CN114861719A - High-speed train bearing fault diagnosis method based on ensemble learning - Google Patents
High-speed train bearing fault diagnosis method based on ensemble learning Download PDFInfo
- Publication number
- CN114861719A CN114861719A CN202210466209.1A CN202210466209A CN114861719A CN 114861719 A CN114861719 A CN 114861719A CN 202210466209 A CN202210466209 A CN 202210466209A CN 114861719 A CN114861719 A CN 114861719A
- Authority
- CN
- China
- Prior art keywords
- model
- classification
- bearing
- frequency
- fault diagnosis
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
- G01M13/045—Acoustic or vibration analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
- G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
- G06F2218/10—Feature extraction by analysing the shape of a waveform, e.g. extracting parameters relating to peaks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Signal Processing (AREA)
- Acoustics & Sound (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
The invention discloses a high-speed train bearing fault diagnosis method based on ensemble learning, which is implemented according to the following steps: carrying out fault marking division after acquiring a noise-containing original signal, then carrying out CEEMDAN decomposition, and carrying out noise reduction processing on the obtained IMF component; performing IMF component reconstruction, and then performing feature extraction; respectively transmitting the extracted features into a first-layer single model of the ensemble learning model to obtain a classification result; and distributing different weights for the first layer of single model of the ensemble learning model according to the classification result, integrating the weights into a training set, transmitting the generated training set into a second layer of random forest model of the ensemble learning model for training, and obtaining a final bearing fault diagnosis result. The fault signal extracted by the method has high precision, the classification accuracy is improved, and the problem of low accuracy of the conventional high-speed train bearing fault diagnosis is solved.
Description
Technical Field
The 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
The rail transit is an important public transportation mode, has the characteristics of large transportation capacity and high speed, has a complex operation environment and large passenger capacity, and is directly related to the life safety of passengers once a fault occurs. The rolling bearing of the high-speed train is an important component of each mechanical system in the running part of the train and one of fault-prone components, supports an axle and bears the load between a wheel set and a train body, has important influence on the running safety of the high-speed train under the running condition, and is the key for ensuring the good running condition of the high-speed train. As a high-speed train needs to experience complex operating conditions such as curves, high speed, severe cold, high temperature and the like, the wheel set bearing is taken as a key bearing component of the train, can bear various impact loads in the long-term service process, and is easy to have fatigue damage and various performance degradation conditions. If the fault information of the wheel pair bearing cannot be timely and correctly detected, hot shafts, burning shafts and even shaft cutting can be caused, so that serious safety accidents such as train derailment and the like can be induced. With the increase of rail transit operation lines and daily passenger flow, the conventional regular maintenance cannot meet the requirements of timely and effective fault monitoring and diagnosis on the bearing in actual engineering due to low diagnosis efficiency, and the bearing diagnosis method needs to be improved. Therefore, the development of the wheel set bearing fault detection method has a particularly important significance for developing high-safety and high-reliability service high-speed trains.
The bearing fault diagnosis problem of the high-speed train is essentially a fault classification problem, and the bearing fault diagnosis method of the high-speed train adopted in the research is as follows: and a vibration signal acquisition device is arranged on the train, and time domain analysis is carried out on the fault vibration signal to judge the fault type of the bearing. The bearing fault diagnosis technology based on the vibration signals is adopted, but the running working condition of the high-speed train is complex, a large amount of noise can be mixed in the vibration signals of the rolling bearing of the train running part collected in the actual engineering, and the noise signals can influence the fault diagnosis effect of the bearing, so that the signals need to be subjected to noise reduction treatment before fault diagnosis is carried out. The research aims to design a bearing fault diagnosis algorithm based on vibration signals, and the bearing fault diagnosis algorithm is used for timely finding faults existing in train axles and ensuring the running safety of trains.
In the prior art, the effect is not ideal when bearing fault data are trained through a single model, diagnosis of different fault signals is unstable, intelligent fault diagnosis is carried out on a high-speed train bearing by using an advanced machine learning technology, and the high-reliability intelligent fault diagnosis method has high practical value. The theoretical significance of the research lies in that the integrated learning method can be applied to bearing fault diagnosis, is high in diagnosis precision and good in classification effect, and is one of the main development directions of future fault diagnosis technologies.
Disclosure of Invention
The invention aims to provide a high-speed train bearing fault diagnosis method based on integrated learning, and solves the problem that the existing high-speed train bearing fault diagnosis accuracy is low.
The technical scheme adopted by the invention is that the high-speed train bearing fault diagnosis method based on integrated learning is implemented according to the following steps:
and 6, transmitting the training set generated in the step 5 into a second-layer random forest model of the integrated learning model for training to obtain a final bearing fault diagnosis result.
