CN111476339A - Rolling bearing fault feature extraction method, intelligent diagnosis method and system - Google Patents

Abstract

The invention provides a rolling bearing fault feature extraction method, an intelligent diagnosis method and a system, including: collecting bearing signal data of an electric drive end, preprocessing the signal data, dividing the signal data into a test set and a training set; According to the characteristics of the signal data, the first characteristic matrix of the training set and the test set are obtained respectively; the wavelet packet-AR spectrum estimation method is used to extract the characteristics of the signal data, and the second characteristic matrix of the training set and the test set are obtained respectively; the EMD-SVD method is used Extract the features of the signal data to obtain the third feature matrix of the training set and the test set respectively; splicing the first, second and third feature matrices, and compressing the data through PCA to obtain the training set and test set after dimension reduction, And input it to the classifier for data analysis and fault diagnosis, which can effectively ensure the completeness and sparsity of signal features, improve the generalization accuracy of feature extraction, and improve the generalization ability of the classifier.

Description

Rolling bearing fault feature extraction method, intelligent diagnosis method and system

technical field

The patent of the present invention belongs to the field of mechanical fault diagnosis, in particular to a rolling bearing fault feature extraction method, an intelligent diagnosis method and a system integrating various feature processing methods and fault identification methods.

Background technique

Mechanical fault diagnosis is a method to automatically supervise and maintain equipment based on measurable signal characteristics within the shortest expected period of failure. Rolling bearings, as the most important and most easily damaged components in rotating machinery, have always attracted the attention of the industry.

The existing bearing fault diagnosis technology usually includes three steps: collecting digital signal, processing digital signal, and classifying by classifier. Among them, the digital signal is mainly acquired by the acceleration sensor, and the classifier usually uses the machine supervised learning algorithm for fault identification. Although such algorithms are quite perfect today, the use of the supervised learning algorithm for fault identification largely depends on the Extracted signal features. The more complete and representative the extracted features are, the stronger the fault identification ability is. However, the traditional single feature extraction method often cannot obtain comprehensive and effective features, so it is difficult for the classifier to perform good generalization ability.

SUMMARY OF THE INVENTION

In view of the above deficiencies in the prior art, the present invention adopts a variety of methods to extract features from different angles, and performs dimensionality reduction processing through certain means, which ensures the completeness and sparseness of the extracted features. In the first stage, for the parameter optimization problem of SVM (Support Vector Machine), it is proposed to use PSO (Particle Swarm Optimization) to speed up the parameter optimization.

In a first aspect, the present invention provides a method for extracting fault features of a rolling bearing. The steps include: collecting bearing signal data at an electric drive end, preprocessing the signal data, and dividing the signal data into a test set and a training set;

The features of the signal data are extracted by the wavelet frequency band energy method, and the first feature matrices of the training set and the test set are obtained respectively;

The wavelet packet-AR spectral estimation method is used to extract the features of the signal data, and the second feature matrices of the training set and the test set are obtained respectively;

The EMD-SVD method is used to extract the features of the signal data, and the third feature matrices of the training set and the test set are obtained respectively;

The first, second and third feature matrices are spliced and processed, and data compression is performed by PCA to obtain the training set and test set after dimensionality reduction, that is, bearing fault features.

In a second aspect, the present invention also provides an intelligent diagnosis method for rolling bearing faults, including using the rolling bearing fault feature extraction method described in the first aspect to obtain a training set and a test set after dimensionality reduction, and the training set after dimensionality reduction is obtained and the test set input classifier for training to diagnose bearing faults;

The specific steps of inputting the dimensionality-reduced training set and test set into the classifier for training include: initializing the particle swarm, and initializing the particle velocity and position;

Calculate the fitness of the particle, and calculate the SVM recognition accuracy under the current penalty term coefficient and kernel function width;

Find the extreme value; update the particle velocity and position; judge whether the SVM classification error meets the termination condition; if the termination condition is met, put the test set into the classifier for classification, and analyze the classification results of the training set and test set; , then return to find the extreme value to continue processing.

