of noise and signal have different intensity distributions in different frequency bands, and the wavelet coefficients corresponding to the noise in each frequency band are removed; in step 3, the wavelet decomposition coefficients of the original signal are retained; the processed coefficients are wavelet reconstructed to obtain a pure signal. Quantization processing is the process of approximating the continuous values of a signal to a finite number of discrete values.

[0077] Extracting the time-domain features of vibration signals includes: dimensional and dimensionless. Dimensional features include mean value, root mean square value, root amplitude, absolute average value, skewness, kurtosis, variance, maximum value, minimum value, and peak-to-peak value; dimensionless features include waveform index, peak index, pulse index, margin index, skewness index, and kurtosis index.

[0078] Step 4: The features extracted in step 3 are respectively input into the first-layer single models of the ensemble learning model to obtain classification results;

[0079] The specific process of step 4 is as follows: The ensemble learning model uses the Stacking ensemble learning model. The first-layer single models of the Stacking ensemble learning model include SVM, KNN, AdaBoost, XGBoost, and LightGBM. The features extracted in step 3 are respectively input into the five single models to obtain five different classification results:

[0080] SVM constructs multiple classifiers by the indirect method to classify fault diagnosis. During training, the samples of a certain category are successively grouped into one category, and the other remaining samples are grouped into another category. During classification, the unknown samples are classified into the one with the maximum classification function value.

[0081] During KNN classification, by calculating the distances between the points to be classified and the points with known categories, sorting them in ascending order of distance, selecting the K points with the smallest distances to the points to be classified, determining the occurrence times of the categories where the first K points are located, and returning the category with the most occurrence times among the first K points as the classification of the points to be classified. KNN classification only determines the category of the samples to be classified based on the categories of one or several nearest samples in the class decision.

[0082] During AdaBoost classification, the idea of iteration is adopted. Only one weak classifier is trained in each iteration, and the trained weak classifier will participate in the next iteration.

[0083] During XGBoost classification, the base learner used is the CART regression tree, and the ensemble model is constructed by gradually adding trees. Assume that the model has a total of K trees integrated, and the sum of the leaf node values corresponding to the K trees is the final classification result of the model. Newton's method is used to solve the extreme value of the loss function, the loss function is Taylor-expanded to the second order, and a regularization term is added to the loss function.

[0084] When using LightGBM for classification, a model with a higher diagnostic rate is built through step-by-step optimization. It adopts the Histogram-based decision tree algorithm, the Leaf-wise leaf growth strategy with depth limit, uses histogram difference acceleration, directly supports categorical features, adopts Cache hit rate optimization, adopts histogram-based sparse feature optimization, and adopts multi-threaded optimization.

[0085] Step 5: Calculate the classification accuracy of the single models in the first layer of the ensemble learning model based on the classification results in Step 4, and integrate them into a training set after assigning weights to the model accuracies.

[0086] Step 5 is specifically implemented as follows: Different weights are assigned to the classification accuracies of the single models SVM, KNN, AdaBoost, XGBoost, and LightGBM in the first layer, denoted as w1, w2, w3, w4, and w5 respectively; The information entropy model is used to construct a function and calculate the weight values of each parameter, and they are integrated into a training set according to the weights.

[0087] Step 5.1: Establish the mathematical model of the system. Assume that X is a known matrix, where represents the j-th index of the i-th evaluation object, construct the data matrix, eliminate the dimension of the data matrix X and perform normalization processing to obtain the matrix Y.

[0088]

[0089] In formula (2), maxx*j and minx*j represent the maximum and minimum values of the j-th column of the data matrix X respectively. is the average value of the data matrix X, and any value in the matrix Y is within [0, 1].

[0090] Step 5.2: The information entropy model establishes a weight matrix P for the bearing fault diagnosis accuracy as the evaluation index, then P j represents the weight of the j-th evaluation index, and the sum of P j is 1 and P j ≥0. Use the entropy value to construct a function and calculate the weight values of each parameter:

[0091] Construct a function H for calculating the matrix Y. The function H has symmetry, so H(x1, x2) = H(x2, x1). When the order of the evaluation objects changes, the weights of the same evaluation index remain unchanged, that is, when any two rows of the matrix Y change, the value of the function remains unchanged; The function H requires monotonic increase, continuity, and additivity, so as to construct the function:

[0092]

[0093] Calculate the entropy value of each parameter. Among them, the entropy value of the j-th index is calculated as:

[0094]

[0095] To ensure that the entropy value is positive, a negative sign is taken. Information entropy is a quantity used in information theory to characterize the degree of information redundancy. The larger the entropy value, the higher the degree of information disorder, and the corresponding higher the information efficiency.