The present invention is also characterized in that,
the step 1 is implemented according to the following steps:
step 1.1, acquiring a vibration noise-containing original signal of a high-speed train bearing, carrying out bearing data feature extraction on the noise-containing original signal to obtain a bearing data feature data set, marking the features of normal bearing data, finding out fault data from life-cycle data, and dividing the fault marks of the bearing fault data; when the 'failure' mark is divided, the bearing is divided into the following parts in sequence according to the failure degree of the bearing from outside to inside: the method comprises the following steps of (1) characteristics of outer ring light fault bearing data, characteristics of outer ring medium fault bearing data, characteristics of outer ring serious fault bearing data, characteristics of inner ring light fault bearing data, characteristics of inner ring medium fault bearing data and characteristics of inner ring serious fault bearing data;
step 1.2, transmitting the original signal marked with 'failure' in the step 1.1 into a CEEMDAN algorithm model, decomposing the signal after CEEMDAN processing to obtain a plurality of IMF components and Res margins, wherein each IMF component corresponds to different frequency components, and distributing the plurality of IMF components according to the sequence from high frequency to low frequency of the frequency components.
The step 2 is implemented according to the following steps:
step 2.1, calculating RMSE values of 17 IMF components according to the sequence from high frequency to low frequency by adopting root mean square error, wherein when the RMSE values of the IMF components gradually increase and the IMF components before increasing decrease monotonically, the IMF components before increasing are high-frequency IMF components and contain interference signals, the rest IMF components after increasing are low-frequency IMF components, and taking the high-frequency IMF components to perform wavelet denoising treatment; in the wavelet denoising process, a noisy model is represented as:
wherein f (k) is a useful signal, s (k) is a noisy signal, e (k) is noise, and ε is the standard deviation of the noise coefficient;
2.2, selecting a threshold value for the high-frequency IMF component subjected to denoising processing in the step 2.1;
threshold selection uses hard threshold quantization to preserve local features of the bearing vibration signal edges.
The specific process of the step 3 is as follows: and (3) quantizing the high-frequency IMF component processed in the step (2) to obtain a high-frequency coefficient, processing the low-frequency IMF component by adopting wavelet decomposition to obtain a low-frequency coefficient, linearly adding the low-frequency coefficient and the high-frequency coefficient to perform wavelet reconstruction of a signal, and extracting the time domain characteristic of the vibration signal.
The specific process of the step 4 is as follows: the ensemble learning model adopts a Stacking ensemble learning model, a first layer single model of the Stacking ensemble learning model comprises an SVM, a KNN, an AdaBoost, an XGBoost and a LightGBM, and the features extracted in the step 3 are respectively transmitted into five single models to obtain five different classification results.
The five single model classification processes in the step 4 are as follows:
the SVM constructs a plurality of classifiers by adopting an indirect method to classify fault diagnosis, samples of a certain class are classified into one class during training, other remaining samples are classified into another class, and unknown samples are classified into the class with the maximum classification function value during classification;
when the classification of the KNN is carried out, the distance between the point to be classified and the point of the known class is calculated, the K points with the minimum distance from the point to be classified are selected according to the ascending order of the distance, the occurrence frequency of the class of the front K points is determined, the class with the maximum occurrence frequency of the front K points is used as the classification of the point to be classified, and the classification of the sample to be classified is determined according to the classification of the nearest sample or samples on the classification decision;
when AdaBoost is classified, an iterative idea is adopted, only one weak classifier is trained in each iteration, and the trained weak classifier participates in the use of the next iteration;
during XGboost classification, a used base learner is a CART regression tree, an integrated model is constructed by gradually adding trees, K trees are assumed to be integrated together by the model, the sum of leaf node values corresponding to the K trees is a final classification result of the model, a Newton method is used when a loss function extreme value is solved, a loss function Taylor is expanded to the second order, and a regularization term is added into the loss function;
during LightGBM classification, a model with higher diagnosis rate is established through gradual optimization, a decision tree algorithm based on Histogram is adopted, a Leaf growth strategy of Leaf-wise with depth limitation is adopted, difference acceleration is carried out by using a Histogram, category characteristics are directly supported, Cache hit rate optimization is adopted, sparse characteristic optimization based on the Histogram is adopted, and multithreading optimization is adopted.
The specific process of the step 5 is as follows:
step 5.1, establishing a mathematical model of the system, assuming that X is a known matrix which represents the jth index of the ith evaluation object, constructing a data matrix, eliminating dimension of the data matrix X and carrying out normalization processing on the data matrix X to obtain a matrix Y,
in the formula (2), maxx j and minx j represent the maximum value and the minimum value of the jth column of the data matrix X respectively,for the average value of the data matrix X, any value in the matrix Y is [0,1 ]]Inner;
step 5.2, the information entropy model establishes a weight matrix P by taking the bearing fault diagnosis accuracy as an evaluation index, and then P j Weight, P, representing the jth evaluation index j Is 1 and P j And (3) more than or equal to 0, constructing a function by utilizing an entropy value and calculating a weight value of each parameter:
constructing a function H of the calculation matrix Y, wherein the symmetry of the function H is H (x) 1 ,x 2 )=H(x 2 ,x 1 ) When the order of the evaluation objects is changed, the weight of the same evaluation index is unchanged, namely when any two rows of the calculation matrix Y are changed, the value of the function is kept unchanged; the function H requires monotonic increase, continuity, additivity, to construct the function:
calculating the entropy value of each parameter, wherein the entropy value of the j index is calculated as:
in order to ensure that the entropy value is a positive negative sign, the information entropy is a quantity used for describing the information unnecessary degree in the information theory, and the larger the entropy value is, the higher the disorder degree of the information is, and the higher the corresponding information efficiency is;
the normalized coefficient is defined as:
calculating the weight value of each parameter by using the entropy value:
the basic unit of the random forest model is a decision tree, each decision tree is a classifier, N classification results are obtained for N trees for input samples, all classification voting results of the N trees are integrated, and the class with the largest voting times is designated as final output to finish diagnosis of the bearing fault of the high-speed train.