In a third aspect, the present invention also provides a rolling bearing fault feature extraction system, including:

Acquisition module: configured to collect the bearing signal data of the electric drive end;

Preprocessing module: configured to preprocess the signal data, and divide the signal data into a test set and a training set;

Feature extraction module: configured to extract the features of the signal data by using the wavelet frequency band energy method to obtain the first feature matrix of the training set and the test set respectively; use the wavelet packet-AR spectrum estimation method to extract the features of the signal data to obtain the training set respectively and the second feature matrix of the test set; use the EMD-SVD method to extract the features of the signal data, and obtain the third feature matrix of the training set and the test set respectively; splicing the first, second and third feature matrices, and pass PCA Perform data compression to obtain the dimensionality-reduced training set and test set.

In a fourth aspect, the present invention further provides a computer-readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method for extracting fault features of a rolling bearing according to the first aspect is completed.

In a fifth aspect, the present invention also provides an intelligent diagnosis system for rolling bearing faults, including a classifier and the acquisition module, preprocessing module, and feature extraction module as described in the third aspect; the classifier is configured to: input The training set and test set after dimensionality reduction of the feature extraction module; initialize the particle swarm, initialize the speed and position of the particle's penalty term coefficient and kernel function width; calculate the SVM recognition accuracy under the current penalty term coefficient and kernel function width; Find the extreme value; update the particle velocity and position; judge whether the SVM classification error meets the termination condition; if the termination condition is met, put the test set into the classifier for classification, and analyze the classification results of the training set and test set; , then return to find the extreme value to continue processing.

Compared with the prior art, the present invention has the following beneficial effects:

1. The present invention uses a variety of methods to perform feature extraction on the signal, and uses PCA (singular value decomposition) for dimensionality reduction processing. Compared with a single method for feature extraction, it can effectively ensure the completeness and sparsity of signal features, and improve the performance of signal features. Generalization accuracy of feature extraction.

2. Aiming at the parameter optimization problem of SVM, the present invention utilizes the PSO algorithm to realize the optimization speed of two important parameters C and g of SVM. Compared with the conventional SVM classifier, the problem of overfitting and underfitting can be effectively avoided, Improve the generalization ability of the classifier.

3. The present invention can extract accurate and representative features from a large amount of data through preprocessing, use a variety of methods to extract features from different angles, and perform dimensionality reduction processing through certain means, which ensures that the extracted features are extracted. The completeness and sparseness of features, and in the classification and identification stage of the classifier, for the parameter optimization problem of SVM (Support Vector Machine), it is proposed to use PSO (Particle Swarm Optimization) to speed up the parameter optimization.

Description of drawings

The accompanying drawings that form a part of the present application are used to provide further understanding of the present application, and the schematic embodiments and descriptions of the present application are used to explain the present application and do not constitute improper limitations on the present application.

1 is a schematic flowchart of bearing vibration signal processing according to the present invention;

Fig. 2 is the wavelet decomposition tree of the present invention;

Fig. 3 is the wavelet packet decomposition tree of the present invention;

FIG. 4 is a flowchart of the SVM-PSO algorithm of the present invention.

Detailed ways:

The present invention will be further described below with reference to the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

In the present invention, terms such as "upper", "lower", "left", "right", "front", "rear", "vertical", "horizontal", "side", "bottom", etc. The orientation or positional relationship is based on the orientation or positional relationship shown in the accompanying drawings, and is only a relational word determined for the convenience of describing the structural relationship of each component or element of the present invention, and does not specifically refer to any component or element in the present invention, and should not be construed as a reference to the present invention. Invention limitations.

In the present invention, terms such as "fixed connection", "connected", "connected", etc. should be understood in a broad sense, indicating that it can be a fixed connection, an integral connection or a detachable connection; it can be directly connected, or it can be connected through the middle media are indirectly connected. For the relevant scientific research or technical personnel in the field, the specific meanings of the above terms in the present invention can be determined according to the specific situation, and should not be construed as a limitation of the present invention.

Example 1

In order to extract accurate and representative features from a large amount of data, the present invention adopts various methods to extract features from different angles, and performs dimensionality reduction processing by certain means, which ensures the completeness and integrity of the extracted features. Sparsity, and in the classification and identification stage of the classifier, for the parameter optimization problem of SVM (support vector machine), it is proposed to use PSO (particle swarm algorithm) to speed up the parameter optimization speed.

