CN107832729A - A kind of bearing rust intelligent diagnosing method - Google Patents
A kind of bearing rust intelligent diagnosing method Download PDFInfo
- Publication number
- CN107832729A CN107832729A CN201711176681.7A CN201711176681A CN107832729A CN 107832729 A CN107832729 A CN 107832729A CN 201711176681 A CN201711176681 A CN 201711176681A CN 107832729 A CN107832729 A CN 107832729A
- Authority
- CN
- China
- Prior art keywords
- frequency
- bearing
- spectrum
- signal
- extracted
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- 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
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Signal Processing (AREA)
- Acoustics & Sound (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biomedical Technology (AREA)
- Evolutionary Biology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Testing Resistance To Weather, Investigating Materials By Mechanical Methods (AREA)
Abstract
The present invention discloses a kind of bearing rust intelligent diagnosing method, by gathering the vibration signal of vibration signal and bearing of the bearing in the case where there are all kinds of different faults states in normal state in advance, and Training RBF Neural Network is removed after carrying out feature extraction to vibration signal, and build corrosion intelligent classification model；The real-time fault diagnosis to bearing can have both been realized using the corrosion intelligent classification model afterwards.
Description
Technical field
The present invention relates to bearing failure diagnosis technical field, and in particular to a kind of bearing rust intelligent diagnosing method.
Background technology
Bearing is a kind of standard component being widely used in all kinds of machineries.Most bearings are all by metal material system
Into it is common problem that metal material produces corrosion in atmospheric environment.Bearing rust can influence its accuracy class, drop
Low service life, or even scrap, or even trigger security incident.Therefore, a simple and effective bearing rust is studied and establishes to examine
Disconnected and disaggregated model has great importance in engineering.
The content of the invention
The present invention provides a kind of bearing rust intelligent diagnosing method, and it can realize that the implementation to bearing rust is monitored with examining
It is disconnected.
To solve the above problems, the present invention is achieved by the following technical solutions：
A kind of bearing rust intelligent diagnosing method, it is as follows to specifically include step：
The vibration signal and bearing of step 1, in advance collection bearing in the case where there are all kinds of different faults states are in normal condition
Under vibration signal, and these vibration signals are divided into two classes, one kind is used as training sample, another kind of to be used as test sample；
Step 2, extract training characteristics from the training sample of step 1；I.e.
Step 2.1, elder generation extract time domain charactreristic parameter from the time domain of vibration signal；FFT is carried out to vibration signal again
Spectrum information is obtained, frequency domain character parameter is extracted from frequency spectrum；Hilbert envelope spectral transformations are carried out to vibration signal afterwards to be wrapped
Network spectrum information, frequency domain character parameter is extracted from envelope spectrum；
Step 2.2, first vibration signal is handled using redundant wavelet packet transform, obtain the decomposed signal of frequency band, and
Time domain charactreristic parameter is extracted respectively from the decomposed signal of each frequency band；Frequency is obtained to the decomposed signal progress FFT of frequency band again
Spectrum information, extract frequency domain character parameter respectively from each frequency spectrum；Hilbert envelope spectral transformations are carried out to the decomposed signal of frequency band afterwards
Envelope spectrum information is obtained, extracts frequency domain character parameter respectively from each envelope spectrum；
Step 2.3, first vibration signal is handled using redundancy multi-wavelet packets transform, obtains the decomposed signal of frequency band,
And time domain charactreristic parameter is extracted respectively from the decomposed signal of each frequency band；The decomposed signal progress FFT of frequency band is obtained again
Spectrum information, extract frequency domain character parameter respectively from each frequency spectrum；Hilbert envelope spectrum changes are carried out to the decomposed signal of frequency band afterwards
Get envelope spectrum information in return, extract frequency domain character parameter respectively from each envelope spectrum；
Step 2.4, the step 2.1-2.3 time domain charactreristic parameters extracted and frequency domain character parameter are formed and train spy
Sign；
Step 3, the susceptibility that all training characteristics obtained by step 2 are carried out is assessed, obtain commenting for each training characteristics
Estimate the factor, and training characteristics are ranked up according to the order of evaluation factor from big to small；
Step 4, that the training sample of training characteristics sort obtained by step 3 and step 1 gained is input into RBF is neural
Network is trained, and the RBF neural being trained is tested using the test sample obtained by step 1, is exported
Relation between classification accuracy and these Characteristic Numbers；
Step 5, the feature that can reach default classification accuracy is selected as the spy most sensitive to corrosion after completing classification
Levy subset；The input feature vector of the most sensitive character subset as RBF neural is selected, establishes corrosion intelligent classification model；
The real-time vibration signal composition of step 6, collection bearing in the course of the work treats diagnostic sample；And diagnostic sample will be treated
It is input in corrosion intelligent classification model, to complete to diagnose.