The invention has the beneficial effects that: the invention adopts a CEEMDAN wavelet threshold value-based combined denoising method to jointly process the bearing vibration signal denoising, separately processes the high-frequency component and the low-frequency component, improves the denoising precision of the original signal, has high precision of the extracted fault signal, obtains a purer bearing fault signal, adopts a Stacking integrated learning model to preprocess the initial data set and train a plurality of models, then combines different classification diagnosis results output by each model of the first layer as the input of the second layer and continues training on the second layer, improves the classification accuracy, diagnoses whether the bearing is in fault and the fault type more accurately, and solves the problem of low accuracy of the bearing fault diagnosis of the existing high-speed train.
Drawings
FIG. 1 is a schematic flow chart of a high-speed train bearing fault diagnosis method based on integrated learning according to the invention;
FIG. 2 is a wavelet threshold denoising flow chart of the high-speed train bearing fault diagnosis method based on ensemble learning according to the invention;
FIG. 3 is a time sequence chart of the CEEMDAN method for diagnosing the bearing fault of the high-speed train based on the integrated learning according to the present invention;
FIG. 4 is an instantaneous frequency chart of the CEEMDAN decomposed high-speed train bearing fault diagnosis method based on the integrated learning;
FIG. 5(a) is a curve diagram before wavelet thresholding is carried out on an original signal by the high-speed train bearing fault diagnosis method based on ensemble learning according to the invention;
FIG. 5(b) is a graph after wavelet thresholding is performed on an original signal by the high-speed train bearing fault diagnosis method based on ensemble learning according to the invention;
FIG. 6(a) is a graph before wavelet thresholding is carried out on an IMF1 component by the high-speed train bearing fault diagnosis method based on integrated learning according to the invention;
FIG. 6(b) is a graph after the wavelet thresholding is performed on the IMF1 component by the high-speed train bearing fault diagnosis method based on the ensemble learning according to the invention;
FIG. 7 is a comparison graph of the accuracy of classification results of models of the high-speed train bearing fault diagnosis method based on ensemble learning.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention discloses a high-speed train bearing fault diagnosis method based on integrated learning, which has a flow shown in figure 1 and is implemented according to the following steps:
the step 1 is implemented according to the following steps:
step 1.1, acquiring a vibration noise-containing original signal of a high-speed train bearing, carrying out bearing data feature extraction on the noise-containing original signal to obtain a bearing data feature data set, marking the features of normal bearing data, finding out fault data from life-cycle data, and dividing the fault marks of the bearing fault data;
when the 'failure' mark is divided, the bearing is divided into the following parts in sequence according to the failure degree of the bearing from outside to inside: the characteristic of the bearing data with slight fault on the outer ring, the characteristic of the bearing data with moderate fault on the outer ring, the characteristic of the bearing data with serious fault on the outer ring, the characteristic of the bearing data with slight fault on the inner ring, the characteristic of the bearing data with moderate fault on the inner ring and the characteristic of the bearing data with serious fault on the inner ring.
And step 1.2, transmitting the original signal marked with the fault in the step 1.1 into a CEEMDAN algorithm model, and decomposing the signal after the CEEMDAN processing to obtain a plurality of IMF components and Res margins. Each IMF component corresponds to different frequency components, and a plurality of IMF components are distributed according to the sequence from high frequency to low frequency of the frequency components.
the step 2 is implemented according to the following steps:
step 2.1, calculating RMSE values of a plurality of IMF components according to the sequence from high frequency to low frequency by adopting root mean square error, wherein when the RMSE values of the IMF components gradually increase and the IMF components before increasing decrease monotonically, the IMF components before increasing are high-frequency IMF components and contain interference signals, and the rest IMF components after increasing are low-frequency IMF components, and taking the high-frequency IMF components to perform wavelet denoising treatment;
the root mean square error reflects the precision of the measurement as a criterion for assessing which IMF components are to be denoised. The wavelet de-noising processing integrates the feature extraction and the low-pass filtering, and the wavelet de-noising reserves the feature extraction part and is superior to the traditional de-noising method in performance.
In the wavelet denoising process, a noisy model is represented as:
s(k)=f(k)+ε*e(k)k=0,1......n-1 (1)
where f (k) is the desired signal, s (k) is the noisy signal, e (k) is the noise, and ε is the standard deviation of the noise figure.
2.2, selecting a threshold value for the high-frequency IMF component subjected to denoising processing in the step 2.1;
the threshold selection adopts hard threshold quantization to reserve the local characteristics of the edge of the bearing vibration signal and avoid the phenomenon of fuzzy distortion of the edge.
the specific process of the step 3 is as follows: and (3) quantizing the high-frequency IMF component processed in the step (2) to obtain a high-frequency coefficient, processing the low-frequency IMF component by adopting wavelet decomposition to obtain a low-frequency coefficient as shown in fig. 2, linearly adding the low-frequency coefficient and the high-frequency coefficient to perform wavelet reconstruction of a signal, and extracting the time domain characteristic of the vibration signal.