In order to achieve the above purpose, the solution of the present invention is as follows:

The method for extracting fault features of a rolling bearing includes the following steps: collecting bearing signal data of an electric drive end, preprocessing the signal data, and dividing the signal data into a test set and a training set;

The features of the signal data are extracted by the wavelet frequency band energy method, and the first feature matrices of the training set and the test set are obtained respectively;

The wavelet packet-AR spectral estimation method is used to extract the features of the signal data, and the second feature matrices of the training set and the test set are obtained respectively;

The EMD-SVD method is used to extract the features of the signal data, and the third feature matrices of the training set and the test set are obtained respectively;

The first, second and third feature matrices are spliced and processed, and data compression is performed by PCA to obtain the training set and test set after dimensionality reduction, that is, bearing fault features.

Further, the step of preprocessing the signal includes: the signal sequences to be used in the bearing data signal are F ₁ to F _i , and F _i is appropriately divided to form one signal sequence, that is, F _i ={f _i1 , f _i2 , f _i3 ...f _il }, where f _i1 is the first subset sequence divided by the i-th fault sequence; a part of the signal subsequence is randomly selected from F _i as the original training set; the other part is used as the original test set.

Further, the specific steps of using the wavelet band energy method to extract the features of the signal data include:

Perform wavelet decomposition on bearing signal data;

Decompose the different components of the bearing signal data into the corresponding frequency bands, and calculate the energy characteristics of each frequency band;

The energy eigenvectors are normalized to obtain the first eigenmatrix of the training set and the test set, respectively.

Further, the specific steps of using the wavelet packet-AR spectrum estimation method to extract the features of the signal data include: decomposing the bearing signal by wavelet packet;

Decompose different components of the signal into corresponding frequency bands, and perform wavelet packet reconstruction on each component to obtain a reconstructed signal;

Perform AR spectrum estimation on the reconstructed signal, and calculate the AR power spectral density of each frequency band;

Calculate the AR spectral energy features of each frequency band, and obtain the second feature matrix of the training set and the test set respectively.

Further, the specific steps of using the EMD-SVD method to extract the features of the signal data include:

Decompose the signal by EMD to obtain several IMF components;

Calculate the energy proportion of the IMF component;

Set a threshold to extract several IMF components whose energy ratio exceeds the threshold;

SVD is used to compress the extracted IMF components, and the compressed IMF components are used as signal features to obtain the third feature matrix of the training set and the test set respectively.

Further, the specific steps of splicing the first, second and third feature matrices include: obtaining training set α and test set β after splicing processing, where α=[α ₁ , α ₂ , α ₃ ] _{A1 ×M} , β=[β ₁ , β ₂ , β ₃ ] _A2×M , A1, A2 represent the number of signals in the training set and test set respectively, M represents the feature dimension α ₁ , α ₂ and α ₃ are the training set and α 3 respectively The first, second and third feature matrices of the set, β ₁ , β ₂ and β ₃ are the first, second and third feature matrices of the test set, respectively.

Further, the specific steps of performing data compression by PCA include: for the training data set α, calculating its covariance matrix; performing singular value decomposition on the covariance matrix; outputting the training set after dimension reduction; similarly, for the test set β Perform the same processing to obtain the dimensionality-reduced test set.

Example 2

The present invention is based on the bearing signal data set of the motor drive end of the Western Reserve University. It should be pointed out that such bearing fault data signals are divided into three categories: inner ring, outer ring, and rolling body, and each category includes diameters of 0.1778, 0.3556, and 0.5334mm. The artificial spark scars and the normal bearing data signals are divided into ten cases in total, so the signal sequence to be used is determined to be F ₁ -Fi , where _i =1, 2...10.

