CN103528836B - The rotary machinery fault diagnosis method of word pattern is prohibited based on symbolic dynamics - Google Patents
The rotary machinery fault diagnosis method of word pattern is prohibited based on symbolic dynamics Download PDFInfo
- Publication number
- CN103528836B CN103528836B CN201310432071.4A CN201310432071A CN103528836B CN 103528836 B CN103528836 B CN 103528836B CN 201310432071 A CN201310432071 A CN 201310432071A CN 103528836 B CN103528836 B CN 103528836B
- Authority
- CN
- China
- Prior art keywords
- sequence
- sample
- class
- parameter
- word pattern
- 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.)
- Expired - Fee Related
Links
Abstract
The invention discloses the rotary machinery fault diagnosis method prohibiting word pattern based on symbolic dynamics, it is characterized in that: comprise the steps: the first step: set up and need the time-domain signal sample of diagnostic device under normal condition and various malfunction; Second step: to each sample, carries out symbolism process, becomes codomain sequence by sample; 3rd step: codomain sequence is become symbol sebolic addressing: the 4th step: by symbol sebolic addressing generating feature vector, specifically comprise four parameter group; 5th step: after standardization and normalized are carried out to proper vector, then carry out PCA dimension-reduction treatment; 6th step: generate the new proper vector after dimensionality reduction, and standardization and normalized are carried out to new proper vector; 7th step: adopt the LibSVM of Layering memory to diagnose, utilize gridding method to obtain g parameter and penalty parameter c parameter, and carry out training modeling; 8th step: utilize model to test test sample book; The present invention can be widely used in the fault diagnosis of various rotating machinery.
Description
Technical field
The present invention relates to method for diagnosing faults, be specifically related to the rotary machinery fault diagnosis method prohibiting word pattern based on symbolic dynamics.
Background technology
Current, rotating machinery has been widely applied in the fields such as Aero-Space, automobile boats and ships, metallurgical petrochemical industry, weapons equipment and energy generating.Comparatively typical rotating machinery as gear, rotary main shaft, bearing etc., the fault diagnosis of rotating machinery is the hot issue of engineering circles application always.Rotary machinery fault diagnosis process is generally: signals collecting, signal transacting and fault diagnosis.Fault diagnosis is actually a kind of mode identification procedure.
The method for diagnosing faults of rotating machinery has a lot.Generally be divided into based on kinetic model, based on dynamic signal analysis and the mode based on data-driven.Based on the diagnostic method of kinetic model, be generally need to set up complete kinetics equation, by math equation, describe the dynamic characteristic of rotating machinery.Under normal circumstances, kinetic model is not easy to set up very much, and when breaking down, its model can change; Dynamic signal analysis method is a kind of method that current application obtains a lot, and its method carries out signal transacting to time domain or frequency-region signal, extracts the special composition that wherein can characterize defect or fault, by mode identification method, has judged whether fault; Based on the method for data-driven, being the various sample of signal based on obtaining, adopting supervision or non-supervisory method, by extracting the feature inside data, setting up the relation between failure modalities and feature, the method for implementation pattern identification fault diagnosis.
The difficulty that current method for diagnosing faults exists is: one. be difficult to set up the kinetic model of accurate rotating machinery.The dynamic characteristic of fault is the most fundamental way of carrying out fault diagnosis.Due to the driving source of system Planar Mechanisms characteristic and inside complexity, machinery is made to be that vibration coupling is comparatively outstanding in work; Meanwhile, face the fluctuation such as shock load, intermittent load variability operating mode during work, this directly results in the non-stationary of response signal.These two factor interactions, make fault be that response signal has obvious nonlinear characteristic.Which results in rotating machinery essence difficulty in diagnostic techniques.Its two: be difficult to effectively overcome response signal signal to noise ratio (S/N ratio) in row actual condition low.The noise of the rotating machinery of actual condition is all relatively more serious, and its characteristic frequency of rotating machinery is distributed in the more of low frequency medium frequency more, and these characteristic frequencies are as easy as rolling off a log by pollution from environmental noise.Its three: characteristic parameter need optimize and diagnostic model robustness need raising.Although have a lot of meaningful achievement in research in fault signature extraction, Fault Pattern Recognition, in some rotating machinery, its characteristic frequency is very complicated and be difficult to determine.In the diagnosis of reality, the optimization of characteristic parameter and the robustness of diagnostic model improve the important path of diagnosis and prediction accuracy, minimizing reject rate and misclassification rate.
Rotating machinery is actually a complicated dynamical system, and its specifically detailed dynamical structure is difficult to understand.When rotating machinery normally works, its dynamical system is stable, namely its kinetic parameter is should be stable within a stable scope, when machinery breaks down, its dynamic characteristic will change, different change directions, and different change degree, correspond to certain fault.Particularly importantly, when system malfunctions, the Nonlinear Dynamical Characteristics of system will there will be, the Nonlinear Dynamical Characteristics that different faults is corresponding different, if these Nonlinear Dynamical Characteristics can be characterized to a certain extent, by the method establishment fault model of data-driven, just fault diagnosis can be carried out very well.Accordingly, this method proposes a kind of diagnostic method based on time-domain signal symbolic dynamics disable word pattern.