The wavelet de-noising is to remove wavelet coefficients corresponding to noise on each frequency band according to the characteristic that wavelet decomposition coefficients of the noise and signals on different frequency bands have different intensity distributions; step 3, preserving wavelet decomposition coefficients of the original signals; and performing wavelet reconstruction on the processed coefficient to obtain a pure signal. The quantization process approximates a continuous value of the signal to a finite number of discrete values.
Extracting the time domain features of the vibration signal comprises: the method comprises the following steps of (1) dimensional and dimensionless, wherein the dimensional comprises a mean value, a root mean square value, a square root amplitude value, an absolute mean value, a skewness, a kurtosis, a variance, a maximum value, a minimum value and a peak-to-peak value; dimensionless includes waveform index, peak index, pulse index, margin index, skewness index, kurtosis index.
the specific process of the step 4 is as follows: the ensemble learning model adopts a Stacking ensemble learning model, a first layer single model of the Stacking ensemble learning model comprises an SVM, a KNN, an AdaBoost, an XGBoost and a LightGBM, and the characteristics extracted in the step 3 are respectively transmitted into five single models to obtain five different classification results:
the SVM adopts an indirect method to construct a plurality of classifiers to classify fault diagnosis, samples of a certain class are sequentially classified into one class during training, other remaining samples are classified into another class, and unknown samples are classified into the class with the maximum classification function value during classification.
And during KNN classification, calculating the distance between the point to be classified and the point of the known class, sorting according to the distance increasing order, selecting the K points with the minimum distance from the point to be classified, determining the occurrence frequency of the class where the previous K points are located, and taking the class with the maximum occurrence frequency of the previous K points as the classification of the point to be classified. The KNN classification only determines the category of the sample to be classified according to the category of the nearest sample or a plurality of samples in the classification decision.
When AdaBoost is classified, the idea of iteration is adopted, only one weak classifier is trained in each iteration, and the trained weak classifier participates in the use of the next iteration.
And in the XGboost classification, the used base learner is a CART regression tree, and an integration model is constructed by gradually adding trees. And assuming that K trees are integrated into the model, wherein the sum of leaf node values corresponding to the K trees is the final classification result of the model. And when solving the extreme value of the loss function, using a Newton method to expand the Taylor of the loss function to the second order, and adding a regularization term into the loss function.
During LightGBM classification, a model with higher diagnosis rate is established through gradual optimization, a decision tree algorithm based on Histogram is adopted, a Leaf growth strategy of Leaf-wise with depth limitation is adopted, difference acceleration is carried out by using a Histogram, category characteristics are directly supported, Cache hit rate optimization is adopted, sparse characteristic optimization based on the Histogram is adopted, and multithreading optimization is adopted.
Step 5.1, establishing a mathematical model of the system, assuming that X is a known matrix which represents the jth index of the ith evaluation object, constructing a data matrix, eliminating dimension of the data matrix X and carrying out normalization processing on the data matrix X to obtain a matrix Y,
in the formula (2), maxx j and minx j represent the maximum value and the minimum value of the jth column of the data matrix X respectively,for the average value of the data matrix X, any value in the matrix Y is [0,1 ]]Internal;
step 5.2, the information entropy model establishes a weight matrix P by taking the bearing fault diagnosis accuracy as an evaluation index, and then P j Weight, P, representing the jth evaluation index j Is 1 and P j And (3) more than or equal to 0, constructing a function by utilizing an entropy value and calculating a weight value of each parameter:
constructing a function H of the calculation matrix Y, wherein the symmetry of the function H is H (x) 1 ,x 2 )=H(x 2 ,x 1 ) When the order of the evaluation objects is changed, the weight of the same evaluation index is unchanged, namely when any two rows of the calculation matrix Y are changed, the value of the function is kept unchanged; the function H requires monotonic increase, continuity, additivity, to construct the function:
calculating the entropy value of each parameter, wherein the entropy value of the j index is calculated as:
in order to ensure that the entropy value is a positive negative sign, the information entropy is a quantity used for describing the information unnecessary degree in the information theory, and the larger the entropy value is, the higher the disorder degree of the information is, and the higher the corresponding information efficiency is.
The normalized coefficient is defined as:
calculating the weight value of each parameter by using the entropy value:
and 6, transmitting the training set generated in the step 5 into a second-layer random forest model of the integrated learning model for training to obtain a final bearing fault diagnosis result.
The basic unit of the random forest model is a decision tree, each decision tree is a classifier, N classification results are obtained for N trees for input samples, all classification voting results of the N trees are integrated, and the class with the largest voting times is designated as final output to finish diagnosis of the bearing fault of the high-speed train. And randomly taking characteristics and samples from the random forest, so that each tree in the forest has similarity and difference.