1 is a flowchart of feature extraction and intelligent diagnosis of bearing vibration signals of the present invention. As shown in FIG. 1 , the feature extraction and intelligent diagnosis method of bearing vibration signals of the present invention specifically includes the following steps:

(1) Data preprocessing

Divide F _i appropriately to form a signal sequence, that is, F _i ={f _i1 , f _i2 , f _i3 ...f _il }, where f _i1 is the first subset sequence divided by the i-th fault sequence . Next, a part of signal subsequences are _randomly selected from Fi as the original training set; the other part is used as the original test set. For convenience, the training set is a matrix of A1*B1, and the test set is a matrix of A2*B2, where A1 and A2 are the number of signals, and B1 and B2 are the signal feature dimensions.

(2) Multi-resolution analysis using wavelet transform to extract the energy characteristics of the signal in different frequency bands

The bearing signal is decomposed by wavelet, so that the different components of the signal are decomposed into corresponding frequency bands, and finally the energy characteristics of each frequency band are calculated. details as follows:

j∈Z(1)为小波基函数，对信号进行分解，具体数学表达如下：(2-1) Wavelet decomposition: select discrete wavelet function

j∈Z(1) is the wavelet basis function, which decomposes the signal. The specific mathematical expression is as follows:

where j is the decomposition scale, x is the position parameter of the decomposition, W _2j f(x) is the wavelet decomposition coefficient of the signal f at the decomposition scale 2j and the position is x.

Based on the mathematical expression of formula (2) wavelet decomposition, the signal is decomposed at M1 level, and the signal features contained in each leaf node are extracted. Assuming the original signal is f, denote the nth node of the mth layer of the wavelet decomposition tree (as shown in Figure 2) by (m, n), where m=0, 1, 2, 3...M1, n=0, 1 ; Each node represents a certain frequency band feature, (0, 0) represents the original signal, (1, 0) and (1, 1) represent the low-frequency coefficient S ₁₀ and high-frequency coefficient S ₁₁ of the wavelet decomposition, respectively, and so on (m, n) represents the n-th node coefficient S _mn of the m-th layer.

(2-2) Calculate the frequency band energy: Calculate the energy E _mn of each leaf node of the signal f, and the calculation formula is as follows:

In the formula, the value of n is constrained by m and 0<m≤M1; when m<M1, n=1; when m=M1, n=0 or n=1 (the actual meaning is that the M1 layer signal decomposition will get a low frequency leaf node and a high frequency leaf node). At this time k represents the feature dimension of S _mn

归一化，各频带的能量比重作为

的一个特征分量，

如下所示：(2-3) Normalization of eigenvectors: the eigenvectors

Normalized, the energy proportion of each frequency band is taken as

a feature component of ,

As follows:

In the formula, the value range of m and n is the same as that of formula (3).

(2-4) Obtain feature matrix: perform feature extraction for each signal according to the above steps, and finally obtain the first A1*M1 α ₁ training set and A2*M1 β ₁ test set.

(3) Using the wavelet packet-AR spectrum estimation method to extract the AR spectrum energy characteristics of the signal in each frequency band

The bearing signal is decomposed by wavelet packet, so that the different components of the signal are decomposed into corresponding frequency bands, and each component is reconstructed. Finally, the AR spectral energy characteristics of each frequency band are calculated by AR spectral estimation. details as follows:

(3-1) Wavelet packet decomposition: Select the appropriate wavelet basis function to determine the number of decomposition layers for the signal. The specific mathematical expression is as follows:

为信号f(x)经过j层小波包分解得到的第m个分解序列；

为原数字信号f(x)。where k is the translation time factor; h is the low-pass filter coefficient; g is the high-pass filter coefficient;

is the mth decomposition sequence obtained by decomposing the signal f(x) through the j-layer wavelet packet;

is the original digital signal f(x).

Based on the mathematical expression of wavelet packet decomposition in formula (5), the signal is decomposed in M2 layer, and the signal features contained in the last layer of leaf nodes are extracted. Assuming the original signal is f, use (m, n) to represent the nth node of the mth layer of the wavelet packet decomposition tree (as shown in Figure 3), where m=0, 1, 2, 3...M2, n=0, 1,...2 ^M2 -1; each node represents a certain frequency band feature, (0, 0) represents the original signal, (1, 0) and (1, 1) represent the low-frequency coefficients S10 and The high-frequency coefficient S11, and so on (m, n) represents the coefficient S _mn representing the n-th node of the m-th layer.

(3-2) Wavelet packet reconstruction: perform signal reconstruction according to the coefficient sequence obtained at the M2th layer. RS _mn represents the reconstructed signal of coefficient S _mn . The mathematical expression of wavelet reconstruction is as follows:

和

分别为h和g的对偶算子，

为j-1层重构形成的第m个重构序列。where:

and

The dual operators of h and g, respectively,

The mth reconstruction sequence formed for the j-1 layer reconstruction.

(3-3) AR spectrum analysis: perform AR spectrum estimation on the RS _mn reconstructed signal to obtain the AR spectrum power density P _mn of each frequency band, as follows:

Assuming that the reconstructed signal sequence is x(n), the autoregressive model of the sequence can be expressed as:

的正态分布白噪声；q为模型的阶次。通过相关的信号处理知识，求得AR模型参数aj(j＝1，2…N)和

然后由自传递函数，计算出信号x(n)的功率谱密度为：where w(x) has zero mean and variance is

The normally distributed white noise; q is the order of the model. Through the relevant signal processing knowledge, the AR model parameters aj (j=1, 2...N) and

Then from the self-transfer function, the power spectral density of the signal x(n) is calculated as:

(3-4) Solving the AR spectral energy of each frequency band signal: Set the energy Ep _{mn corresponding to P mn ,} then:

In the formula, k is the characteristic dimension; m=M2; n=0, 1, 2...2 ^M2 -1.

各频带的能量作为

的一个特征分量，

如下所示：(3-5) Obtain eigenvectors

The energy of each frequency band is as

a feature component of ,

As follows:

(3-6) Obtain feature matrix: perform feature extraction for each signal according to the above steps, and finally obtain α ₂ training set of size A1*(2 ^M2 -1) and β ₂ test of A2*(2 ^M2 -1) set.

(4) Signal feature extraction method based on EMD-SVD

The signal is decomposed by EMD to obtain several IMF components; the energy proportion of the IMF components is calculated; several IMF components with a large energy proportion are extracted; SVD is used to compress the extracted IMF components, and the compressed IMF components are used as signal features.

details as follows:

(4-1) Use EMD decomposition to extract the IMF components that are ranked in the front and whose energy sum exceeds 98%, as follows:

Find all local extreme points of the original signal f(t)

Using the interpolation algorithm, fit all the local extreme points found to obtain the upper envelope f _max (t) and the lower envelope f _min (t)

Obtain the mean value a ₁ from the upper and lower envelopes

Further through h ₁ =f(t)-a ₁ , it is judged whether h ₁ satisfies the two conditions of IMF. If so, h ₁ is the first-order IMF component; if not, h ₁ is used as the new f(t ), repeat steps a) to d) to obtain h ₁₁ =h ₁ -a ₁₁ , where a ₁₁ is the mean value of the upper and lower envelopes of h1; if h ₁₁ is still not satisfied, continue to repeat the above steps until h _1k satisfies the condition, Denoted as C ₁ =h _1k is the first-order IMF component

After the first-order IMF component is separated, r ₁ =f(t)-C ₁ is obtained, and r ₁ is taken as the new f(t), and then steps a) to d) are repeated, and the n-order IMF components and The residual r _n , and finally realizes the EMD decomposition, the formula is as follows:

Calculate the energy Eci of different IMF components Ci, where i=1, 2, 3..n, and then find the IMF component S _m which accounts for a large proportion of energy and whose energy sum accounts for more than 98%.

(4-2) Obtain the IMF characteristic matrix U _m×τ : each IMF component is a characteristic component of U m× _τ , and U _m×τ is as follows:

In the U _m×τ matrix, m represents the number of vectors, and τ represents the feature dimension.