Symbolic dynamics is a branch of system dynamics, since eighties of last century seventies, has been applied in the middle of engineering reality gradually from pure research.The thought of symbolic dynamics is carried out " coarse grain " to time-domain signal sequence exactly and is changed process, forms symbol sebolic addressing.When considering to embed peacekeeping delay parameter, simplification and abstract is carried out to system.According to the symbolism method of Bants & Pompe, time-domain signal can be analyzed to ergodic word pattern sequence.The method of usual symbolization dynamics research can be divided into word pattern statistical information, ordinary word pattern and disable word pattern.The dynamic characteristic of symbol sebolic addressing energy retention system, not necessarily can meet embedding theorems, and be easy to carry out data modeling process, therefore symbolic dynamics is in fault diagnosis and prediction etc. simultaneously, has started to be subject to the concern of people and has carried out correlative study.The method of word pattern statistical information is expressed the parameter of uncertainty, as entropy, complexity etc., carrys out the dynamic characteristic of analytic system.Nineteen ninety-five, the WernerEbeling symbolization entropy of Germany studies the statistical distribution information in a large amount of text; The MihajloGrbovic of Temple university of the U.S. etc. 2012 propose the distribution sensor method for diagnosing faults of a kind of Based PC A and maximum entropy fuzzy decision; Argentine L.Zunino is by the Stochastic sum chaotic characteristic of multiple dimensioned complexity-entropy plane discrimination system; The Shuen-DeWu of the state-run normal university in Taiwan in 2012 builds the characteristic parameter of diagnosis by multiple dimensioned sequence entropy, then utilize SVM to carry out the fault diagnosis of automobile bearing; Domestic aspect: acting like a bully etc. of Institute of Automation, CAS in 2005 proposes a kind of division methods of self-adaptive symbol interval, and thus to the diagnosing malfunction of motor; Within 2007, Xi'an hands over large Li Chongsheng etc. to utilize the information entropy of chaos and sequence number sequence to carry out rolling bearing fault diagnosis; The Chen Xiao equality of Jiangsu University proposes the YongyongHe of the Fault Diagnosis of Roller Bearings Tsing-Hua University based on symbol entropy and SVM, waits 2012 and proposes approximate entropy as fault diagnosis nonlinear characteristic parameters; The BingYu of Nanjing space flight and aviation university proposed the characteristic parameter of Wavelet Entropy as sensor fault diagnosis in 2011; The Zheng Jinde of Hunan University etc. propose a kind of Fault Diagnosis of Roller Bearings based on multi-scale entropy, embed the Sample Entropy of the sequence of peacekeeping tolerance limit by obtaining the difference after coarse, and recycling SVM carries out fault diagnosis; The Tang Youfu etc. of Shanghai Communications University in 2012 proposes the compressor Representative Faults Diagnosis method of non linear complexity, after signal is carried out binary symbol sequence process, the LZC of the sequence of calculation, is divided by the characteristic interval of the LZC of different faults state and diagnoses; The Xue Tinggang of Xi'an University of Technology etc. extract the packet multi-scale information entropy of the EMD decomposed signal of hydraulic turbine tail pipe as the characteristic parameter of Fault Pattern Recognition, analyze complicated pulsating pressure signal, etc.
Disable word pattern, refers to the word pattern not occurring in the ergodic process of symbol sebolic addressing or seldom occur.At present based on the research of disable word pattern, be mainly used in system features identification, and research is not still deeply.2000, there are related notion and the expression method of disable word in literature research symbolic dynamics; 2007, literature research is had to determine the distribution situation of true in stochastic system, pseudo-disable word pattern; There is the application of literature research disable word pattern in stock market, have studied the feature of system disable word pattern and the impact on stock behavior under different delay; 2008, there is the distribution situation of disable word pattern in a large amount of economical time serial data of document analysis; 2012, the people such as OsvaldoA.Rosso of Brazil have studied in entropy-complexity plane, have studied the sensitivity of disable word pattern to sequence length and noise.Up to the present, there is not yet the document adopting the method for disable word pattern to carry out fault diagnosis.
The characteristic present of the disable word pattern during symbol the is ergodic non-linear dynamic feature of system.Therefore, when adopting disable word model to have the fault of different Nonlinear Dynamical Characteristics as research, there is natural undeniable advantage.Inherently see, when system malfunctions, its dynamic behavior will change, and nonlinear characteristic will occur or occur state and change, and now the disable word model of sequence number sequence also changes thereupon; From the sensitivity to noise, when the symbol sebolic addressing length analyzed meets certain condition, disable word pattern is to Gaussian noise, coloured noise and f
-λnoise pollution will be completely immune; Even length is inadequate, when there is certain level in noise, noise on the impact of disable word pattern also by statistical study process, the impact of stress release treatment; From operability or algorithm complex, the quantity of disable word pattern is far smaller than ordinary pattern, therefore adopt the mode of data-driven to study relation between disable word pattern and fault, carry out characteristic parameter optimization, it is practicable for building sane diagnostic model.
Summary of the invention
Problem to be solved by this invention is to provide the rotary machinery fault diagnosis method prohibiting word pattern based on symbolic dynamics.