Randomly selecting m samples from the training sets synthesized in the step 5 by using a bootstrapping method, and carrying out n _ tree times of sampling to generate n _ tree training sets; respectively training n _ tree decision tree models for n _ tree training sets, assuming the feature number of training samples to be n for a single decision tree model, selecting the best feature to split according to the Keyney coefficient during each splitting, splitting each tree in such a way until all the training samples of the node belong to the same class, and forming a random forest by using a plurality of generated decision trees without pruning in the splitting process of the decision trees. And voting according to a plurality of tree classifiers to determine a final fault classification result.
Examples
The method for diagnosing the bearing fault of the high-speed train based on the integrated learning is implemented according to the following steps:
the noise-containing original signal is subjected to fault marking and CEEMDAN decomposition, as shown in FIG. 3, the CEEMDAN algorithm is an improved method aiming at the modal aliasing phenomenon existing in the decomposition process of the EEMD algorithm, can complete better intrinsic mode function separation, accurately reconstructs the original signal and has lower operation cost. After the signal is processed by CEEMDAN, a complex original signal can be decomposed into a series of intrinsic mode components IMF, each IMF component comprises different frequency components, the signal can be accurately separated by denoising and preprocessing the signal by adopting a CEEMDAN algorithm, the original signal is transmitted into a CEEMDAN algorithm model, 17 IMF components and 1 Res margin are obtained by decomposition, the 17 components are distributed according to the sequence from high frequency to low frequency, the IMF1 frequency is the highest, and the IMF17 frequency is the lowest. The 17 components respectively represent each layer of signal components obtained after the signals are decomposed, so that a foundation is laid for next step of feature extraction, and a res margin always exists after the decomposition is finished because a bearing fault data set is not the signal with the average value of 0, and the margin also contains a small amount of fault information, so that the margin is reserved. As shown in fig. 4, the instantaneous frequency distribution of the signal after CEEMDAN decomposition is 17 IMF components, and the vibration amplitude of the signal decreases in turn;
comparing the original signals before and after wavelet threshold processing, as shown in fig. 5(a) and 5(b), removing useless signals after the wavelet threshold processing is carried out on the original signals, and making the signals sparse; comparing the IMF1 component before and after wavelet threshold processing, as shown in fig. 6(a) and fig. 6(b), removing useless signals after wavelet threshold processing is carried out on the IMF1 component, and enabling signals to be sparse, so that the judgment of the next step is facilitated;
the root mean square error RMSE for each IMF component is calculated, which is a good reflection of the precision of the measurement. The root mean square error can be used as a criterion to assess which IMF components are denoised. The calculated values of the RMSE of the first 5 IMF components are gradually decreased, and the RMSE of the 6 th IMF component to the last res component is gradually increased, so that the 5 IMF components before the 6 th IMF are taken for the next wavelet denoising treatment. Denoising the high-frequency component and selecting a threshold, which is specifically implemented according to the following steps: calculating the RMSE values of the first five IMFs to be monotonously reduced, and considering the RMSE values to be high frequency and contain interference signals, so that wavelet denoising processing is carried out on the first 5 high frequency IMF components, wavelet reconstruction is carried out on signals according to low-frequency coefficients from the 6 th layer to the 17 th layer of wavelet decomposition and high-frequency coefficients from the 1 st layer to the 5 th layer after quantization processing, and then time domain characteristics of vibration signals are extracted.
Reconstructing the obtained IMF component to obtain a reconstructed signal and then extracting the characteristics; respectively transmitting the extracted features into a first-layer single model of the ensemble learning model to obtain a classification result; distributing different weights for the accuracy of the first layer single model of the ensemble learning model according to the classification result, and integrating the weights into a training set; and transmitting the generated training set into a second layer of random forest model of the integrated learning model for training to obtain a final bearing fault diagnosis result.
The accuracy of each model is calculated as follows in table 1:
TABLE 1 statistical table of accuracy data of each model
The first five models in the upper table are the first-layer models of the invention, which are all single models, the five models are respectively carried out in parallel to obtain five different results, the five results are integrated into a new training set according to the weights of the five results and are transmitted into a random forest model, and the two layers of models are collectively called as integrated models. As shown in fig. 7, the horizontal axis of the line graph is the model name, the vertical axis is the model accuracy, among the five single models, the SVM model has the lowest accuracy, the LightGBM model has the highest accuracy, and among the 7 models, the ensemble learning model has the highest accuracy, which indicates that the best ensemble learning fault diagnosis effect is obtained.
Through experiments, the accuracy of five single models (SVM, KNN, AdaBoost, XGboost and LightGBM) is respectively as follows: 73.2%, 76.5%, 83.5%, 87.3% and 88.6%, presenting increasing trend, wherein the random forest accuracy rate is 88.2%, the accuracy rate of the whole integrated model is 97.6%, and the integrated learning method is greatly improved compared with the first 6 models, and experimental results show that the integrated learning method is better suitable for bearing fault diagnosis.
The weight assignment for each single model is as follows in table 2:
TABLE 2 weight assignment for single model
Model name | SVM | KNN | AdaBoost | XGBoost | LightGBM |
Weight of | 0.176 | 0.183 | 0.197 | 0.238 | 0.206 |
And according to the fault diagnosis accuracy of the five single models, different weights are distributed to each model, and the different weights are integrated into a new training set.