(4-3) The eigenvalue decomposition is further carried out by SVD to realize the compression of the U _m×τ matrix. The decomposition formula is as follows:

右奇异矩阵；B_m×τ为奇异值矩阵，且其为对角矩阵，具体如下所示：In the above formula, Q _m×m is a left singular matrix,

Right singular matrix; B _m×τ is a singular value matrix, and it is a diagonal matrix, as follows:

且维数为M3，因此可得：

Here, the first few singular values with relatively large values will be taken as the characteristics of the signal according to the situation. Let the extracted feature vector be

And the dimension is M3, so we can get:

(4-4) Obtaining a feature matrix: perform feature extraction for each signal according to the above steps, and finally obtain an α ₃ training set of A1*M3 and a β ₃ test set of A2*M3.

(5) Use PCA to reduce the dimensionality of the obtained data set

(5-1) Obtain training set α and test set β, where α=[α ₁ , α ₂ , α ₃ ] _A1×M , β=[β ₁ , β ₂ , β ₃ ] _A2×M , A1, A2 Respectively represent the number of signals in the training set and test set, M represents the feature dimension, and M=M1+2 ^M2 -1+M3 The further expansion of α and β is expressed as follows:

where Tri and Tei represent the vectors of A1*1 and A2*1, respectively.

(5-2) Compress α and β using PCA

details as follows:

For the training data set α, calculate its covariance matrix Q, the formula is as follows:

Perform singular value decomposition on the covariance matrix: [U, S, V]=SVD(Q), where U represents the left singular matrix, V represents the right singular matrix, and S represents the singular value matrix.

其中ev_i代表Q的第i个奇异值。Find the feature ratio and the corresponding k value when the threshold is met. The judgment conditions are as follows:

where ev _i represents the ith singular value of Q.

Select the first k columns of U as the projection matrix Ptr=[u1, u1,...uk], where the kth column is the eigenvector corresponding to the kth singular value.

Output the training data set after dimensionality reduction Trian _A1×k =Ptr×α, k＜M

The same processing is performed on the test set β to obtain the final test set Test _A2×r =Pte×β, r<M

(6) Using the SVM method to classify the fault signal

(6-1) Classifier training

As shown in the flowchart of SVM-PSO algorithm in Figure 4, the specific explanation is as follows:

(a) Initialize the particle swarm, the size of the particle swarm is N; the number of iterations It; the position of the penalty term coefficient ci is Pci, the velocity is Vci; the position of the kernel function width gi is Pgi, the velocity is Vgi; define the SVM classification error error

(b) Calculate the fitness Fit[i] of the particle, that is, calculate the accuracy of the SVM classification under the current ci and gi;

(c) Compare Fit[i] with the individual optimal Pbest[i], if Fit[i]>Pbest[i], then Pbest[i]=Fit[i]; compare Fit[i] with the individual optimal Gbest[i] is compared, if Fit[i]>Gbest[i], then Gbest[i]=Fit[i]

(d) Update (Pci, Vci) and (Pgi, Vgi), which is equivalent to updating ci and gi; then calculate the fitness Fit[i] of the particle, and the update formula is shown in (13) (14):

v _im =w· _vim (k)+c ₁ r ₁ ·(P _im (k)-x _id (k))+c ₂ r ₂ ·(P _gm (k)-x _im (k)) (13 )

x _im (k+1)=x _im (k)+v _im (k+1) (14)

In the above formula, w is called inertia weight, c1 and c2 represent acceleration constants, r1 and r2 are random numbers between (0, 1), and k is the number of iterations.

(e) Judging that the current error is small enough or the current iteration iter is greater than It, then exit, output the c and g parameters, and the target parameter values of w and b in the objective function (15); otherwise, return to the second step.

Subject to: y _i (W ^T x _i +b)≥1-ξ _i , ξ _i ≥0

(6-2) Put the test set into the classifier for classification

(6-3) Analyze the classification results of the training set Train and the test set Test

The above is the specific flow of the present invention.

Other specific embodiments also provide:

An intelligent diagnosis method for rolling bearing faults, comprising using the rolling bearing fault feature extraction method as described in Embodiment 1 to obtain a training set and a test set after dimension reduction, and inputting the training set and test set after dimension reduction into a classifier for training, Diagnose bearing faults.