[0096] The normalization coefficient is defined as:

[0097]

[0098] Use the entropy value to calculate the weight value of each parameter:

[0099]

[0100] Step 6: Input the training set generated in Step 5 into the second-layer random forest model of the ensemble learning model for training to obtain the final bearing fault diagnosis result.

[0101] The basic unit of the random forest model is a decision tree. Each decision tree is a classifier. For an input sample, N trees will have N classification results. By integrating all the classification voting results of the N trees, the category with the most votes is designated as the final output to complete the diagnosis of high-speed train bearing faults. The random forest randomly selects features and samples, making each tree in the forest have both similarities and differences.

[0102] Use the Bootstraping method to randomly sample m samples with replacement from the training set synthesized in Step 5, and perform n_tree samplings in total to generate n_tree training sets; train n_tree decision tree models for the n_tree training sets respectively. For a single decision tree model, assume the number of training sample features is n. Each time of splitting, the best feature is selected for splitting according to the Gini coefficient, and each tree is split like this until all the training examples at this node belong to the same class. Pruning is not required during the splitting process of the decision tree. The multiple generated decision trees are combined into a random forest. The final fault classification result is determined by voting of multiple tree classifiers.

[0103] Embodiment

[0104] The high-speed train bearing fault diagnosis method based on ensemble learning in this embodiment is specifically implemented according to the following steps:

[0105] Perform fault marking and CEEMDAN decomposition on the noisy original signal, as Figure 3As shown, the CEEMDAN algorithm is a method improved to address the mode mixing phenomenon in the decomposition process of the EEMD algorithm. It can achieve better separation of intrinsic mode functions, accurately reconstruct the original signal, and has a lower computational cost. After the signal is processed by CEEMDAN, the complex original signal is decomposed into a series of intrinsic mode components IMF. Each IMF component contains different frequency components. Using the CEEMDAN algorithm for denoising preprocessing of the signal can accurately separate the signal. The original signal is input into the CEEMDAN algorithm model. After decomposition, 17 IMF components and 1 Res residue are obtained. The 17 components are distributed in the order from high frequency to low frequency. IMF1 has the highest frequency and IMF17 has the lowest frequency. The 17 components respectively represent the signal components at each layer obtained after the signal is decomposed, laying a foundation for the next feature extraction. Since the bearing fault dataset is often not a signal with a mean of 0, there will always be a res residue at the end of its decomposition. This residue also contains a small amount of fault information, so it is retained. As Figure 4 shown, the instantaneous frequency distribution of the signal after CEEMDAN decomposition is 17 IMF components, and the signal vibration amplitude decreases in turn;

[0106] Comparing before and after wavelet threshold processing of the original signal, as shown in Fig. 5(a) and Fig. 5(b), after wavelet threshold processing of the original signal, useless signals are removed and the signal becomes sparse; comparing before and after wavelet threshold processing of the IMF1 component, as shown in Fig. 6(a) and Fig. 6(b), after wavelet threshold processing of the IMF1 component, useless signals are removed and the signal becomes sparse, which is beneficial for the next judgment;

[0107] Calculate the root mean square error RMSE of each IMF component. The root mean square error can well reflect the precision of the measurement. Therefore, the root mean square error can be used as a criterion for evaluating which IMF components to perform denoising processing on. After calculation, the RMSE values of the first 5 IMF components gradually decrease, and the RMSE values of the 6th IMF component to the last res component gradually increase. Therefore, take the first 5 IMF components before the 6th IMF for the next wavelet denoising processing. Denoise the high-frequency components and select the threshold. The specific implementation is as follows: Calculate that the RMSE values of the first five IMF are monotonically decreasing, then it is considered that they are high-frequency and contain interference signals. Therefore, perform wavelet denoising processing on the first 5 high-frequency IMF components. According to the low-frequency coefficients of the 6th to 17th layers of wavelet decomposition and the high-frequency coefficients of the 1st to 5th layers after quantization processing, perform wavelet reconstruction of the signal, and then extract the time-domain features of the vibration signal.