In above-mentioned steps 2, the time domain charactreristic parameter extracted includes And p5=max | x (n) |；Wherein x (n) is time-domain signal, n=1,2 ... N；N
It is sample points.
In above-mentioned steps 2, the frequency domain character parameter extracted includes： WithWherein s (k) is time-domain signal x (n) frequency spectrum, k=1,2 ...,
K；K is spectral line number；fkIt is the frequency scale of kth bar spectral line.
Compared with prior art, the present invention is believed by gathering vibration of the bearing in the case where there are all kinds of different faults states in advance
Number and the vibration signal of bearing in normal state, and to vibration signal carry out feature extraction after remove Training RBF Neural Network, and
Build corrosion intelligent classification model；It can both be realized using the corrosion intelligent classification model afterwards and the real time fail of bearing was examined
It is disconnected.
Brief description of the drawings
Fig. 1 is characterized sensitiveness schematic diagram.
Fig. 2 is RBF neural topological structure.
Fig. 3 is bearing rust intelligent classification model construction flow chart.
The evaluation factor that Fig. 4 is characterized.
Fig. 5 is classification accuracy corresponding to preceding 120 features.
Fig. 6 is that bearing rust diagnostic model verifies flow.
Fig. 7 is data B test results.
Embodiment
For the object, technical solutions and advantages of the present invention are more clearly understood, below in conjunction with instantiation, and with reference to attached
Figure, the present invention is described in more detail.
A kind of bearing rust intelligent diagnosing method, it is as follows to specifically include step：
The vibration signal and bearing of step 1, in advance collection bearing in the case where there are all kinds of different faults states are in normal condition
Under vibration signal, and these vibration signals are divided into two classes, one kind is used as training sample, another kind of to be used as test sample；
Step 2, extract training characteristics from the training sample of step 1；I.e.
Step 2.1, elder generation extract time domain charactreristic parameter from the time domain of vibration signal；FFT is carried out to vibration signal again
Spectrum information is obtained, frequency domain character parameter is extracted from frequency spectrum；Hilbert envelope spectral transformations are carried out to vibration signal afterwards to be wrapped
Network spectrum information, frequency domain character parameter is extracted from envelope spectrum；
Step 2.2, first vibration signal is handled using redundant wavelet packet transform, obtain the decomposed signal of frequency band, and
Time domain charactreristic parameter is extracted respectively from the decomposed signal of each frequency band；Frequency is obtained to the decomposed signal progress FFT of frequency band again
Spectrum information, extract frequency domain character parameter respectively from each frequency spectrum；Hilbert envelope spectral transformations are carried out to the decomposed signal of frequency band afterwards
Envelope spectrum information is obtained, extracts frequency domain character parameter respectively from each envelope spectrum；
Step 2.3, first vibration signal is handled using redundancy multi-wavelet packets transform, obtains the decomposed signal of frequency band,
And time domain charactreristic parameter is extracted respectively from the decomposed signal of each frequency band；The decomposed signal progress FFT of frequency band is obtained again
Spectrum information, extract frequency domain character parameter respectively from each frequency spectrum；Hilbert envelope spectrum changes are carried out to the decomposed signal of frequency band afterwards
Get envelope spectrum information in return, extract frequency domain character parameter respectively from each envelope spectrum；
Step 2.4, the step 2.1-2.3 time domain charactreristic parameters extracted and frequency domain character parameter are formed and train spy
Sign；
Step 3, the susceptibility that all training characteristics obtained by step 2 are carried out is assessed, obtain commenting for each training characteristics
Estimate the factor, and training characteristics are ranked up according to the order of evaluation factor from big to small；
Step 4, that the training sample of training characteristics sort obtained by step 3 and step 1 gained is input into RBF is neural
Network is trained, and the RBF neural being trained is tested using the test sample obtained by step 1, is exported
Relation between classification accuracy and these Characteristic Numbers；
Step 5, the feature that can reach default classification accuracy is selected as the spy most sensitive to corrosion after completing classification
Levy subset；The input feature vector of the most sensitive character subset as RBF neural is selected, establishes corrosion intelligent classification model；
The real-time vibration signal composition of step 6, collection bearing in the course of the work treats diagnostic sample；And diagnostic sample will be treated
It is input in corrosion intelligent classification model, to complete to diagnose.
Below to key technology involved in the present invention, it is further elaborated：
(1) character subset is built
It can see from the characteristics of corrosion, corrosion has the structure distribution of dimension index, frequency spectrum and envelope spectrum to time domain more
Sensitivity, the character subset of intelligent classification model is selected and established below in terms of these.
In terms of time domain (time domain charactreristic parameter)：Select average：Variance：Root amplitude：Root-mean-square value：Peak value：p5=max | x (n) |, this five have dimension index.Formula
In：X (n) is time-domain signal sequence, n=1,2 ... N；N is sample points.
In terms of frequency spectrum and envelope spectrum (frequency domain character parameter), it have selected respectively This 13 characteristic parameter (p12~
p24).In formula：S (k) is signal x (n) frequency spectrum, k=1,2 ..., K；K is spectral line number；fkIt is the frequency scale of kth bar spectral line.
They describe the change of the size conversion of spectrum energy, point spread of distribution and main band position in frequency spectrum, such as frequency domain respectively
Characteristic parameter p12Reflect the size of frequency domain vibrational energy；p16And p18~p20Reflect the change of main band position；p13~p15, p17With
p21~p24Characterize the scattered or intensity of frequency spectrum.
In order to improve the accuracy rate of bearing rust feature extraction, and reflect the feature of bearing rust state, the present invention comprehensively
The corrosion bearing features subset of structure mainly includes following three parts：
1) 5 time domain charactreristic parameter (p are extracted from the vibration equipment original time domain signal collected1~p5), from original letter
Number frequency spectrum in extract 13 frequency domain character parameter (p12~p24) and extract 13 from the Hilbert envelope spectrums of primary signal
Individual frequency domain character parameter (p12~p24), i.e., 31 characteristic parameters altogether are extracted from primary signal, they fully reflect original
The time domain general picture and frequency domain general picture of vibration signal.
2) in order to catch more fault messages and corrosion character, using redundant wavelet packet transform to primary signal at
Reason.Redundancy resolution ensure that the uniformity and translation invariance of frequency resolution before and after decomposition, also overcome Gibbs phenomenons；It is small
Ripple packet transform then realizes the multiresolution analysis of the full frequency band of signal.Here orthogonal wavelet Db8 is selected to carry out 3 to primary signal
Layer redundant wavelet bag decomposes, and obtains the decomposed signal of 8 frequency bands.The feature extraction of similar primary signal, it is of the invention from 8 frequency bands
5 time domain charactreristic parameters are extracted in component respectively, obtain 8 × 5 characteristic values；Equally FFT is carried out to 8 band signals to obtain
To spectrum information, 13 frequency domain character parameters are extracted from frequency spectrum, obtain 8 × 13 characteristic values；Finally 8 band signals are entered
Row Hilbert envelope spectral transformations, 13 frequency domain character parameters are extracted from envelope spectrum, obtain 8 × 13 characteristic values.It is small from redundancy
248 characteristic parameters are obtained in ripple packet transform altogether.
3) m ultiwavelet has the incomparable advantage of many single wavelets, and in order to give full play to the advantage of m ultiwavelet, extraction is more
More useful informations, select redundancy multi-wavelet packets to be decomposed twice here, obtain the decomposed signal of 8 frequency bands.With it is small above
Ripple bag decompose it is identical, to extracting 8 × 5 temporal signatures values in this 8 subband-signal components, the frequency domain after 8 × 13 FFTs
Frequency domain character value after characteristic value and 8 × 13 Hilbert envelope spectral transformations.Decomposed from redundancy multi-wavelet packets and obtain 248 altogether
Individual characteristic parameter.
Therefore, the present invention is extracted 31+248+248=527 character subset altogether, and they can fully reflect vibration letter
The feature of number time domain, frequency spectrum and envelope spectrum.
(2) sensitive features select
(2.1) sensitive features selection general introduction
Fault diagnosis is substantially the identification problem of pattern, and during this fault signature selection and extraction be problem
Key, only select and extract can maximally utilise fault message, abundant faults essence it is most sensitive
Feature, it just can ensure that the accuracy and validity of fault diagnosis.The fault signature that different fault types is showed is often not
Equally, and redundancy and incoherent feature diagnostic accuracy can not be provided, new information can not be increased, these opposite features increase
Big data set and the burden calculated, the strong influence real-time of status monitoring simultaneously add cost.Therefore, how to reduce
Incoherent feature and the redundancy for reducing feature, i.e., select in multiple features of how comforming most sensitive feature will have it is important
Practical significance.