In order to solve the problems of the technologies described above, first technical scheme of the present invention is, prohibits the rotary machinery fault diagnosis method of word pattern, it is characterized in that: comprise the steps: based on symbolic dynamics
The first step: set up and need the time-domain signal sample of diagnostic device under normal condition and various malfunction to be respectively no less than 30, the sampling length of each sample is not less than 4096 points;
Second step: to each sample, carries out symbolism process, becomes codomain sequence samples by time domain sequences; Be divided into M region by each sample equably by its peak-to-peak value scope, M is 6; Carry out region to each point on each sample to choose, the region at the value place namely on sample is the value in codomain sequence;
Thus, time domain sequences becomes codomain sequence; M: signal coarse degree discrimination; Namely between signal codomain peak-to-peak value, the interval number of equidistant segmentation; Coarse degree is different, also different to the simulation degree of original signal; If too little, some signal detail will be submerged, if too large, then coarse has little significance; According to the previous experiments research of related data and this method, think that the situation of M=6 is applicable to the signal coarse process in rotary machinery fault diagnosis;
3rd step, symbol sebolic addressing are formed:
The formation of symbol sebolic addressing is by delay parameter τ and go through length I and determine, start in the codomain sequence obtained in second step from the point that sampling instant is minimum, using (L-I* τ+1) individual point before the sampled point of sample is got according to sampling instant order from small to large as start point signal, start point signal is added the starting point time delay τ with signal, 2 τ, 3 τ ... (I-1) point that τ is corresponding forms one group of sequence respectively, and distinguish time delay τ by start point signal in sequence and with start point signal, 2 τ, 3 τ ... (I-1) point that τ is corresponding carries out symbol definition successively, again according to time-domain signal amplitude putting in order from big to small, symbol sebolic addressing set or word set, wherein
L: the length of sample of signal;
τ: delay parameter, τ get the natural number of 1 to 6; τ was 1 expression time delay one cycle, and namely last starting point and a rear starting point are adjacent sampled point, and τ is 2 expression time delay two cycles, namely last starting point and a rear starting point are spaced a sampled point, the like, τ was 6 expression time delay six cycles, and namely last starting point and a rear starting point are spaced five sampled points;
Go through length I and get 6;
4th step: generating feature vector, comprising:
4th. a step: generate parameter group one, is specially and calculates total taboo word rate f
rwith the information entropy S of symbol sebolic addressing
h; Wherein:
Wherein:
Total disable word rate:
N
ffor all disable word pattern-word quantity occurred;
Mt is word pattern sum; M
t=M unequal to 720;
4th. two steps: generate parameter group two, be specially:
For in whole word pattern sum, be divided into 30 groups in order, each group has 24 patterns, to every class symbol sequence, adds up respectively in 24 patterns, the number N that disable word pattern occurs
i; By the N occurred inside each group
idivided by total disable word quantity, obtain 30 disable word rate f
1, f
2..., f
30;
Parameter group two is actually and represents disable word pattern and distribute in the one of certain scale;
4th. three steps: generate parameter group three, be specially:
In whole word pattern sum, be divided into six groups in order, each group has 120 patterns, to every class symbol sequence, adds up the disable word pattern count N ' occurred inside 120 patterns respectively
i, then forbid number of words N divided by total
f, obtain six disable word rate f
b1, f
b2..., f
b6; Wherein:
In fact parameter group three represents the one distribution of disable word pattern on larger scale;
4th. four steps: generate parameter group four, be specially:
Every class symbol sequence is divided into two sections, front and back, and every section is then made up of three characters; According to the size position relationship of these three characters, namely every three characters have four kinds of position relationships, so two sections, front and back, have 16 kinds of patterns, inside whole word pattern sum, to every class symbol sequence, add up the disable word pattern quantity N occurred under 16 kinds of patterns respectively
ci, then divided by total disable word pattern quantity, form disable word rate f
ci;
In word set, M
t, also referred to as dictionary table.Each sample of signal is generated to the word pattern set obtained, length L adds up the pattern do not appeared at inside mod table, or with the pattern that very low probability occurs, be called disable word pattern.Disable word pattern is not only relevant with the dynamic characteristic of signal, also relevant with the noise signal be loaded on signal.But no matter be white noise, pink noise or f
-λnoise, is loaded on signal, after L reaches certain length, is not large to the quantity of the disable word pattern of signal, distribution influence.Therefore, after adopting the method for disable word pattern, to antimierophonic very capable.
5th step, standardization and normalized are carried out to proper vector after, and carry out PCA dimension-reduction treatment;
New proper vector after 6th step, generation dimensionality reduction, and standardization and normalized are carried out to new proper vector;
The LibSVM of the 7th step, employing Layering memory diagnoses, and kernel function adopts RBF, utilizes gridding method to obtain RBF nuclear parameter g parameter and the penalty parameter c parameter of LibSVM, and carries out training modeling;
8th step: utilize model to test test sample book, judges whether every class accuracy is greater than 90%; If every class accuracy is greater than 90%, judge whether merging class; If do not merge class, model construction is complete, can carry out diagnostic work; If any merging class, then π=X is set, τ=1, M=6, returns second step; Do you if there is accuracy to be less than 90%, judge τ < 6? as τ < 6, then establish τ=τ+1, return second step; As τ=6, then judge to have the classification of accuracy >90%, as not having, then this diagnoses unsuccessfully, cannot modeling, can not diagnose; As there is the classification of accuracy >90% τ=6, then observe when τ=1, the class mutually obscured not reaching accuracy 90% is merged into a class, and arranges M=6, π=π-X+1, τ=1, returns second step; Wherein: π is for identifying object, and N is total fault category number to be sorted; X is the classification number of the class merged.
The beneficial effect prohibiting the rotary machinery fault diagnosis method of word pattern based on symbolic dynamics of the present invention is: the disable word pattern inside symbolization dynamics of the present invention, multiple disable word mode statistical parameter is adopted to be proper vector parameter, utilize Layering memory LibSvm method, carry out multistage fault diagnosis, can when strong noise pollution, identify mechanical fault, recognition correct rate is higher, can be widely used in rotary machinery fault diagnosis.
Accompanying drawing explanation
Fig. 1 is the rotary machinery fault diagnosis method process flow diagram prohibiting word pattern based on symbolic dynamics of the present invention.
Fig. 2 is second and third step exemplary plot of rotary machinery fault diagnosis method of prohibiting word pattern based on symbolic dynamics of the present invention.
Fig. 3 (a) is the wherein a kind of position relationship generating three characters in parameter group four.
Fig. 3 (b) is the wherein the second position relationship generating three characters in parameter group four.
Fig. 3 (c) is wherein the third position relationship generating three characters in parameter group four.
Fig. 3 (d) is wherein the 4th kind of position relationship generating three characters in parameter group four.