[0108] Reconstruct the obtained IMF components, extract features after obtaining the reconstructed signal; transmit the extracted features into the first-layer single models of the ensemble learning model respectively to obtain classification results; assign different weights to the accuracies of the first-layer single models of the ensemble learning model according to the classification results, and then integrate them into a training set; transmit the generated training set into the second-layer random forest model of the ensemble learning model for training to obtain the final bearing fault diagnosis result.

[0109] After calculation, the accuracies of each model are as shown in Table 1 below:

[0110] Table 1 Statistical table of the accuracy data of each model

[0111]

[0112] The first five models in the above table are the first-layer models of the invention, all of which are single models. The five models are carried out in parallel respectively to obtain five different results. According to the weights of the five results, the five results are integrated into a new training set and transmitted into the random forest model. The two-layer model is collectively called an ensemble model. As Figure 7 shown, the horizontal axis of the line chart is the model name, and the vertical axis is the model accuracy. Among the five single models, the SVM model has the lowest accuracy, and the LightGBM model has the highest accuracy. Among the 7 models, the ensemble learning model has the highest accuracy, indicating that the ensemble learning fault diagnosis effect is the best.

[0113] After experiments, the accuracies of the five single models (SVM, KNN, AdaBoost, XGBoost, LightGBM) are 73.2%, 76.5%, 83.5%, 87.3%, and 88.6% respectively, showing an increasing trend. The accuracy of the random forest is 88.2%, and the accuracy of the entire ensemble model is 97.6%, which has a significant improvement compared to the previous 6 models. The experimental results show that the ensemble learning method is well applicable to bearing fault diagnosis.

[0114] The weight assignment of each single model is as shown in Table 2 below:

[0115] Table 2 Weight assignment of single models

[0116] Model Name SVM KNN AdaBoost XGBoost LightGBM Weight 0.176 0.183 0.197 0.238 0.206

[0117] According to the fault diagnosis accuracies of the five single models, different weights are assigned to each model, and the different weights are integrated into a new training set.

[0118] The present invention adopts a method for denoising bearing vibration signals by combining CEEMDAN-wavelet threshold joint denoising. First, the signal is decomposed by CEEMDAN, and then the IMF components with more noise-containing signals are selected according to the RMSE value. These components are processed by wavelet threshold denoising. By separately processing and linearly adding the high-frequency components and low-frequency components, a reconstructed signal is obtained, which is a pure fault signal. The denoising accuracy of the original signal is improved, laying a foundation for the next step of fault diagnosis.

[0119] The integrated learning model of the present invention is divided into two layers. It adopts an integrated framework of hierarchical models for integrating several different algorithms. First, the initial data set is preprocessed and then several models are trained. The trained models serve as the first layer of the integrated learning model. Then, the outputs of each model in the first layer are combined as the input of the second layer of the integrated learning model and continue to be trained in the second layer, thus obtaining a complete integrated learning model. The model uses the classification diagnosis results of each different algorithm as the training data for the next layer of the model, combines the advantages of each model well, improves the classification accuracy. At the same time, since the importance of each feature of the original data is different for different algorithms, integrating several classifiers through the integrated learning model can learn the knowledge in the data more fully.

[0120] The high-speed train bearing fault diagnosis method based on integrated learning of the present invention adopts a two-layer fault detection model: in the first layer, five single models of SVM, KNN, AdaBoost, XGBoost, and LightGBM are used in parallel for fault diagnosis. In the second layer, the detection results of multiple single models are combined to generate a new training set, which is then input into the random forest model for secondary diagnosis. The two-layer fault diagnosis model can accurately and quickly diagnose faults. The signal denoising method in the present invention is applicable to non-stationary and non-linear axle random vibration signals, can accurately diagnose whether the bearing is faulty and the type of fault, provides convenience for the subsequent inspection and maintenance of workers, so as to be able to take necessary safeguard measures earlier, avoid more serious faults, and save manpower, material resources and financial resources.