According to the method, a CEEMDAN-wavelet threshold combined denoising and joint processing bearing vibration signal denoising method is adopted, CEEMDAN decomposition is firstly carried out on signals, IMF components with more noise signals are selected according to RMSE values, wavelet threshold denoising processing is carried out on the components, high-frequency components and low-frequency components are processed separately and linearly added, and a reconstructed signal is obtained, wherein the signal is a pure fault signal. The denoising precision of the original signal is improved, and a foundation is laid for the next fault diagnosis.
The ensemble learning model is divided into two layers, a layered model integration framework with the ensemble learning model for fusing different algorithms is adopted, firstly, an initial data set is preprocessed, then a plurality of models are trained, the trained models serve as a first layer of the ensemble learning model, then outputs of all models of the first layer are combined to serve as inputs of a second layer of the ensemble learning model, and training is continued on the second layer, so that a complete ensemble learning model is obtained. The model uses the classification diagnosis result of each different algorithm as the training data of the next layer of the model, the advantages of each model are well combined, the classification accuracy is improved, and meanwhile, because the importance of each characteristic of the original data to different algorithms is different, several classifiers are integrated through an integrated learning model, so that the knowledge in the data can be more fully learned.
The invention discloses a high-speed train bearing fault diagnosis method based on integrated learning, which adopts two layers of fault detection models: the first layer adopts five single models of SVM, KNN, AdaBoost, XGboost and LightGBM to perform fault diagnosis in parallel, the second layer combines the detection results of a plurality of single models to generate a new training set, and then the new training set is transmitted to a random forest model to perform secondary diagnosis. The two-layer fault diagnosis model can accurately and quickly diagnose faults. The signal denoising method is suitable for non-stable and non-linear axle random vibration signals, can diagnose whether the bearing has faults and fault types accurately, and provides convenience for the follow-up staff to check and maintain, so that necessary guarantee measures can be carried out in advance, more serious faults are avoided, and manpower, material resources and financial resources are saved.
Claims (10)
1. The high-speed train bearing fault diagnosis method based on integrated learning is characterized by comprising the following steps:
step 1, acquiring a vibration original signal of a high-speed train bearing, then dividing a fault mark, and then performing CEEMDAN decomposition to obtain a series of intrinsic modal components IMF;
step 2, performing noise reduction processing on the IMF component obtained in the step 1;
step 3, reconstructing the IMF component processed in the step 2 to obtain a reconstructed signal and then extracting characteristics;
step 4, respectively transmitting the features extracted in the step 3 into a first layer of single model of the ensemble learning model to obtain a classification result;
step 5, calculating the classification accuracy of the first layer of single models of the ensemble learning model according to the classification result of the step 4, and integrating the models into a training set after weights are distributed to the model accuracy;
and 6, transmitting the training set generated in the step 5 into a second-layer random forest model of the integrated learning model for training to obtain a final bearing fault diagnosis result.
2. The integrated learning-based high-speed train bearing fault diagnosis method according to claim 1, wherein the step 1 is specifically implemented according to the following steps:
step 1.1, acquiring a vibration noise-containing original signal of a high-speed train bearing, carrying out bearing data feature extraction on the noise-containing original signal to obtain a bearing data feature data set, marking the features of normal bearing data, finding out fault data from life-cycle data, and dividing the fault marks of the bearing fault data; when the 'failure' mark is divided, the bearing is divided into the following parts in sequence according to the failure degree of the bearing from outside to inside: the method comprises the following steps of (1) characteristics of outer ring light fault bearing data, characteristics of outer ring medium fault bearing data, characteristics of outer ring serious fault bearing data, characteristics of inner ring light fault bearing data, characteristics of inner ring medium fault bearing data and characteristics of inner ring serious fault bearing data;
step 1.2, transmitting the original signal marked with 'failure' in the step 1.1 into a CEEMDAN algorithm model, decomposing the signal after CEEMDAN processing to obtain a plurality of IMF components and Res margins, wherein each IMF component corresponds to different frequency components, and distributing the plurality of IMF components according to the sequence from high frequency to low frequency of the frequency components.
3. The integrated learning-based high-speed train bearing fault diagnosis method according to claim 2, wherein the step 2 is implemented specifically according to the following steps:
step 2.1, calculating RMSE values of 17 IMF components according to the sequence from high frequency to low frequency by adopting root mean square error, wherein when the RMSE values of the IMF components gradually increase and the IMF components before increasing decrease monotonically, the IMF components before increasing are high-frequency IMF components and contain interference signals, the rest IMF components after increasing are low-frequency IMF components, and taking the high-frequency IMF components to perform wavelet denoising treatment; in the wavelet denoising process, a noisy model is represented as:
s(k)=f(k)+ε*e(k)k=0,1......n-1 (1)
wherein f (k) is a useful signal, s (k) is a noisy signal, e (k) is noise, and ε is the standard deviation of the noise coefficient;
2.2, selecting a threshold value for the high-frequency IMF component subjected to denoising processing in the step 2.1;
threshold selection uses hard threshold quantization to preserve local features of the bearing vibration signal edges.