Feature subset selection is also in sensitive features selection, and it refers to selects N number of sensitive features to cause from existing M feature
The specific indexes optimization of system is that some most effective features are selected from primitive character to reduce intrinsic dimensionality, simplify number
According to process, be data prediction step crucial in pattern-recognition, be improve learning algorithm performance an important means.
As shown in figure 1, sensitive features subset is in bosom, its outside also has many completely unrelated features to refer to
Mark, if not rejecting no correlated characteristic, these extraneous features will disturb diagnostic result during diagnosis, increase calculate cost or
Person causes dimension disaster.Therefore, in order to improve the performance of grader, fault diagnosis efficiency is improved, sensitive features select the pole that seems
Its is important.
(2.2) feature evaluation and selection technique based on distance
527 characteristic values of bearing rust, a portion feature is closely related with failure, and another part is then nothing
The either redundancy feature closed, will be with fault signature the most by feature evaluation technology before feature set is input into grader
Related sensitive features extract.
Feature evaluation technology is feature-sensitive degree to be assessed apart from size between feature based, and its principle is：No
Similar class pitch characteristics distance is maximum, and characteristic distance is minimum in of a sort class, can meet the feature quilt of this principle
Be considered most sensitive feature, i.e., the inhomogeneous between class distance of a certain feature is bigger, of a sort inter- object distance is smaller, then this
Feature is more sensitive., must also be by the otherness of aggregation extent between class and class, and the difference of different between class distances when classification is more
The opposite sex takes in.
Assuming that one has C classes, M sample and per a kind of feature set containing J feature：
{qm,c,j, m=1,2 ..., Mc；C=1,2 ..., C；J=1,2 ..., J }
In formula：qm,c,j--- j-th of feature of m-th of sample of c classes.
Then comprised the following steps that based on the feature selection approach apart from assessment technology：
1) average distance in the class of all samples of same class is calculated
Then the average value of C inter- object distance is obtained
2) average value of each feature of all samples of same class is calculated
Then the average distance between inhomogeneity is obtained
3) evaluation factor of j-th of feature is calculated
It is clear that αjSize reflect the complexity that j-th of feature is classified to C class, bigger αjValue
Represent that j-th of feature is higher to the sensitiveness of failure, it can conform better to assess principle, so as to be easier to C classification
Classified.
Complete all features in feature set susceptibility assess, to carry out the selection of sensitive features subset also need combine point
The classifying quality of class device determines, i.e., by evaluation factor αjThe subsequent classifier that sequentially inputs from big to small is trained and tested,
The relation between classification accuracy and evaluation factor is obtained, further according to the selection criterion for needing setting sensitive features subset, is such as selected
Select classification accuracy and reach the threshold value of setting and (be such as arranged to a certain numerical value between 95%~100%, concrete numerical value is according to tool
Body problem is configured) when or Characteristic Number increase continuously N (it is more suitable during N=5 rule of thumb, to take) it is individual when classify it is accurate
When rate does not have any improvement, optimal sensitive features subset is obtained.Using the sensitive features subset that these choose as later right
The foundation that this C class is classified, so as to avoid the repeatability of extraction and selection feature.
(3) feature based assesses the corrosion bearing intelligent disaggregated model with neutral net
(3.1) radial basis function neural network
RBF (Radial Basis Function, RBF) neutral net has that simple in construction, training is succinct, learns
Habit speed is fast, has many advantages, such as local map feature and nonlinear function approximation capability are strong, can avoid local minimum
And global convergence is realized, the nonlinear function of complexity can be also approached with arbitrary accuracy, therefore, achieves and is widely applied.
RBF neural is a kind of three-layer forward networks, as shown in Fig. 2 comprising input layer, hidden layer and output layer, it is defeated
Enter space and transform to implicit sheaf space by Nonlinear Mapping, implicit sheaf space transforms to output layer sky by Linear Mapping again
Between.Wherein, the nonlinear mapping function that hidden layer uses is RBF, and it is a kind of local distribution to central point radial direction
The non-negative nonlinear function of balance attenuation.