Claims

1. A high-speed train bearing fault diagnosis method based on ensemble learning, characterized in that, The implementation is specifically carried out according to the following steps: Step 1: After obtaining the original vibration signal of the high-speed train bearing, conduct fault marking division, and then perform CEEMDAN decomposition to obtain a series of intrinsic mode function components IMF; Step 2: Denoise the IMF components obtained in Step 1; Step 3: Reconstruct the IMF components processed in Step 2, and extract features after obtaining the reconstructed signal; Step 4: Input the features extracted in Step 3 into the single models of the first layer of the ensemble learning model respectively to obtain classification results; Step 5: Calculate the classification accuracy of the single models of the first layer of the ensemble learning model according to the classification results in Step 4, assign weights to the model accuracies, and integrate them into a training set; Step 6: Input the training set generated in Step 5 into the random forest model of the second layer of the ensemble learning model for training to obtain the final bearing fault diagnosis result; The specific process of Step 3 is as follows: After quantization processing of the high-frequency IMF components processed in Step 2, high-frequency coefficients are obtained, and wavelet decomposition is used to process the low-frequency IMF components to obtain low-frequency coefficients. The low-frequency coefficients and high-frequency coefficients are linearly added for wavelet reconstruction of the signal, and the time-domain features of the vibration signal are extracted; The specific process of Step 4 is as follows: The ensemble learning model adopts the Stacking ensemble learning model. The single models of the first layer of the Stacking ensemble learning model include SVM, KNN, AdaBoost, XGBoost, and LightGBM. Input the features extracted in Step 3 into the five single models respectively to obtain five different classification results; In Step 5, different weights are assigned to the classification accuracies of the single models of the first layer, namely SVM, KNN, AdaBoost, XGBoost, and LightGBM, which are denoted as w1, w2, w3, w4, and w5 respectively; Use the information entropy model to construct a function and calculate the weight values of each parameter, and integrate them into a training set according to the weights.

2. The method for diagnosing high-speed train bearing faults based on ensemble learning according to claim 1, wherein The specific implementation of Step 1 is carried out according to the following steps: Step 1.1: Obtain the original noisy vibration signal of the high-speed train bearing, extract the bearing data features from the noisy original signal to obtain a bearing data feature dataset, mark the features of the normal bearing data, and then find the fault data from the full-life data, and divide the "fault" marks of the bearing fault data; When dividing the "fault" marks, they are successively divided according to the fault degree of the bearing from the outside to the inside into: the features of the outer ring mild fault bearing data, the features of the outer ring moderate fault bearing data, the features of the outer ring severe fault bearing data, the features of the inner ring mild fault bearing data, the features of the inner ring moderate fault bearing data, and the features of the inner ring severe fault bearing data; Step 1.2: Input the original signal marked with "fault" in Step 1.1 into the CEEMDAN algorithm model. After CEEMDAN processing, the signal is decomposed into several IMF components and a Res residue. Each IMF component corresponds to a different frequency component, and the several IMF components are distributed in the order of high frequency to low frequency of the frequency components.

3. The method for diagnosing high-speed train bearing faults based on ensemble learning according to claim 2, characterized in that, The specific implementation of Step 2 is carried out according to the following steps: Step 2.1: Calculate the RMSE values of 17 IMF components in the order from high frequency to low frequency using the root mean square error. When the RMSE values of the IMF components gradually increase and the IMF component before the increase is monotonically decreasing, the IMF component before the increase is the high-frequency IMF component and contains interference signals, and the remaining IMF components after the increase are low-frequency IMF components. Select the high-frequency IMF components for wavelet denoising. When performing wavelet denoising, a noisy model is expressed as: （1） where \(f(k)\) is the useful signal, \(s(k)\) is the noisy signal, \(e(k)\) is the noise, and \(\varepsilon\) is the standard deviation of the noise coefficient; Step 2.2: Perform threshold selection on the high-frequency IMF components after the denoising process in Step 2.1; The threshold selection uses hard threshold quantization to retain the local features of the bearing vibration signal edge.