4. The integrated learning-based high-speed train bearing fault diagnosis method according to claim 3, wherein the specific process of the step 3 is as follows: and (3) quantizing the high-frequency IMF component processed in the step (2) to obtain a high-frequency coefficient, processing the low-frequency IMF component by adopting wavelet decomposition to obtain a low-frequency coefficient, linearly adding the low-frequency coefficient and the high-frequency coefficient to perform wavelet reconstruction of a signal, and extracting the time domain characteristic of the vibration signal.
5. The integrated learning-based high-speed train bearing fault diagnosis method according to claim 1, wherein the specific process in the step 4 is as follows: the ensemble learning model adopts a Stacking ensemble learning model, a first layer single model of the Stacking ensemble learning model comprises an SVM, a KNN, an AdaBoost, an XGBoost and a LightGBM, and the features extracted in the step 3 are respectively transmitted into five single models to obtain five different classification results.
6. The integrated learning-based high-speed train bearing fault diagnosis method according to claim 5, wherein the five single model classification processes of the step 4 are as follows:
the SVM constructs a plurality of classifiers by adopting an indirect method to classify fault diagnosis, samples of a certain class are classified into one class during training, other remaining samples are classified into another class, and unknown samples are classified into the class with the maximum classification function value during classification;
when the classification of the KNN is carried out, the distance between the point to be classified and the point of the known class is calculated, the K points with the minimum distance from the point to be classified are selected according to the ascending order of the distance, the occurrence frequency of the class of the front K points is determined, the class with the maximum occurrence frequency of the front K points is used as the classification of the point to be classified, and the classification of the sample to be classified is determined according to the classification of the nearest sample or samples on the classification decision;
when AdaBoost is classified, an iterative idea is adopted, only one weak classifier is trained in each iteration, and the trained weak classifier participates in the use of the next iteration;
during XGboost classification, a used base learner is a CART regression tree, an integrated model is constructed by gradually adding trees, K trees are assumed to be integrated together by the model, the sum of leaf node values corresponding to the K trees is a final classification result of the model, a Newton method is used when a loss function extreme value is solved, a loss function Taylor is expanded to the second order, and a regularization term is added into the loss function;
during LightGBM classification, a model with higher diagnosis rate is established through gradual optimization, a decision tree algorithm based on Histogram is adopted, a Leaf growth strategy of Leaf-wise with depth limitation is adopted, difference acceleration is carried out by using a Histogram, category characteristics are directly supported, Cache hit rate optimization is adopted, sparse characteristic optimization based on the Histogram is adopted, and multithreading optimization is adopted.
7. The integrated learning-based high-speed train bearing fault diagnosis method as claimed in claim 5, wherein the step 5 assigns different weights to the classification accuracy of the first layer of single model SVM, KNN, AdaBoost, XGboost and LightGBM, which are respectively denoted as w1, w2, w3, w4 and w 5; and constructing a function by using the information entropy model, calculating the weight value of each parameter, and integrating the functions into a training set according to the weights.
8. The integrated learning-based high-speed train bearing fault diagnosis method according to claim 7, wherein the specific process of the step 5 is as follows:
step 5.1, establishing a mathematical model of the system, assuming that X is a known matrix which represents the jth index of the ith evaluation object, constructing a data matrix, eliminating dimension of the data matrix X and carrying out normalization processing on the data matrix X to obtain a matrix Y,
in the formula (2), maxx j and minx j represent the maximum value and the minimum value of the jth column of the data matrix X respectively,the average value of the data matrix X is any value in the matrix Y [0,1 ]]Internal;
step 5.2, the information entropy model establishes a weight matrix P by taking the bearing fault diagnosis accuracy as an evaluation index, and then P j Weight, P, representing the jth evaluation index j Is 1 and P j And (3) more than or equal to 0, constructing a function by utilizing an entropy value and calculating a weight value of each parameter:
constructing a function H of the calculation matrix Y, wherein the symmetry of the function H is H (x) 1 ,x 2 )=H(x 2 ,x 1 ) When the order of the evaluation objects is changed, the weight of the same evaluation index is unchanged, namely when any two rows of the calculation matrix Y are changed, the value of the function is kept unchanged; the function H requires monotonic increase, continuity, additivity, to construct the function:
calculating the entropy value of each parameter, wherein the entropy value of the j index is calculated as:
in order to ensure that the entropy value is a positive negative sign, the information entropy is a quantity used for describing the information unnecessary degree in the information theory, and the larger the entropy value is, the higher the disorder degree of the information is, and the higher the corresponding information efficiency is;
the normalized coefficient is defined as:
calculating the weight value of each parameter by using the entropy value:
9. the integrated learning-based high-speed train bearing fault diagnosis method according to claim 1, wherein a basic unit of a random forest model is a decision tree, each decision tree is a classifier, for input samples, N classification results are available for N trees, and by integrating all classification voting results of the N trees, the class with the largest voting number is designated as a final output, so that the high-speed train bearing fault diagnosis is completed.