Fig. 2 show the RBF neural of n-h-m structures, i.e., containing n input, h hidden node and m output.Its
In, x=(x1,x2,...,xn)TFor the input vector of network, Γi() is the activation primitive of i-th hidden layer node, W ∈ Rh ×mTo export weight matrix, export the Σ in node layer and represent the linear activation primitive that output layer neuron uses.Here footpath
To basic function Γi() uses the most frequently used Gaussian function, such as following formula：
In formula：X --- input vector；The central value of c --- Gaussian function；σi--- the extension constant of i-th of basic function
Or width；||x-ci||——x-ciNorm, i.e. x and ciThe distance between.
The output of radial basis function neural network is：
In general, the learning process of RBF neural is divided into two stages, and the first stage is determined by input sample distribution
Parameter (the center c of the Gaussian function of fixed each implicit nodeiWith sound stage width σi), it belongs to unsupervised learning；Second stage uses line
Property optimized algorithm obtains the weights before hidden layer and output layer, and it accelerates pace of learning and avoids local optimum, and belonging to has
Supervised learning.
The advantages of based on RBF neural, the present invention select it as the grader for diagnosing corrosion bearing here.
(3.2) bearing rust intelligent classification model
The present invention combines redundant wavelet packet transform technology, redundancy multi-wavelet packets transform technology, feature evaluation technology and RBF god
Corrosion bearing fault intelligent classification model is established through network, as shown in Figure 3.First, 3 are decomposed from primary signal, redundant wavelet bag
8 band signals and redundancy multi-wavelet packets after layer, which decompose extraction 5 in 8 band signals after 2 layers, has dimension time domain special
Sign, and 13 frequencies of extraction reflection band position, degree of scatter and energy size etc. respectively from their frequency spectrum and envelope spectrum
Characteristic of field, constitute a union feature collection；Then feature evaluation is carried out by feature evaluation technology, and presses their evaluation factors
It is ranked up from big to small；These features and training sample are input into RBF neural again to be trained, and to test data
Tested, the relation between output category accuracy rate and these Characteristic Numbers；Complete classification after select can complete 95% with
The feature of upper classification accuracy is as the character subset most sensitive to corrosion；This most sensitive character subset is selected as RBF
The input feature vector of neutral net, establish corrosion intelligent classification model.
(3.3) corrosion bearing sensitive features subset is extracted
In experimentation, vibration acceleration sensor is arranged on exports end bearing with the close epicyclic gearbox of faulty bearings
At seat, and 60 sample numbers under normal bearing state are gathered, the data length of each sample is 10000.6 made will be tested
Individual corrosion bearing fault part is divided into two groups, and every group 3, each failure part collecting sample number is 10, and the data of each sample are grown
Spend for 10000, i.e. sample number under every group of corrosion state is 30.By 30 samples of the bearing test data of normal condition and
The experimental data of first group of corrosion bearing forms a database containing 2 kinds of states, 60 samples, and this database is referred to as counting
According to collection A, wherein 30 samples are used for RBF network trainings, 30 samples are used for RBF network tests in addition；Under normal condition in addition
The database for 60 samples Han 2 kinds of states that the experimental data of the bearing signal of 30 samples and second group of corrosion bearing is formed claims
For data set B.Data set A mainly for assessment of with determine most sensitive character subset, and data set B be mainly used in checking should
The accuracy and validity of model and sensitive features subset.
Below according to bearing rust diagnostic model and experiment, bearing rust sensitive features are extracted, are the knot of corrosion bearing
By offer premise and basis.
1) each training sample original vibration signals of data set A, 8 redundant wavelet bag decomposition frequency band signals and 8 are extracted
5 temporal signatures indexs, 13 frequency spectrum frequency domain characteristic indexs and 13 envelopes of individual redundancy multi-wavelet packets decomposition frequency band signal
Frequency domain characteristic index is composed, 527 features are obtained, decomposition frequency band corresponding to each feature or the characteristic index of expression such as table 1
It is shown.
Feature set selected by the corrosion of table 1
2) susceptibility assessment is carried out to 527 features, its susceptibility size is as shown in Figure 4.The classification of sensitive features selection
Accuracy rate threshold value is arranged to 95%, and the evaluation factor α that will be obtained after feature evaluationjThe value of (j=1,2 ..., 527) press from
Small order is arrived greatly by input RBF neural.When reaching termination setting thresholding, RBF input feature vectors are 13, that is, are evaluated
13 sensitive features, with the 13rd feature evaluation value, i.e., with α432=11.34 be that boundary draws a line, then sensitive features are located at
Horizontal line and on.