Embodiment
See Fig. 1, prohibit the rotary machinery fault diagnosis method of word pattern based on symbolic dynamics, carry out as follows:
The first step: set up and need the time-domain signal sample of diagnostic device under normal condition and various malfunction to be respectively no less than 30, the sampling length L of each sample is not less than 4096 points;
Second step, to each sample, carry out symbolism process, become codomain sequence by sampled point; Be divided into M region by each sample equably by its peak-to-peak value scope, M is 6; Carry out region to each point on each sample to choose, the region at the value place namely on sample is the value in symbol sebolic addressing; Thus, time domain sequences becomes codomain sequence; M: signal coarse degree discrimination; Namely between signal codomain peak-to-peak value, the interval number of equidistant segmentation;
Specifically see Fig. 2, get M=6, between peak-to-peak value, namely divide 6 equal altitudes intervals; Then the codomain sequence of 11 sampled points in time domain waveform foremost is 15462534661;
3rd step: codomain sequence is become symbol sebolic addressing:
The formation of symbol sebolic addressing is by delay parameter τ and go through length I and determine; Delay parameter τ be not more than 6 natural number; Using (L-I* τ+1) individual point before the sampled point of sample is got according to sampling instant order from small to large as start point signal; Start point signal is added and the starting point time delay τ of signal, 2 τ, 3 τ ... (I-1) point that τ is corresponding forms one group of sequence respectively, and distinguish time delay τ, 2 τ, 3 τ by start point signal in sequence and with start point signal ... (I-1) point that τ is corresponding carries out symbol definition successively, again according to time-domain signal amplitude putting in order from big to small, symbol sebolic addressing set or word set; Wherein, τ: delay parameter, τ get the natural number of 1 to 6; τ was 1 expression time delay one cycle, and namely last starting point and a rear starting point are adjacent sampled point, and τ is 2 expression time delay two cycles, namely last starting point and a rear starting point are spaced a sampled point, the like, τ was 6 expression time delay six cycles, and namely last starting point and a rear starting point are spaced five sampled points;
Go through length I and get 6 points;
In the specific implementation, definable goes through length I=M, by minimum sampling instant in the codomain sequence obtained in second step corresponding o'clock as the starting point of first group of sequence, the starting point of first group of sequence is added this starting point time delay τ, 2 τ, 3 τ ... (I-1) some formation first group of sequence that τ is corresponding, again using corresponding with the starting point time delay τ of first group of sequence o'clock starting point as second group of sequence, the starting point of the second class symbol sequence is added and the starting point time delay τ of the second class symbol sequence, 2 τ, 3 τ ... (I-1) some formation second group of sequence that τ is corresponding, the like, finally is organized the starting point of sequence at corresponding with starting point time delay (L-I* τ+1) τ of first group of sequence o'clock as (L-I* τ+1), the starting point of (L-I* τ+1) class symbol sequence is added the starting point time delay τ of (L-I* τ+1) class symbol sequence, 2 τ, 3 τ ... (I-1) some formation (L-I* τ+1) the group sequence that τ is corresponding, and distinguish time delay τ by start point signal in sequence and with start point signal, 2 τ, 3 τ ... (I-1) point that τ is corresponding carries out symbol definition successively, again according to time-domain signal amplitude putting in order from big to small, final formation (L-I* τ+1) individual symbol sebolic addressing, namely symbol sebolic addressing set or word set are { s
i,
Specifically see Fig. 2, totally 11 points, the delay parameter between 2, front and back is τ; Length I=M=6 is gone through in definition; τ=1; Using t1 corresponding o'clock as the starting point of first group of sequence, six points that t1 and t2 to t6 is corresponding form first group of sequence, and the codomain of these six points is 154625; T1 to t6 is set to " 1 ", " 2 ", " 3 ", " 4 ", " 5 ", " 6 "; Also other symbols can be set to; And according to time-domain signal amplitude putting in order from big to small, obtain the first class symbol sequence s
1=426321; In like manner, by t
2the corresponding some starting point as second group of sequence, six points that t2 and t3 to t7 is corresponding form second group of sequence, and the codomain of these six points is 546253; T2 to t7 is set to " 1 ", " 2 ", " 3 ", " 4 ", " 5 ", " 6 "; And according to time-domain signal amplitude putting in order from big to small, obtain the second class symbol sequence S2=315264; The like, the codomain of the 3rd group of sequence is 462534, obtains the 3rd class symbol sequence S3=241653; The codomain of the 4th group of sequence is 625346, obtains the 4th class symbol sequence S4=163542; After be not repeated.In fig. 2, as τ=2, should with some formation first group of sequence corresponding to t1, t3, t5, t7, t9, t11; Second group of sequence should be that starting point is with t3, t5, t7, t9 with t3 ... form second group of sequence Deng six corresponding points, remaining is tired states.4th step, generating feature vector, comprising:
4th. a step, generation parameter group one: parameter group one considers the overall objective of symbol sebolic addressing, and one has two, is total taboo word rate f respectively
rwith the information entropy S of sequence
h:
Total disable word rate:
N
ffor all disable word pattern-word quantity occurred;
Mt is word pattern sum; M
t=M unequal to 720;
S
hinformation entropy for symbol sebolic addressing:
The wherein probability that occurs for sampled point of p (π); π ∈ ∏ (L) is the arrangement of certain word arbitrarily;
S
halgorithm consistent with general Shannon entropy algorithm;
4th. two steps, generation parameter group two:
Parameter group two considers that, in whole word pattern sum, be divided into 30 groups in order, each group has m
t1=M
t/ 30=240 pattern; To every class symbol sequence, add up respectively in 240 patterns, the number N that disable word pattern occurs
i; To often organize the N of the inside appearance
idivided by total disable word quantity, obtain 30 disable word rate f
1, f
2..., f
30;
Parameter group two is actually and represents disable word pattern and distribute in the one of certain scale;
4th. three steps, generation parameter group three: parameter group three considers that, in whole word pattern sum, be divided into six groups in order, each group has m
t2=M
t/ 6=120 pattern, to every class symbol sequence, adds up the disable word pattern count N ' occurred in 120 patterns respectively
i, then forbid number of words N divided by total
f, obtain six disable word rate f
b1, f
b2..., f
b6;
In fact parameter group three represents the one distribution of disable word pattern on larger scale;
4th. four steps, generation parameter group four: parameter group four is considered in whole dictionary table, the morphology distribution of the disable word pattern of appearance; Every class symbol sequence is divided into two sections, front and back, and every section is then made up of three characters; According to the size position relationship of these three characters, namely every three characters have four kinds of position relationships, and so two sections, front and back, have 16 kinds of patterns,
Inside whole word pattern sum, to every class symbol sequence, add up the disable word pattern quantity N occurred under 16 kinds of patterns respectively
ci, then divided by total disable word pattern quantity, form disable word rate f
ci;
Morphological characterization i.e. four kinds of position relationships of three characters of disable word pattern are shown in that Fig. 3 (a) is to Fig. 3 (d);
The design of above-mentioned four groups of parameters carries out investigating from disable word mode profile three aspects the disable word mode profile, morphology of the integral power characteristic of sequence, different resolution level.But in fact there is some coupling between these characteristic parameters, therefore, need to carry out dimension-reduction treatment to characteristic parameter, the method for process is PCA method.After dimensionality reduction, according to the mapping of proper vector in new dimension space, obtain new proper vector, adopt SVM method to carry out model training, finally obtain training pattern.Utilize this model just can carry out the diagnosis of new samples.
5th step, standardization and normalized are carried out to proper vector after, and carry out PCA dimension-reduction treatment; The vector parameter dimension of certain signal of sign that above-mentioned four parameter group obtain is 54.These parameters, from Global Information, segmentation prohibition word pattern information, disable word morphology 3 aspects of signal, describe the Nonlinear Dynamical Characteristics of system comprehensively.But between these features, more or less there is coupling.This method adopts the method for principal component analysis (PCA) PCA to carry out feature vectors dimensional down.PCA is the universal method of a standard, in PCA method, is more than or equal to 85% for standard, generates the proper vector of new dimension space with total contribution rate.Before carrying out PCA, because parameters dimension is different, carry out standardization, normalized to this 54 dimension parameter, after process, all parameters are all between-1 to 1.
New proper vector after 6th step, generation dimensionality reduction, and standardization and normalized are carried out to new proper vector;
7th step, adopt multi-level SVM recognition methods, obtain the parameter of LibSVM, mainly RBF nuclear parameter g parameter and penalty parameter c parameter recently, and carry out training modeling;
In the training process, for given normal condition, fault 1, fault 2 ... fault N, has altogether N+1 pattern.Then corresponding total m* (N+1) organizes training sample, has n* (N+1) to organize test sample book.This method requires that number of training is respectively no less than 30.
The major parameter that LibSVM carries out fault diagnosis and pattern-recognition is set to:
SVM type selecting: select C-SVM;
Kernel function type: select RBF kernel function;
Degree parameter in kernel function: the default parameters 3 being set to SVM;
Penalty parameter c and nearest RBF nuclear parameter g: the scope that arranges is for [2^ (-10), 2^ (10)];
CV process: CV process is default value 3 folding.The stepping-in amount of c parameter and g parameter is 0.5
Because this method is applicable to the fault diagnosis of small sample, the method for trellis traversal is adopted to obtain optimum parameter c and g;
Cross-verification model n-fold: be set to 3;
According to the penalty parameter c obtained and nearest RBF nuclear parameter g, training sample is organized to m* (N+1) and trains, obtain model.
In rotary machinery fault diagnosis process, the phenomenon of a lot of fault is like comparing class, as rotor unbalance and misaligning; Inner ring fault in rolling bearing and normal condition etc.Especially these faults are at the fault initial stage, and the Nonlinear Dynamical Characteristics showed is not clearly.Therefore, although according to this method, be extracted multiple parameters of disable word pattern, and carried out PCA Dimensionality reduction, find in our experiment, if directly adopt the method for multiclass many points to carry out SVM training and identification, when it is tested test sample book, usually be difficult to reach satisfied recognition result, if discrimination is not higher than 90% etc.Therefore, as a kind of general method, need to adopt Layering memory SVM recognition methods.
8th step, utilize model to test test sample book, judge whether every class accuracy is greater than 90%; If every class accuracy is greater than 90%, judge there is merging class; If do not merge class, model construction is complete, can carry out diagnostic work; If any merging class, then π=X is set, τ=1, M=6, returns second step; Do you if there is accuracy to be less than 90%, judge τ < 6? as τ < 6, then establish τ=τ+1, return second step; As τ=6, then judge to have the classification of accuracy >90%, as not having, then this diagnoses unsuccessfully, cannot modeling, can not diagnose; As there is the classification of accuracy >90% τ=6, then observe when τ=1, the class mutually obscured not reaching accuracy 90% is merged into a class, and arranges M=6, π=π-X+1, τ=1, returns second step; Wherein: π is for identifying object, and N is total fault category number to be sorted; τ is timing_delay estimation parameter; X is the classification number of the class merged.
This method is actually a kind of modeling pattern of recursion by different level, utilizes disable word schema extraction proper vector, utilizes the LibSVM modeling method of Layering memory, finally can realize the modeling under complex object and diagnosis.
Below, the rolling bearing fault data in U.S.'s CaseWesternReserveUniversity electrical engineering laboratory (http://csegroups.case.edu/bearingdatacenter/home) are utilized to carry out checking and the test of said new method.