Further, the specific steps of inputting the dimensionality-reduced training set and test set into the classifier for training include: initializing the particle swarm, initializing the speed and position of the particle's penalty term coefficient ci and the kernel function width gi; Under the term coefficient ci and the kernel function width gi, calculate the particle fitness (SVM recognition accuracy); find the extreme value; update the particle speed and position; judge whether the SVM classification error meets the termination condition; if the termination condition is met, put the test set into Enter the classifier for classification, and analyze the classification results of the training set and the test set; if not satisfied, return to finding the extreme value to continue processing.

A rolling bearing fault feature extraction system, comprising:

Acquisition module: configured to collect the bearing signal data of the electric drive end;

Preprocessing module: configured to preprocess the signal data, and divide the signal data into a test set and a training set;

Feature extraction module: configured to extract the features of the signal data by using the wavelet frequency band energy method to obtain the first feature matrix of the training set and the test set respectively; use the wavelet packet-AR spectrum estimation method to extract the features of the signal data and obtain the training set respectively and the second feature matrix of the test set; use the EMD-SVD method to extract the features of the signal data, and obtain the third feature matrix of the training set and the test set respectively; splicing the first, second and third feature matrices, and pass PCA Perform data compression to obtain the dimensionality-reduced training set and test set.

A computer-readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method for extracting fault features of a rolling bearing as described in Embodiment 1 is completed.

An intelligent diagnosis system for rolling bearing faults, comprising a classifier and the acquisition module, preprocessing module, and feature extraction module described in the above embodiments; the classifier is configured to: input a training set after dimension reduction by the feature extraction module and test set; initialize the particle swarm, initialize the speed and position of particles c and g; under the current penalty term coefficient c and kernel function width g, calculate particle fitness (SVM recognition accuracy); find extreme values; update particle speed and position; judge whether the SVM classification error satisfies the termination condition; if it satisfies the termination condition, put the test set into the classifier for classification, and analyze the classification results of the training set and test set; if not, return to find the extreme value Continue processing.

Although the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, they do not limit the scope of protection of the present invention. Those skilled in the art should understand that on the basis of the technical solutions of the present invention, those skilled in the art do not need to pay creative efforts. Various modifications or deformations that can be made are still within the protection scope of the present invention.

Claims (10)

1. A rolling bearing fault feature extraction method is characterized by comprising the following steps: acquiring bearing signal data of the electric driving end, preprocessing the signal data, and dividing the signal data into a test set and a training set;

extracting the characteristics of signal data by adopting a wavelet band energy method to respectively obtain a first characteristic matrix of a training set and a first characteristic matrix of a testing set;

extracting the characteristics of the signal data by adopting a wavelet packet-AR spectrum estimation method to respectively obtain second characteristic matrixes of the training set and the test set;

extracting the characteristics of the signal data by adopting an EMD-SVD method to respectively obtain a third characteristic matrix of the training set and the test set;

and splicing the first characteristic matrix, the second characteristic matrix and the third characteristic matrix, and performing data compression through PCA to obtain a training set and a test set after dimensionality reduction, namely the bearing fault characteristics.

2. The rolling bearing fault signature extraction method of claim 1, wherein the signal preprocessing step comprises: the signal sequence to be used in the bearing data signal is F₁～F_iWill F_iAre appropriately divided to form I signal sequences, i.e. F_i＝{f_i1，f_i2，f_i3…f_ilIn which f_i1Dividing a first subset sequence into ith fault sequences; random slave F_iSelecting a part of signal subsequence as an original training set; the other part is used as the original test set.

3. The rolling bearing fault feature extraction method according to claim 1, wherein the specific step of extracting the features of the signal data by using the wavelet band energy method comprises: carrying out wavelet decomposition on the bearing signal data;

decomposing different components of the bearing signal data into corresponding frequency bands, and calculating the energy characteristics of each frequency band;

and normalizing the energy feature vector to respectively obtain a first feature matrix of the training set and a first feature matrix of the test set.

4. The rolling bearing fault feature extraction method according to claim 1, wherein the specific step of extracting the features of the signal data by using a wavelet packet-AR spectrum estimation method comprises: carrying out wavelet packet decomposition on the bearing signal;

decomposing different components of the signal into corresponding frequency bands, and performing wavelet packet reconstruction on each component to obtain a reconstructed signal;

performing AR spectrum estimation on the reconstructed signal, and calculating AR spectrum energy characteristics of each frequency band;

and obtaining a feature vector of the AR spectrum energy feature, and respectively obtaining a second feature matrix of the training set and the test set.