The descending arrangement of this 13 sensitive features assessed values and the feature corresponding to them, as shown in table 2.Can by table
See, sensitive features are concentrated mainly among time domain has dimension index, and only have in terms of frequency domain redundancy m ultiwavelet decompose after envelope
There are 3 features for representing envelope spectrum main band change in location and intensity in spectrum.And the frequency-domain index of FFT spectrums is to the shape of corrosion
State classification all shows as insensitive.It can see simultaneously from table, sensitive features number will be distal to more after multi-wavelet packets redundancy resolution
Redundant wavelet bag decompose after sensitive features number, also reflected from this respect m ultiwavelet due to multiple wavelet basis functions with
And many characteristics excellent compared with single wavelet so that it can show effect more more preferable than single wavelet.Either using wavelet packet also
It is that multi-wavelet packets decomposition can obtain more fault characteristic informations, while also obtains substantial amounts of garbage, only passes through spy
Sign, which is assessed, can just draw to useful sensitivity of classifying, and remove not useless feature, and so as to reduce network size, it is accurate to improve classification
Rate.
The corrosion sensitive features of table 2 and corresponding index
3) according to the descending order of evaluation factor, as the input feature vector of RBF neural, make from 1 and increasing one by one
527 are added to, and it is trained and tested, makes the relation of classification accuracy and input feature vector number.Because from 107 spies
It is all 50% that value indicative, which starts to 527 characteristic value its classification accuracies, becomes apparent from only giving preceding 120 characteristic value pair to show
The classification accuracy answered.
From fig. 5, it is seen that classification accuracy corresponding to 13 sensitive features values above has all reached 100%, when
After characteristic is more than 13, the decline of classification accuracy can be caused on the contrary by increasing the input of characteristic, and this is due to increased below
Classification of the feature to this 2 states is insensitive, increase their input can so that the difference between state becomes more to obscure, from
And cause the reduction of classification accuracy.
(3.4) corrosion bearing intelligent disaggregated model experimental verification
Corrosion bearing intelligent disaggregated model has been established above, and has been trained and tested by experimental data set A, has obtained
The sensitive features of 13 corrosions, in order to verify that selected sensitive features distinguish the ability and corrosion axle of normal bearing and corrosion bearing
Hold the validity of intelligent classification model.The corrosion for selecting data set B sensitive features to be trained as input to data set A below
Diagnostic model is tested, and flow is as shown in Figure 6.
It is normal that state 2 is corrosion that data set B, which includes normal and corrosion two states, i.e. state 1,.Select data set B
In this 13 sensitive features the RBF neural trained through data set A is tested as input feature vector, test result
As shown in Figure 7.
As can be seen from Figure 7,30 samples of state 1 can be distinguished 29 when being tested with the model data B, 30 shapes
The sample of state 2 also separates 29, and two states 2 mistakes occur and divide sample altogether, and obtained test accuracy rate is 96.67%.
Therefore, from checking above it can be seen that 13 sensitive features subsets that the present invention extracts are distinguishing normal and corrosion
It is validity that is accurate and effective, while also demonstrating the corrosion intelligent classification model during the two states.
It should be noted that although embodiment of the present invention is illustrative above, but it is to the present invention that this, which is not,
Limitation, therefore the invention is not limited in above-mentioned embodiment.Without departing from the principles of the present invention, it is every
The other embodiment that those skilled in the art obtain under the enlightenment of the present invention, it is accordingly to be regarded as within the protection of the present invention.
Claims (3)
1. a kind of bearing rust intelligent diagnosing method, it is characterized in that, it is as follows to specifically include step：
The vibration signal and bearing of step 1, in advance collection bearing in the case where there are all kinds of different faults states is in normal state
Vibration signal, and these vibration signals are divided into two classes, one kind is used as training sample, another kind of to be used as test sample；
Step 2, extract training characteristics from the training sample of step 1；I.e.