Inside its Mishap Database, measuring and calculation object: motor Hp1, FE hold, rpm:1797
;data normal_0 (normal condition), inner ring fault (IR007_1.mat), outer ring fault (B007_1.mat) and rolling body fault (OR0073_1.mat).Sample mode totally four kinds.Namely total identification object is π=4, comprises three kinds of malfunctions and a kind of normal condition; Because these sample data amounts are very large, the sample of single type has 120,000 multiple spots.Adopt the first half data to be training sample, and later half is as test sample book.
Concrete diagnostic procedure is as follows:
The first step: training sample and test sample book respectively in 60,000 data points, first three ten thousand as training sample region, rear 30,000 as volumes sample areas; In respective region, stochastic generation starting point, gets continuous 4096 points as sample; Training and testing sample number is often kind of pattern 40;
Second step: adopt M=6, symbolism process is carried out to all test sample books and training sample, namely becomes codomain sequence samples;
3rd step: adopt τ=1, go through length I=M; Codomain sequence samples is formed symbol sebolic addressing set;
4th step: 320 sampling feature vectors symbol sebolic addressing set being generated 54 dimensions; Comprise:
Generate parameter group one: parameter group one is total taboo word rate f respectively
rwith the information entropy S of sequence
h;
Wherein:
S
halgorithm consistent with general Shannon entropy algorithm;
Generate parameter group two: in whole word pattern sum, be divided into 30 groups in order, each group has 24 patterns, to every class symbol sequence, adds up respectively in 24 patterns, the number N that disable word pattern occurs
i; To often organize the N of the inside appearance
idivided by total disable word quantity, obtain 30 disable word rate f
1, f
2..., f
30;
Generate parameter group three: in whole word pattern sum, be divided into six groups in order, each group has 120 patterns, to every class symbol sequence, adds up the disable word pattern count N ' occurred inside 120 patterns respectively
i, then forbid number of words N divided by total
f, obtain six disable word rate f
b1, f
b2..., f
b6; Wherein:
Generate parameter group four: every class symbol sequence is divided into two sections, front and back, and every section is then made up of three characters; According to the size position relationship of these three characters, namely every three characters have four kinds of position relationships, so two sections, front and back, have 16 kinds of patterns, inside whole word pattern sum, to every class symbol sequence, add up the disable word pattern quantity N occurred under 16 kinds of patterns respectively
ci, then divided by total disable word pattern quantity, form disable word rate f
ci;
5th step: standardization and normalized are carried out to these 320 sampling feature vectors;
6th step: adopt PCA method to carry out dimensionality reduction to 320 samples, contribution rate is set to 85%.Now obtaining new dimension is 26; According to mapping relations, generate the sampling feature vectors of 320 26 dimensions;
7th step: according to the method to set up of LibSVM parameter, obtains the RBF nuclear parameter g in the parameter of LibSVM and penalty parameter c, generates training pattern model1 thus;
8th step: test sample book is tested according to model1, test accuracy result is: normal 87%, inner ring 80%, outer ring 91%, rolling body 93%;
Regulation time controling parameters τ=2, until 6, test accuracy table is table 1:
Table 1: the accuracy table of four classes under different time control parameter values
Find from table 1, desirable diagnostic result cannot be reached.The analysis found that, normal and inner ring fault may differentiate is not high, namely the most mutually obscure.
9th step: normal and inner ring fault are merged into a class, the sample after respective random choose 20 training samples and test sample book symbolism serializing.Now, general classification number is 3.Generate 240 54 dimensional feature vectors, standardization normalization, PCA dimensionality reduction.Now dimensionality reduction drops to 30 dimensions, regeneration 30 dimensional feature vector, standardization normalization.Generate c and g parameter.Generate training pattern model2 thus; Test 3 class objects, result is as table 2:
The accuracy table of table 2 three class under different time control parameter values
| τ | Normally+inner ring fault (%) | Outer ring fault (%) | Rolling body fault (%) |
| τ＝1 | 91 | 95 | 94 |
Find from table 2, reach requirement.Success is three classes separately.Now merging class is 2. will distinguish normal and inner ring fault.
9th step: select each 30 of sample after normal and inner ring fault symbolism serializing, generates the proper vector of 54 dimension parameters.Standardization, normalization.PCA dimensionality reduction, this stylish dimension is 33, generates the proper vector under new dimension; Training generates parameter c and g, the generation model model3 of LibSVM needs.Test test sample book, test result is in table 3:
The accuracy table of table 3 two class under different time control parameter values
Diagnosis finds, gets τ=2 and just can reach requirement.
Note: the recognition correct rate of table 3 be actually three classes normal+inner ring recognition correct rate a1, be multiplied by two class identification LibSVM recognition correct rate a2.Namely a1.a2.
Thus, whole diagnostic model can have two, be respectively in training pattern model2 and model model3 in τ=2 time model; To any one new test sample book, first τ=1 is set, judges it is normal+inner ring fault according to model2, or outer ring fault or rolling body fault.If belong to outer ring fault and rolling body fault, directly diagnostic result can be obtained; If normal+inner ring fault, arranges τ=2, again obtains proper vector, bring model3 into and differentiate, just can know normal or inner ring fault.According to this thinking, total diagnostic result accuracy is 95.2%.
