CN112163472A - Rolling bearing diagnosis method based on multi-view feature fusion - Google Patents

Abstract

The invention researches the characteristic fusion process of the multi-view characteristic set of the vibration signal of the rolling bearing and provides the characteristic fusion based on random forest characteristic selection and self-encoder dimension reduction. Firstly, extracting multi-view characteristics of a rolling bearing by utilizing a statistical characteristic and a time sequence signal spectrum analysis method; 2) outputting the feature importance of the high-dimensional features by using a random forest, and eliminating invalid features based on the feature importance to reduce feature dimensions; 3) aiming at redundant features with the same feature importance in the feature set, the self-encoder is utilized to perform nonlinear dimension reduction on the redundant features, so that a small redundant low-dimensional feature set capable of clearly expressing the bearing state difference is obtained.

Description

Rolling bearing diagnosis method based on multi-view feature fusion

Technical Field

The invention belongs to the field of bearing vibration signal processing, and relates to a rolling bearing diagnosis method based on multi-view feature fusion.

Background

Rolling bearings are one of the key components in the drive train of a wind turbine. Due to the harsh operating environment of wind turbines, rolling bearing failures often occur. Statistically, 30% of rotating machinery failures are caused by rolling bearings, while 80% of wind turbine gearbox failures are caused by bearing failures. Therefore, bearing fault diagnosis is crucial for efficient and reliable operation of wind turbines. Traditionally, fault diagnosis of wind turbine rolling bearings is based on spectral analysis of vibration signals. The key technology is to extract fault characteristic frequency from noise signal. The spectral analysis methods are fourier transform, hilbert transform and some joint time-frequency analysis methods such as Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD). The traditional method only researches a bearing vibration signal from a specific angle, and manually extracts features according to prior knowledge of a signal processing technology and diagnosis professional knowledge, so that the requirements of real-time performance and portability of fault diagnosis cannot be met. In the big data age.

In order to comprehensively analyze the difference between the failures, it is necessary to check the vibration signal of the bearing from a plurality of angles to grasp the overall state of the bearing. In the invention, the vibration signal is subjected to spectrum analysis and time-frequency analysis. Features are then extracted from the time domain, frequency domain (spectral and envelope spectra) and time-frequency domain (EMD). Although the multi-view functions are highly complementary, these features are often redundant, which is detrimental to fault diagnosis. Therefore, feature selection and feature fusion prior to classification are necessary.

Disclosure of Invention

The invention aims to improve the accuracy and robustness of fault diagnosis of a rolling bearing under complex conditions, and provides a rolling bearing vibration characteristic construction method.

The technical scheme is as follows: in order to solve the technical problem, the invention provides a rolling bearing diagnosis method based on multi-view feature fusion, and a multi-view function set can fully utilize information of original signals to reflect the difference between states. The extraction of multi-view features and the reduction of redundancy by feature fusion are two basic parts of the model. The method comprises the following steps:

(1) and (3) combining a signal processing method and statistical characteristics, extracting vibration characteristics of the rolling bearing from the angles of a time domain, a frequency domain and a time-frequency domain, and constructing a multi-source high-dimensional characteristic set.

(2) Inputting the bearing feature set constructed in the step (1) into a random forest model for model training, outputting the feature importance of each feature by using the trained random forest model, then selecting more important features (setting the number of selected features), and rejecting other invalid features.

(3) Inputting the more important features reserved in the step (2) into a self-coding network, performing nonlinear dimension reduction on the screened features by using a self-coding model, and outputting a feature set after dimension reduction as the finally constructed bearing features.

Preferably, in step (1): and (4) extracting features, and inspecting the multi-view analysis result of the original vibration signal, namely a time domain signal, a frequency domain signal and a time frequency signal. 1) The original vibration signal contains information including a normal vibration component, a faulty vibration component and ambient noise. To reduce the dependence on a priori knowledge, a number of statistical features are extracted from the time-series waveform. Part of the feature calculation formula is as follows: mean, maximum amplitude, maximum peak, standard deviation, trivial root amplitude, skewness, kurtosis, and the like. 2) The extraction of the frequency domain features must first transfer the data points in the time domain to the frequency domain by means of fourier and hilbert transforms before the signal can be analyzed in the frequency domain. The frequency domain features include frequency-structured features including average frequency, center frequency, root mean square of frequency, frequency deviation. 3) EMD adaptively decomposes a signal into a finite number of signals defined as eigenmode functions (IMFs) from high frequencies to low frequencies of unequal bandwidth. The internal fluctuations of the signal are reflected in the extracted IMFs, which contain the actual physical information of the signal. In addition to all the basic features in all the basic time domains, the IMF will also be considered to extract entropy features and fractal dimensions.

Preferably, in step (2): in this step, a random forest is applied to the training data, and the results of all decision trees in the forest are then aggregated based on the highest voting weights of all data pairs for classification. The feature importance evaluation is to calculate the contribution of each feature to each tree in the random forest, then take the average value, and finally compare the contributions of different features.

Preferably, in step (3): after the random forest model feature selection, partial redundant features exist in the multi-view feature set, namely, a plurality of features have similar or same relevance degrees with the label, the features only represent the same aspect of the fault difference, and the redundant features account for too much so that other features cannot be regarded as important in the model. In the step, the redundancy in the feature set is further reduced by adopting the self-encoder, and because the self-encoder is a neural network model, the convergence speed and the training effect can be effectively improved by standardizing before data input. All feature vectors are normalized by removing the mean and scaling to unity variance.

Has the advantages that: compared with the prior art, the invention has the advantages and effects that:

1. in the feature extraction, a signal processing method and a statistical formula are combined to construct a high-dimensional mixed feature set of equipment, so that the dependence on professional knowledge in the fault diagnosis feature extraction process is reduced.

2. And (4) sequencing the features according to the importance by adopting a random forest model pair, and replacing the standard considered to be constructed in the traditional method by using an automatic screening mode.

3. The method creatively combines feature selection and feature fusion, and eliminates invalid features and redundant features of the equipment multi-view feature set in two steps.

Drawings

FIG. 1 is a flow chart of a multi-view based feature fusion method;

FIG. 2-bench view of the bearing apparatus;

FIG. 3-time domain plot of different fault vibration signals;

FIG. 4-non-inner ring fault sample spectrogram;

FIG. 5-inner circle fault sample envelope spectrum;

FIG. 6-5 IMF components before inner ring failure;

FIG. 7-importance of random forest features;

FIG. 8-autocoder hyper-parametric trellis optimization diagram.

Detailed Description

Fig. 1 is a flow chart of a multi-view feature extraction and fusion method. The technical solution of the present invention is further explained below with reference to the drawings and the embodiments.

The method comprises the following steps: the data set was published using rolling bearings of the University of kiss west reservoir (Case Western Reserve University). As shown in fig. 2(a), the test stand consists of a 2 horsepower motor, torque sensor/encoder, power meter and control electronics (not shown) arranged on the left, middle and right sides of the test stand, respectively. Fig. 3 is a waveform diagram of vibration signals of a rolling bearing under four 0.007 feet crack conditions. The feature extraction of the vibration signal comes from three aspects: time domain, frequency domain and time-frequency domain. Extracting the characteristics of the vibration signals from a plurality of visual angles, introducing the time-frequency domain analysis result of the signals on the basis of a spectrogram and an envelope spectrogram, decomposing the original signals by using an empirical mode decomposition mode, and generating intrinsic mode components IMFs from high frequency to low frequency. And then, extracting the features of the sequence by adopting a mode of solving the sequence statistical features. Because the quantity of IMFs decomposed among different signals is different, in order to ensure that the characteristic quantity of the sample is consistent and main characteristic information is reserved, the IMFs obtained after the first 5 EMD decompositions are selected for characteristic extraction. The spectra, envelope spectra and IMFs plots of the EMD decomposition of the inner ring fault samples are shown in fig. 4, 5 and 6.

The time domain extracts 43-dimensional statistical features. The frequency domain features are based on spectral and envelope spectra, including the underlying 43 statistical features and also the 4 frequency-dependent features. The time domain features include statistical features, three entropy features, and box-dimensional features. As shown in table 1 in detail,

TABLE 1 high dimensional feature set for rolling bearings

Step two: fig. 7 shows the importance of each feature in four different feature sets, including the time domain feature (T), the frequency domain feature consisting of a spectrum (FFT) and an envelope spectrum (HT), the time-frequency domain (EMD) from the IMF. To facilitate viewing, the first picture had features removed that were less than 0.0001 importance. As can be seen from fig. 7(a), the features extracted from the original signal waveform and spectrum have relatively high features, and the features extracted from EMD are inferior. The time domain characteristics have better performance in different types of fault identification, namely the difference of vibration oscillograms of the outer ring fault and the inner ring fault of the bearing is obvious, so that the importance of the time domain characteristics is improved. But in the classification task of the same type and different crack sizes and fault orientations, the time-frequency characteristics and the envelope spectrum characteristics have obvious advantages. The distribution of all feature importance is given in fig. 7 (b). As can be seen from the figure, the importance of most features is low, and the importance of only 12 features is significantly higher than that of other features. In order to more clearly reveal the differences between states, less important features need to be culled. As is apparent from fig. 7(b), only a few features are very important. If only the head features with higher importance are needed to be utilized for fault diagnosis, the time cost and the computing resources for training and testing the recognition model are greatly reduced. The determination of the specific retained feature quantity needs to refer to the accuracy of SVM classification.

Step three: as an unsupervised machine learning model, there are two important parameters to optimize for an autoencoder: the number of nodes of the hidden layer and the iteration number (epoch) are optimized through grid search. As shown in fig. 8, there are three combinations that can make the SVM achieve good classification results, labeled as points A, B and C. The number of hidden layer nodes is 10, 20 and 30, and the Epoch is 450, 250 and 450, respectively. To maximize classification accuracy, point C is selected. Therefore, the number of hidden layer nodes and the number of epochs are determined as (30, 450).

Step four: the accuracy of classification of SVM classification test sets for 4 different feature sets (temporal, total, RF selection, and RF and self-encoding fusion) was 87.02%, 89.10%, 96.71%, and 99.10%, respectively. The analysis model classification error is generated, and the error mainly comes from the classification error of the sample under the 4 outer ring fault states. As can be seen from table 2, the method of feature fusion of random forest and self-encoder used in the present invention achieves the best classification result. The features of PCA fusion are the worst in SVM classification, which is only 87.59%, and even the classification accuracy (89.10%) of the full feature set before fusion can not be obtained. In contrast, KPCA better reflects the differences between states, with an accuracy that is improved by 2.3% compared to the original features. The accuracy of LLE is greatly improved, mainly due to its good non-linear mapping capability. Among the models, the RF + AutoE model proposed by the invention has the highest accuracy, which further illustrates the robustness of the invention in extracting effective information and reducing feature redundancy. While the invention has been shown and described with respect to the preferred embodiments, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the scope of the invention as defined in the following claims.

TABLE 2 comparison chart of SVM classification accuracy by multiple feature fusion method

Claims (4)

1. A rolling bearing diagnosis method based on multi-view feature fusion is characterized by comprising the following steps:

(1) and combining a signal processing method and statistical characteristics, extracting the vibration characteristics of the rolling bearing from the angles of time domain, frequency domain and time-frequency domain, and constructing a multi-source high-dimensional characteristic set.

(2) Inputting the bearing feature set constructed in the step (1) into a random forest model for model training, outputting feature importance of each feature by using the trained random forest model, presetting the importance, selecting the important feature according to the importance, and eliminating other invalid features;

(3) inputting the important features reserved in the step (2) into a self-coding network, carrying out nonlinear dimension reduction on the screened features by using a self-coding model, and outputting a feature set after dimension reduction as finally constructed bearing features;

(4) collecting the finally constructed bearing characteristics and the corresponding fault types as a training set, and inputting the training set into a neural network model for training to obtain a rolling bearing fault diagnosis model;

(5) and (4) acquiring a new bearing vibration signal, processing the bearing vibration signal according to the methods in the steps (1) to (3), obtaining bearing characteristics, and inputting the bearing characteristics into the rolling bearing fault diagnosis model in the step (4) to output a fault type.

2. The method for diagnosing a rolling bearing based on multi-view feature fusion according to claim 1, wherein the method in the step (1) is as follows:

1) the method comprises the steps that information contained in an original vibration signal comprises a normal vibration component, a fault vibration component and environmental noise, a plurality of statistical characteristics are extracted from a time sequence waveform, and the average value, the maximum amplitude, the maximum peak value, the standard deviation, the trivial root amplitude, the skewness and the kurtosis are calculated for the extracted characteristics;

2) extracting frequency domain characteristics, namely firstly transmitting data points in a time domain to a frequency domain through Fourier transformation and Hilbert transformation, and then inspecting a spectrum analysis signal, wherein the frequency domain characteristics comprise characteristics of frequency construction, including average frequency, central frequency, frequency root mean square and frequency deviation;

3) EMD adaptively decomposes signals into a limited number of signals which are defined as Inherent Mode Functions (IMF) from high frequency to low frequency with unequal bandwidths, and time-frequency information of vibration signals can be obtained by considering IMFs obtained by EMD decomposition, so that rolling bearing characteristics comprise time-domain characteristics, frequency-spectrum frequency-domain characteristics and EMD time-frequency characteristics.

3. The rolling bearing diagnosis method based on multi-view feature fusion according to claim 1, wherein in the step (2): in the step, a random forest is applied to training data, then the random forest model can give the importance of all features in the feature set according to the contribution of each feature to the decision tree branches in the training process, and the features are chosen based on the feature importance.

4. The rolling bearing diagnosis method based on multi-view feature fusion of claim 1, wherein in the step (3), the training process of the self-coding is an unsupervised training, and the trained self-coding performs nonlinear dimension reduction on the features according to the distribution characteristics of the feature space to reduce the occupation ratio of redundant features.

Images