Step 2.1, elder generation extract time domain charactreristic parameter from the time domain of vibration signal；FFT is carried out to vibration signal again to obtain
Spectrum information, frequency domain character parameter is extracted from frequency spectrum；Hilbert envelope spectral transformations are carried out to vibration signal afterwards and obtain envelope spectrum
Information, frequency domain character parameter is extracted from envelope spectrum；
Step 2.2, first vibration signal is handled using redundant wavelet packet transform, obtain the decomposed signal of frequency band, and from each
Time domain charactreristic parameter is extracted respectively in the decomposed signal of frequency band；Frequency spectrum letter is obtained to the decomposed signal progress FFT of frequency band again
Breath, extracts frequency domain character parameter respectively from each frequency spectrum；The decomposed signal progress Hilbert envelope spectral transformations of frequency band are obtained afterwards
Envelope spectrum information, extract frequency domain character parameter respectively from each envelope spectrum；
Step 2.3, first vibration signal is handled using redundancy multi-wavelet packets transform, obtain the decomposed signal of frequency band, and from
Time domain charactreristic parameter is extracted respectively in the decomposed signal of each frequency band；Frequency spectrum is obtained to the decomposed signal progress FFT of frequency band again
Information, extract frequency domain character parameter respectively from each frequency spectrum；The decomposed signal progress Hilbert envelope spectral transformations of frequency band are obtained afterwards
To envelope spectrum information, frequency domain character parameter is extracted respectively from each envelope spectrum；
Step 2.4, the step 2.1-2.3 time domain charactreristic parameters extracted and frequency domain character parameter formed into training characteristics；
Step 3, the susceptibility that all training characteristics obtained by step 2 are carried out is assessed, obtain the assessments of each training characteristics because
Son, and training characteristics are ranked up according to the order of evaluation factor from big to small；
Step 4, the training sample of training characteristics sort obtained by step 3 and step 1 gained is input to RBF neural
It is trained, and the RBF neural being trained is tested using the test sample obtained by step 1, output category
Relation between accuracy rate and these Characteristic Numbers；
Step 5, the feature that can reach default classification accuracy is selected as feature most sensitive to corrosion after completing classification
Collection；The input feature vector of the most sensitive character subset as RBF neural is selected, establishes corrosion intelligent classification model；
The real-time vibration signal composition of step 6, collection bearing in the course of the work treats diagnostic sample；And it will treat that diagnostic sample inputs
Into corrosion intelligent classification model, to complete to diagnose.
2. a kind of bearing rust intelligent diagnosing method according to claim 1, it is characterized in that, in step 2, extracted when
Characteristic of field parameter includes And p5=
max|x(n)|；Wherein x (n) is time-domain signal, n=1,2 ... N；N is sample points.
3. a kind of bearing rust intelligent diagnosing method according to claim 1, it is characterized in that, in step 2, the frequency that is extracted
Characteristic of field parameter includes： WithWherein s (k) is time-domain signal x (n) frequency spectrum, k=1,2 ...,
K；K is spectral line number；fkIt is the frequency scale of kth bar spectral line.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711176681.7A CN107832729A (en) | 2017-11-22 | 2017-11-22 | A kind of bearing rust intelligent diagnosing method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711176681.7A CN107832729A (en) | 2017-11-22 | 2017-11-22 | A kind of bearing rust intelligent diagnosing method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107832729A true CN107832729A (en) | 2018-03-23 |
Family
ID=61652295
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711176681.7A Pending CN107832729A (en) | 2017-11-22 | 2017-11-22 | A kind of bearing rust intelligent diagnosing method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107832729A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108544303A (en) * | 2018-03-30 | 2018-09-18 | 上海交通大学 | A kind of main shaft of numerical control machine tool fault diagnosis method and system |
| CN109520738A (en) * | 2018-10-25 | 2019-03-26 | 桂林电子科技大学 | Rotating machinery Fault Diagnosis of Roller Bearings based on order spectrum and envelope spectrum |
| CN110334562A (en) * | 2018-03-30 | 2019-10-15 | 北京金风慧能技术有限公司 | Bear vibration operating status prediction model training method and prediction technique, device |
| CN112487910A (en) * | 2020-11-24 | 2021-03-12 | 中广核工程有限公司 | Fault early warning method and system for nuclear turbine system |
| CN115406656A (en) * | 2022-08-29 | 2022-11-29 | 桂林电子科技大学 | Bearing corrosion intelligent diagnosis method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100023307A1 (en) * | 2008-07-24 | 2010-01-28 | University Of Cincinnati | Methods for prognosing mechanical systems |