In order to verify the noise resisting ability of this method, we respectively to often kind of test sample book add signal to noise ratio (S/N ratio) be the white Gaussian noise of 5,2,1,0.1,0.01 and 0.001 in table 4, and add that pink noise is in table 5.And training sample remains the sample not adding any noise, the overall accuracy of identification is following (doing the mean value after identifying for 10 times)
Table 4:
| Signal to noise ratio (S/N ratio) 5 | 2 | 1 | 0.1 | 0.01 | 0.001 |
| 100％ | 100％ | 100％ | 95.2％ | 93.4％ | 86.1％ |
Table 5:
| Signal to noise ratio (S/N ratio) 5 | 2 | 1 | 0.1 | 0.01 | 0.001 |
| 96.1％ | 95.2％ | 95.1％ | 95.2％ | 93.4％ | 86.1％ |
Between signal to noise ratio (S/N ratio) 5 to 0.1, little on identification impact.After signal to noise ratio (S/N ratio) 0.001, identify that stability declines.
Therefore, this method shows, the disable word pattern inside symbolization dynamics, adopts multiple disable word mode statistical parameter to be proper vector parameter, utilizes Layering memory LibSvm method, carry out multistage fault diagnosis.When strong noise pollution, mechanical fault can be identified.Recognition correct rate is higher.
Above the specific embodiment of the present invention is described, but, the scope being not limited only to embodiment of the present invention's protection.
Claims (1)
1. prohibit the rotary machinery fault diagnosis method of word pattern based on symbolic dynamics, it is characterized in that: comprise the steps:
The first step: set up and need the time-domain signal sample of diagnostic device under normal condition and various malfunction to be respectively no less than 30, the sampling length L of each sample is not less than 4096 points;
Second step: to each sample, carries out symbolism process, becomes codomain sequence samples by time domain sequences; Be divided into M region by each sample equably by its peak-to-peak value scope, M is 6; Carry out region to each point on each sample to choose, the region at the value place namely on sample is the value in codomain sequence;
3rd step: codomain sequence is become symbol sebolic addressing:
The formation of symbol sebolic addressing is by delay parameter τ and go through length I and determine; Using (L-I* τ+1) individual point before the sampled point of sample is got according to sampling instant order from small to large as start point signal, start point signal is added and the starting point time delay τ of signal, 2 τ, 3 τ ... (I-1) point that τ is corresponding forms one group of sequence respectively, and distinguish time delay τ, 2 τ, 3 τ by start point signal in sequence and with start point signal ... (I-1) point that τ is corresponding carries out symbol definition successively, again according to time-domain signal amplitude putting in order from big to small, obtain symbol sebolic addressing set or word set; Wherein:
L: the length of sample of signal;
τ: delay parameter, τ get the natural number of 1 to 6;
4th step: generating feature vector, comprising:
4.1st step: generate parameter group one, is specially and calculates total disable word rate f
rwith the information entropy S of symbol sebolic addressing
h; Wherein:
Total disable word rate:
N
ffor the total disable word quantity occurred;
M
tfor word pattern sum; M
t=M unequal to 720;
The wherein probability that occurs for sampled point of p (π); π ∈ Π (L) is the arrangement of certain word arbitrarily;
4.2nd step: generate parameter group two, be specially:
In whole word pattern sum, be divided into 30 groups in order, each group has 24 patterns, to every class symbol sequence, adds up respectively in 24 patterns, the number N that disable word pattern occurs
i; To often organize the N of the inside appearance
idivided by total disable word quantity, obtain 30 disable word rate f
1, f
2..., f
30;
4.3rd step: generate parameter group three, be specially:
In whole word pattern sum, be divided into six groups in order, each group has 120 patterns, to every class symbol sequence, adds up the disable word pattern count N occurred inside 120 patterns respectively
i', then divided by total disable word quantity N
f, obtain six disable word rate f
b1, f
b2..., f
b6; Wherein:
4.4th step: generate parameter group four, be specially:
Every class symbol sequence is divided into two sections, front and back, and every section is then made up of three characters; According to the size position relationship of these three characters, namely every three characters have four kinds of position relationships, so two sections, front and back, have 16 kinds of patterns, inside whole word pattern sum, to every class symbol sequence, add up the disable word pattern quantity N occurred under 16 kinds of patterns respectively
ci, then divided by total disable word quantity, form disable word rate f
ci;
5th step: after standardization and normalized are carried out to proper vector, then carry out PCA dimension-reduction treatment;
6th step: generate the new proper vector after dimensionality reduction, and standardization and normalized are carried out to new proper vector;
The LibSVM of the 7th step, employing Layering memory diagnoses, and kernel function adopts RBF, utilizes gridding method to obtain RBF nuclear parameter g and the penalty parameter c of LibSVM, and carries out training modeling;
8th step: utilize model to test test sample book, judges whether every class accuracy is greater than 90%; If every class accuracy is greater than 90%, judge whether the merging class of setting; If do not merge class, model construction is complete, can carry out diagnostic work; If any merging class, then π=X is set, τ=1, M=6, returns second step; Do you if there is accuracy to be less than 90%, judge τ < 6? as τ < 6, then establish τ=τ+1, return second step; As τ=6, then judge whether the classification of accuracy >90%, as not having, then this diagnoses unsuccessfully, cannot modeling, can not diagnose; As there is the classification of accuracy >90% τ=6, then observe when τ=1, the class mutually obscured not reaching accuracy 90% is merged into a class, and arranges M=6, π=π-X+1, τ=1, returns second step; Wherein: π is for identifying object, and N is total fault category number to be sorted; X is the classification number of the class merged.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310432071.4A CN103528836B (en) | 2013-09-22 | 2013-09-22 | The rotary machinery fault diagnosis method of word pattern is prohibited based on symbolic dynamics |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310432071.4A CN103528836B (en) | 2013-09-22 | 2013-09-22 | The rotary machinery fault diagnosis method of word pattern is prohibited based on symbolic dynamics |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103528836A CN103528836A (en) | 2014-01-22 |