10. The integrated learning-based high-speed train bearing fault diagnosis method according to claim 9, wherein m samples are selected from the training set synthesized in the step 5 by using a bootstrapping method at random in the step 6, and n _ tree times of sampling are performed in total to generate n _ tree training sets; respectively training n _ tree decision tree models for n _ tree training sets, assuming the feature number of training samples to be n for a single decision tree model, selecting the best feature to split according to the Keyny coefficient during each splitting, splitting each tree in such a way until all the training samples of the node belong to the same class, forming a random forest by a plurality of generated decision trees without pruning in the splitting process of the decision trees, and voting according to a plurality of tree classifiers to determine the final fault classification result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210466209.1A CN114861719A (en) | 2022-04-29 | 2022-04-29 | High-speed train bearing fault diagnosis method based on ensemble learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210466209.1A CN114861719A (en) | 2022-04-29 | 2022-04-29 | High-speed train bearing fault diagnosis method based on ensemble learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114861719A true CN114861719A (en) | 2022-08-05 |
Family
ID=82635392
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210466209.1A Pending CN114861719A (en) | 2022-04-29 | 2022-04-29 | High-speed train bearing fault diagnosis method based on ensemble learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114861719A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115545086A (en) * | 2022-11-07 | 2022-12-30 | 海纳科德(湖北)科技有限公司 | Migratable feature automatic selection acoustic diagnosis method and system |
CN115828140A (en) * | 2022-12-13 | 2023-03-21 | 中国民航大学 | Neighborhood mutual information and random forest fusion fault detection method, system and application |
CN117435981A (en) * | 2023-12-22 | 2024-01-23 | 四川泓宝润业工程技术有限公司 | Method and device for diagnosing operation faults of machine pump equipment, storage medium and electronic equipment |
-
2022
- 2022-04-29 CN CN202210466209.1A patent/CN114861719A/en active Pending
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115545086A (en) * | 2022-11-07 | 2022-12-30 | 海纳科德(湖北)科技有限公司 | Migratable feature automatic selection acoustic diagnosis method and system |
CN115545086B (en) * | 2022-11-07 | 2023-04-18 | 海纳科德(湖北)科技有限公司 | Migratable feature automatic selection acoustic diagnosis method and system |
CN115828140A (en) * | 2022-12-13 | 2023-03-21 | 中国民航大学 | Neighborhood mutual information and random forest fusion fault detection method, system and application |
CN115828140B (en) * | 2022-12-13 | 2024-04-09 | 中国民航大学 | Method, system and application for detecting fault by fusing neighborhood mutual information and random forest |
CN117435981A (en) * | 2023-12-22 | 2024-01-23 | 四川泓宝润业工程技术有限公司 | Method and device for diagnosing operation faults of machine pump equipment, storage medium and electronic equipment |
CN117435981B (en) * | 2023-12-22 | 2024-03-01 | 四川泓宝润业工程技术有限公司 | Method and device for diagnosing operation faults of machine pump equipment, storage medium and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN114861719A (en) | High-speed train bearing fault diagnosis method based on ensemble learning | |
CN111369563B (en) | Semantic segmentation method based on pyramid void convolutional network | |
CN110057584B (en) | Degradation monitoring method for locomotive traction motor bearing | |
CN106895975B (en) | Method for Bearing Fault Diagnosis based on Stacked SAE deep neural network | |
CN103728551B (en) | A kind of analog-circuit fault diagnosis method based on cascade integrated classifier | |
CN111222458A (en) | Rolling bearing fault diagnosis method based on ensemble empirical mode decomposition and convolutional neural network | |
CN110827260B (en) | Cloth defect classification method based on LBP characteristics and convolutional neural network | |
CN111161814A (en) | DRGs automatic grouping method based on convolutional neural network | |
CN111126868B (en) | Road traffic accident occurrence risk determination method and system | |
CN112163640A (en) | Planetary gearbox fault diagnosis method and system based on deep convolution migration learning | |
CN103048329A (en) | Pavement crack detecting method based on active contour model | |
CN113704868B (en) | Fault diagnosis method based on multi-objective optimization flywheel energy storage system bearing | |
CN113869208A (en) | Rolling bearing fault diagnosis method based on SA-ACWGAN-GP | |
CN110232415B (en) | Train bogie fault identification method based on biological information characteristics | |
CN113609569B (en) | Distinguishing type generalized zero sample learning fault diagnosis method | |
CN115905976B (en) | Attention mechanism-based fault diagnosis method, system and equipment for Highway Bi-LSTM bearing | |
CN114371009A (en) | High-speed train bearing fault diagnosis method based on improved random forest | |
CN114118138A (en) | Bearing composite fault diagnosis method based on multi-label field self-adaptive model | |
CN114386452B (en) | Nuclear power circulating water pump sun gear fault detection method | |
CN113569920B (en) | Second neighbor anomaly detection method based on automatic coding | |
CN112380932B (en) | Vibration signal characteristic value selection method and elevator health state evaluation or fault diagnosis method | |
CN111982514B (en) | Bearing fault diagnosis method based on semi-supervised deep belief network | |
CN111444969B (en) | Weakly supervised IVOCT image abnormal region detection method | |
CN112697435A (en) | Rolling bearing fault diagnosis method based on improved SELD-TCN network | |
Zhuang et al. | A neural-network approach to better diagnosis of defect pattern in wafer bin map |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |