CN108520234A - A kind of blind source extracting method of rail cracks signal based on the constraint of various features amount - Google Patents
A kind of blind source extracting method of rail cracks signal based on the constraint of various features amount Download PDFInfo
- Publication number
- CN108520234A CN108520234A CN201810309626.9A CN201810309626A CN108520234A CN 108520234 A CN108520234 A CN 108520234A CN 201810309626 A CN201810309626 A CN 201810309626A CN 108520234 A CN108520234 A CN 108520234A
- Authority
- CN
- China
- Prior art keywords
- signal
- vector
- characteristic quantity
- various features
- blind source
- 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.)
- Granted
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
- G06F2218/22—Source localisation; Inverse modelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Abstract
A kind of Blind Signal Extraction method based on various features amount, the present invention propose simultaneously to extract fanaticism number using one or more characteristic quantities, to the various constraints of tradition blind source extracting method addition and are improved.The present invention step be:One, the blind source extraction algorithm of tradition is improved using variable or various features value addition constraint.Two, initial rights amount is selected using scheduled each characteristic quantity.Three, it is iterated using initial rights amount and extraction result is calculated.The present invention carries out constraint extraction to a signal simultaneously using various features amount, it allows the process of blind source extraction to be no longer limited to fixed characteristic quantity to be judged, but becoming alterable features amount or various features amount is carried out at the same time judgement, wide usage is stronger, optimizes blind source extracting method.
Description
Technical field
The present invention relates to the methods that field is extracted in blind source, and in particular to a kind of blind source extraction based on the constraint of various features amount
Method.
Background technology
Blind signal processing is broadly divided into blind source separating, and blind source is extracted, blind deconvolution, blind equalization, blind discrimination, and in the present invention
The blind source extraction mentioned is then one kind in blind source processing.In order to effective in the case where source signal is unknown or hybrid mode is unknown
Ground extracts source signal, it is proposed that the method for blind source extraction, early in last century the nineties, this method is just proposed out
Come, such an approach achieves the extractions of signal source number source signal when unknown, and with the gradual development of this method, 2002
The blind source extracting method based on fourth order cumulant is proposed when year.Be in the present invention based on blind source separation method,
And it is further processed on this basis, the method for obtaining blind source extraction.For blind source separating algorithm early in last generation
Discipline mid-term just it has been proposed that only do not attract attention for a long time, until ability in 1991 is proposed again, by development for many years,
In 2004, a kind of method of dependent Component Analysis was suggested, and converted the Research Thinking of blind source separating since then.
Rail crack detection technology is developed in the country such as Sino-America Russia of India, is had so far from 2007 more
A scholar studies in this field, the main acoustic emission detection method of mode of rail cracks non-destructive testing, Pulsed eddy current testing
Method, Magnetic Flux Leakage Inspecting method and optical imagery detection method etc., the present invention involved in detection mode be mainly acoustic emission detection side
Formula.
The present invention constrains algorithmic procedure based on various features amount using blind source separation algorithm as core algorithm, proposes
Based on the blind source extracting method of rail cracks signal of various features amount constraint, using the difference of various characteristic quantities between signal come complete
Freely extracting for pairs of unlike signal, is improved existing blind source extracting method, method is allowed to be not limited to a kind of single spy
Sign amount, but various features amount extracts a kind of signal under the influence of simultaneously.
Invention content
It is an object of the invention to propose a kind of blind source extracting method of the rail cracks signal constrained based on various features amount.
Breach the limitation extracted using single characteristics quantity in traditional blind source extracting method.
The purpose of the present invention is what is be achieved through the following technical solutions:First to the signal that is likely related to and its all kinds of
Characteristic quantity and its corresponding initial weight are calculated, and it is spare to obtain volume of data, further according to preset required feature
Amount and threshold value and the data of all signals are compared, and select the initial weight corresponding to suitable output data, then with
This initial weight is that direction is iterated and isolates required signal, achievees the purpose that blind source extraction.
The flow chart of the present invention is as follows as shown in Figure 1, be divided into three steps:
Step 1:The blind source extraction algorithm of tradition is improved using variable or various features value addition constraint.
1) first against j group source signals, the column vector that initial weight vector is j as one group of length is set, in principle initial power
Vector is chosen for randomly selecting, it is contemplated that the effect of different weights is different, and weights first have in calculating process
By normalized, therefore herein in order to protrude different extraction signal after normalization, one group selected herein to
Amount isWherein α=sgn | | sgn (n-k) | -1 |, k is vectorial serial number herein,
N is each vectorial interior element serial number, n, k=1,2,3 ..., j, this j vector corresponding one group of source signal respectively, and subsequent work
Work is exactly to correspond this j groups weights with every group of signal respectively.