| CN103926097A (en) * | 2014-04-03 | 2014-07-16 | 北京工业大学 | Method for collecting and extracting fault feature information of low-speed and heavy-load device |
| CN104198184A (en) * | 2014-08-11 | 2014-12-10 | 中国人民解放军空军工程大学 | Bearing fault diagnostic method based on second generation wavelet transform and BP neural network |
-
2017
- 2017-11-22 CN CN201711176681.7A patent/CN107832729A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100023307A1 (en) * | 2008-07-24 | 2010-01-28 | University Of Cincinnati | Methods for prognosing mechanical systems |
| CN103926097A (en) * | 2014-04-03 | 2014-07-16 | 北京工业大学 | Method for collecting and extracting fault feature information of low-speed and heavy-load device |
| CN104198184A (en) * | 2014-08-11 | 2014-12-10 | 中国人民解放军空军工程大学 | Bearing fault diagnostic method based on second generation wavelet transform and BP neural network |
Non-Patent Citations (2)
| Title |
|---|
| SHUILONG HE ET.AL: "Multifractal entropy based adaptive multiwavelet construction and its application for mechanical compound-fault diagnosis", 《MECHANICAL SYSTEMS AND SIGNAL PROCESSING》 * |
| 雷亚国: "混合智能技术及其在故障诊断中的应用研究", 《万方数据 知识服务平台》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108544303A (en) * | 2018-03-30 | 2018-09-18 | 上海交通大学 | A kind of main shaft of numerical control machine tool fault diagnosis method and system |
| CN110334562A (en) * | 2018-03-30 | 2019-10-15 | 北京金风慧能技术有限公司 | Bear vibration operating status prediction model training method and prediction technique, device |
| CN110334562B (en) * | 2018-03-30 | 2022-10-28 | 北京金风慧能技术有限公司 | Bearing vibration operation state prediction model training method, prediction method and prediction device |
| CN109520738A (en) * | 2018-10-25 | 2019-03-26 | 桂林电子科技大学 | Rotating machinery Fault Diagnosis of Roller Bearings based on order spectrum and envelope spectrum |
| CN112487910A (en) * | 2020-11-24 | 2021-03-12 | 中广核工程有限公司 | Fault early warning method and system for nuclear turbine system |
| CN115406656A (en) * | 2022-08-29 | 2022-11-29 | 桂林电子科技大学 | Bearing corrosion intelligent diagnosis method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107832729A (en) | A kind of bearing rust intelligent diagnosing method | |
| CN109612708B (en) | Power transformer on-line detection system and method based on improved convolutional neural network | |
| CN109768985B (en) | Intrusion detection method based on flow visualization and machine learning algorithm | |
| CN104712542B (en) | A kind of reciprocating compressor sensitive features based on Internet of Things are extracted and method for diagnosing faults | |
| Yu et al. | A bearing fault diagnosis technique based on singular values of EEMD spatial condition matrix and Gath-Geva clustering | |
| CN109033612A (en) | A kind of Diagnosis Method of Transformer Faults based on vibration noise and BP neural network | |
| CN106897738A (en) | A kind of pedestrian detection method based on semi-supervised learning | |
| CN108388927A (en) | Small sample polarization SAR terrain classification method based on the twin network of depth convolution | |
| CN109186973A (en) | A kind of mechanical failure diagnostic method of unsupervised deep learning network | |
| CN106323636A (en) | Adaptive extraction and diagnosis method for degree features of mechanical fault through stack-type sparse automatic coding depth neural network | |
| CN108710889A (en) | A kind of scarce cylinder method for diagnosing faults of automobile engine | |
| CN105841961A (en) | Bearing fault diagnosis method based on Morlet wavelet transformation and convolutional neural network | |
| CN104052612B (en) | A kind of Fault Identification of telecommunication service and the method and system of positioning | |
| CN106599920A (en) | Aircraft bearing fault diagnosis method based on coupled hidden semi-Markov model | |
| CN110044623A (en) | The rolling bearing fault intelligent identification Method of empirical mode decomposition residual signal feature | |
| CN108827605A (en) | A kind of mechanical breakdown characteristic automatic extraction method based on improvement sparseness filtering | |
| CN106628097A (en) | Ship equipment fault diagnosis method based on improved radial basis function neutral network | |
| CN108445868A (en) | A kind of automobile intelligent fault diagnosis system and method based on modern signal processing technology | |
| CN110501122A (en) | A kind of adaptive leakage detection method of pressure vessel based on cluster | |
| CN105160359B (en) | A kind of labyrinth damage collaboration recognition methods based on supersonic guide-wave | |
| Himawan et al. | Deep Learning Techniques for Koala Activity Detection. | |
| CN110020637A (en) | A kind of analog circuit intermittent fault diagnostic method based on more granularities cascade forest | |
| CN110110809A (en) | The construction method of fuzzy automata based on Machine Fault Diagnosis | |
| CN109813542A (en) | The method for diagnosing faults of air-treatment unit based on production confrontation network | |
| CN106443380B (en) | A kind of distribution cable local discharge signal recognition methods and device |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180323 |