5. The rolling bearing fault feature extraction method according to claim 1, wherein the specific step of extracting the features of the signal data by using the EMD-SVD method comprises:

EMD decomposition is carried out on the signals to obtain a plurality of IMF components;

calculating the energy ratio of the IMF components;

setting a threshold value, and extracting a plurality of IMF components with energy ratio exceeding the threshold value;

and compressing the extracted IMF components by adopting SVD, and taking the compressed IMF components as signal characteristics to respectively obtain a third characteristic matrix of the training set and the test set.

6. The rolling bearing fault feature extraction method according to claim 1, wherein the concrete step of splicing the first, second and third feature matrixes comprises obtaining a training set α and a test set β after splicing, wherein α ═ α₁，α₂，α₃]_A1×M，β＝[β₁，β₂，β₃]_A2×MTables A1 and A2Representing the number of signals in the training set and test set, M representing the feature dimension α₁、α₂And α₃First, second and third feature matrices of the training set, β, respectively₁、β₂And β₃First, second and third feature matrices of the test set, respectively;

the specific steps of data compression through PCA include that a covariance matrix of a training data set α is calculated, singular value decomposition is conducted on the covariance matrix, a training set after dimension reduction is output, and the same processing is conducted on a test set β in the same mode to obtain a test set after dimension reduction.

7. An intelligent diagnosis method for rolling bearing faults is characterized by comprising the steps of obtaining a training set and a test set after dimensionality reduction by adopting the rolling bearing fault feature extraction method according to claims 1-6, inputting the training set and the test set after dimensionality reduction into a classifier for training, and diagnosing the bearing faults;

the specific steps of inputting the training set and the test set after the dimensionality reduction into the classifier for training comprise: initializing a particle swarm, and initializing a penalty term coefficient of the particle and the speed and position of the kernel function width; calculating the SVM identification accuracy under the current punishment item coefficient and the kernel function width;

searching an extreme value; updating the particle speed and position; judging whether the SVM classification error meets a termination condition or not;

if the termination condition is met, putting the test set into a classifier for classification, and analyzing the classification results of the training set and the test set; if not, returning to the position where the extreme value is found for continuous processing.

8. A rolling bearing fault feature extraction system comprising:

an acquisition module configured to: collecting bearing signal data of the electric driving end;

a pre-processing module configured to: preprocessing signal data, and dividing the signal data into a test set and a training set;

a feature extraction module configured to: extracting the characteristics of signal data by adopting a wavelet band energy method to respectively obtain a first characteristic matrix of a training set and a first characteristic matrix of a testing set; extracting the characteristics of the signal data by adopting a wavelet packet-AR spectrum estimation method to respectively obtain second characteristic matrixes of the training set and the test set; extracting the characteristics of the signal data by adopting an EMD-SVD method to respectively obtain a third characteristic matrix of the training set and the test set; and splicing the first characteristic matrix, the second characteristic matrix and the third characteristic matrix, and performing data compression through PCA to obtain a training set and a test set after dimensionality reduction.

9. A computer readable storage medium storing computer instructions which, when executed by a processor, perform the rolling bearing fault feature extraction method of claims 1-6.

10. An intelligent diagnosis system for rolling bearing faults is characterized by comprising: an acquisition module, a preprocessing module, a feature extraction module in a classifier and the feature extraction system of claim 8; the classifier configured to: inputting a training set and a test set after dimension reduction of a feature extraction module; initializing a particle swarm, and initializing a penalty term coefficient of the particle and the speed and position of the kernel function width; calculating the SVM identification accuracy under the current punishment item coefficient and the kernel function width; searching an extreme value; updating the particle speed and position; judging whether the SVM classification error meets a termination condition or not; if the termination condition is met, putting the test set into a classifier for classification, and analyzing the classification results of the training set and the test set; if not, returning to the position where the extreme value is found for continuous processing.

Images