4. The method for diagnosing high-speed train bearing faults based on ensemble learning according to claim 1, wherein The five single-model classification processes in Step 4 are as follows: SVM constructs multiple classifiers for fault diagnosis classification using the indirect method. During training, the samples of a certain category are successively grouped into one category, and the remaining samples are grouped into another category. During classification, the unknown samples are classified into the one with the maximum classification function value; When performing KNN classification, calculate the distance between the point to be classified and the points of the known categories, sort them in ascending order of distance, select the K points with the smallest distance to the point to be classified, determine the number of occurrences of the categories where the first K points are located, and return the category with the most occurrences of the first K points as the classification of the point to be classified. KNN classification only determines the category of the sample to be classified based on the category of one or several nearest samples in the class decision; When performing AdaBoost classification, use the idea of iteration. Only one weak classifier is trained in each iteration, and the trained weak classifier will be used in the next iteration; When performing XGBoost classification, the base learner used is the CART regression tree. An ensemble model is constructed by gradually adding trees. Assume that the model has a total of K trees integrated. The sum of the leaf node values corresponding to the K trees is the final classification result of the model. When solving the extreme value of the loss function, the Newton method is used, and the loss function is Taylor-expanded to the second order, and a regularization term is added to the loss function; When performing LightGBM classification, a model with a higher diagnostic rate is established through gradual optimization. The Histogram-based decision tree algorithm is used, the Leaf-wise leaf growth strategy with depth limit is adopted, the histogram difference is used for acceleration, categorical features are directly supported, the Cache hit rate optimization is adopted, the histogram-based sparse feature optimization is adopted, and multi-thread optimization is adopted.

5. The method for diagnosing high-speed train bearing faults based on ensemble learning according to claim 1, wherein The specific process of Step 5 is as follows: Step 5.1: Establish the mathematical model of the system. Assume that X is a known matrix, where represents the j-th index of the i-th evaluation object, construct the data matrix, eliminate the dimension of the data matrix X and perform normalization processing to obtain the matrix Y, （2） In formula (2), maxx*j and minx*j respectively represent the maximum and minimum values of the j-th column of the data matrix X, *j is the average value of the data matrix X, and any value in matrix Y is within [0, 1]; Step 5.2, the information entropy model establishes a weight matrix P with the bearing fault diagnosis accuracy rate as the evaluation index, where P j represents the weight of the j-th evaluation index, and the sum of P j is 1 and P j ≥0. Use the entropy value to construct a function and calculate the weight value of each parameter: Construct the function H for calculating the matrix Y. If the function H is symmetric, , when the order of the evaluation objects changes, the weights of the same evaluation index remain unchanged, that is, when any two rows of the matrix Y change, the value of the function remains unchanged; the function H is required to be monotonically increasing, continuous, and additive, so as to construct the function: （3） Calculate the entropy value of each parameter. Among them, the entropy value of the j-th index is calculated as: （4） Take the negative sign to ensure that the entropy value is positive. Information entropy is a quantity used in information theory to describe the degree of information redundancy. The larger the entropy value, the higher the degree of information disorder and the higher the corresponding information efficiency; The normalization coefficient is defined as: （5） Calculate the weight value of each parameter using the entropy value: （6）。 6. The method for diagnosing high-speed train bearing faults based on ensemble learning according to claim 1, wherein The basic unit of the random forest model is a decision tree. Each decision tree is a classifier. For an input sample, N trees will have N classification results. By integrating all the classification voting results of N trees, the category with the most votes is designated as the final output to complete the diagnosis of high-speed train bearing faults.

7. The method for diagnosing high-speed train bearing faults based on ensemble learning according to claim 6, characterized in that, In step 6, the Bootstraping method is used to randomly sample m samples with replacement from the training set synthesized in step 5, and sampling is performed n_tree times to generate n_tree training sets. For the n_tree training sets, n_tree decision tree models are trained respectively. For a single decision tree model, assuming the number of training sample features is n, the best feature is selected for splitting according to the Gini coefficient each time of splitting. Each tree keeps splitting like this until all the training examples of the current node belong to the same class. Pruning is not required during the splitting process of the decision tree. The multiple generated decision trees are combined into a random forest, and the final fault classification result is determined by the voting of multiple tree classifiers.

Images