| CN103528836B true CN103528836B (en) | 2016-04-20 |
Family
ID=49931028
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310432071.4A Expired - Fee Related CN103528836B (en) | 2013-09-22 | 2013-09-22 | The rotary machinery fault diagnosis method of word pattern is prohibited based on symbolic dynamics |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103528836B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104077231B (en) * | 2014-07-16 | 2017-02-08 | 国家电网公司 | Transformer maintenance optimization method based on symbol dynamics and LS-SVM |
| CN104863840A (en) * | 2015-03-16 | 2015-08-26 | 北京化工大学 | Reciprocating compressor intelligent diagnosis method based on EMD-PCA |
| CN107991097A (en) * | 2017-11-16 | 2018-05-04 | 西北工业大学 | A kind of Method for Bearing Fault Diagnosis based on multiple dimensioned symbolic dynamics entropy |
| CN109034028B (en) * | 2018-07-16 | 2021-09-24 | 河北工业大学 | Symbolic and TF-IDF-based mechanical equipment fault feature extraction method |
| CN109088770B (en) * | 2018-08-21 | 2020-03-31 | 西安交通大学 | Electromechanical system interactive network modeling method based on self-adaptive symbol transfer entropy |
| CN112666918B (en) * | 2020-12-01 | 2022-06-14 | 沈阳化工大学 | Method for self-adaptive industrial process fault detection based on online compressed KECA |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101201295A (en) * | 2006-12-13 | 2008-06-18 | 上海海事大学 | Method and device for predicting grey failure of rotating machinery wavelet |
| CN102998118A (en) * | 2012-11-29 | 2013-03-27 | 西安交通大学 | Bearing quantitative diagnosis method based on morphological filtering and complexity measure |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH06273286A (en) * | 1993-03-18 | 1994-09-30 | Hitachi Ltd | Process state diagnosing system |
| EP1374167B1 (en) * | 2001-03-08 | 2010-05-12 | California Institute Of Technology | Real-time spatio-temporal coherence estimation for autonomous mode identification and invariance tracking |
-
2013
- 2013-09-22 CN CN201310432071.4A patent/CN103528836B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101201295A (en) * | 2006-12-13 | 2008-06-18 | 上海海事大学 | Method and device for predicting grey failure of rotating machinery wavelet |
| CN102998118A (en) * | 2012-11-29 | 2013-03-27 | 西安交通大学 | Bearing quantitative diagnosis method based on morphological filtering and complexity measure |
Non-Patent Citations (3)
| Title |
|---|
| 基于符号动力学信息熵与改进神经网络的风机故障诊断研究;王松岭等;《华北电力大学学报》;20130731;第40卷(第4期);第51-58页 * |
| 基于符号化时间序列分析的电动机轴承故障诊断;马东玲等;《微电机》;20081231;第41卷(第10期);第73-74、86页 * |
| 居于符号熵与支持向量机的滚动轴承故障诊断;陈晓平等;《中国机械工程》;20100915;第21卷(第17期);第2079-2082页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103528836A (en) | 2014-01-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103528836B (en) | The rotary machinery fault diagnosis method of word pattern is prohibited based on symbolic dynamics | |
| Qiao et al. | Deep convolutional and LSTM recurrent neural networks for rolling bearing fault diagnosis under strong noises and variable loads | |
| Gao et al. | FEM simulation-based generative adversarial networks to detect bearing faults | |
| Fu et al. | Fault feature selection and diagnosis of rolling bearings based on EEMD and optimized Elman_AdaBoost algorithm | |
| CN104596767B (en) | Method for diagnosing and predicating rolling bearing based on grey support vector machine | |
| Lin et al. | Hyper-spherical distance discrimination: A novel data description method for aero-engine rolling bearing fault detection | |
| Cui et al. | Improved fault size estimation method for rolling element bearings based on concatenation dictionary | |
| Fei et al. | Quantitative diagnosis of rotor vibration fault using process power spectrum entropy and support vector machine method | |
| CN109187021B (en) | Multi-source Wind turbines Method for Bearing Fault Diagnosis based on entropy | |
| CN102156873B (en) | Chaos-based method for detecting and classifying early single-point faults of mechanical component | |
| Yongbo et al. | Review of local mean decomposition and its application in fault diagnosis of rotating machinery | |
| Fei | Fault diagnosis of bearing based on wavelet packet transform-phase space reconstruction-singular value decomposition and SVM classifier | |
| CN104849050A (en) | Rolling bearing fault diagnosis method based on composite multi-scale permutation entropies | |
| CN112633098B (en) | Rotary machine fault diagnosis method, system and storage medium | |
| Wang et al. | A novel bearing fault diagnosis methodology based on SVD and one-dimensional convolutional neural network | |
| Mohanty et al. | Adaptive fault identification of bearing using empirical mode decomposition–principal component analysis‐based average kurtosis technique | |
| Zhang et al. | A multi-module generative adversarial network augmented with adaptive decoupling strategy for intelligent fault diagnosis of machines with small sample | |
| Yu et al. | Machine fault diagnosis based on Gaussian mixture model and its application | |
| Prosvirin et al. | Global and local feature extraction using a convolutional autoencoder and neural networks for diagnosing centrifugal pump mechanical faults | |
| CN113256443B (en) | Nuclear power water pump guide bearing fault detection method, system, equipment and readable storage medium | |
| Sun et al. | Application of a novel improved adaptive CYCBD method in gearbox compound fault diagnosis | |
| Zhao et al. | Fault diagnosis of rolling bearings with noise signal based on modified kernel principal component analysis and DC‐ResNet | |
| Jiang et al. | Application of EMD and 1.5‐dimensional spectrum in fault feature extraction of rolling bearing | |
| CN111380680A (en) | Check valve fault feature extraction method based on improved permutation entropy | |
| Hao et al. | New fusion features convolutional neural network with high generalization ability on rolling bearing fault diagnosis |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20160420 Termination date: 20160922 |