2) signal that may possibly be separated out is calculated, it is S to enable mixed signal, is utilized respectively j kind vectors as just
Beginning weight vector is iterated calculating, stops after iteration h times, calculates separately its characteristic value.Extract iterative process in traditional blind source
The formula utilized is
Wherein W is previous weight vector,For the weight vector after iteration, β=E { WTSg(WT), g is non-linear letter
Number, usually takes g (y)=y3, g (y)=tanh (y) etc..There is prodigious randomness by the result that this formula iteration obtains, and
The requirement for the signal for directionally extracting needs in blind source extracting method cannot be met, therefore should be carried out on this basis further
Derivation so that it is met the requirement of blind source extraction.
G (y)=tanh (y) is taken herein, is exactly initial power for the uncertain maximum reason of this iterative process result
The value of vectorial W is simultaneously not fixed, and the result that the difference of W values is obtained for different in the case of also and differs, therefore repeatedly
Before instead of should to W processing, and the 1 of step 1) in this weight vector has been handled, so leading in this step
If being selected W and being classified.Therefore first sampled point is first taken
Wherein α=sgn | | sgn (n-k) | -1 |, stop after calculating h times, at this time store each characteristic quantity spare.Therefore, herein
It is as follows to the calculation formula that characteristic quantity is preliminary:
Wherein p is sampled point serial number, and k is vectorial serial number, and n is each vectorial interior element serial number, n, k=1,2,3 ..., j, α
=sgn | | sgn (n-k) | -1 |,The each characteristic quantity t that will be obtainedl, l=1,2,3 ... storage it is standby
With.
It when needing to take out related initial weight vector, is illustrated by taking two predetermined characteristic values as an example, if there are two characteristic values
Range is respectively a1<T1<b1,a2<T2<b2, at this point for WkSelection follow the principles
a1<tk1<b1,a2<tk2<b2
At this point, iterative formula becomes
The characteristic value used in the present invention is Ring-down count and kurtosis, and Ring-down count is one fixed threshold value of setting, right
In more than threshold value sampled point counted, for kurtosis then utilize formulaIt calculates, herein y
For arbitrary target signal.If iteration all recalculates the obvious waste of resource of entire signal and operation time each time, therefore is
Simplified calculating introduces constant f, f herein<1, it enables Wherein k is present sample
Points,
m4(y (k+1))=f × m4(y(k))+(1-f)×y(k)4
m2(y (k+1))=f × m2(y(k))+(1-f)×y(k)2
Finally obtained data and expection are compared, obtain result.
Step 2:Initial weight vector is selected using scheduled each characteristic quantity.
Each characteristic quantity of each vector and its corresponding initial gross separation that are obtained in previous step is stored,
It is related to taking out the characteristic quantity that these have been stored when predetermined characteristic amount and is compared, finding qualified characteristic quantity can
To find its corresponding vector, and it is that initial weight vector is iterated calculating with this vector.
Step 3:It is iterated using initial weight vector and extraction result is calculated.
For the initial weight vector obtained in previous step, it is iterated calculating, the formula of calculating is
Formula is used identical with step 1, and wherein p is sampled point serial number, and k is vectorial serial number, and n is member in each vector
Plain serial number, n, k=1,2,3 ..., j, m are preset judgement point, and it acts as reset just when deviation occurs in judging result
Beginning weight vector, α=sgn | | sgn (n-k) | -1 |,WkFor previous weight vector,After iteration
Weight vector.For theoretically, the result for selecting obtained weight vector to be iterated for the first time is exactly final result, but some
When have the case where each characteristic quantity is completely superposed and exist, therefore in order to avoid obtained result is wrong, in calculating process
It can be judged to when fixed node, the mode of judgement is to calculate the characteristic quantity of each separating resulting at this time, if characteristic quantity
It meets the requirements, continues interative computation, converting initial weight vector if characteristic quantity is unsatisfactory for expected require re-starts repeatedly
Generation, such result ensure that avoid when characteristic quantity range is completely superposed and extract the wrong situation of result caused by meeting and go out
It is existing.
The present invention has the following advantages that compared with prior art:
The present invention carries out constraint extraction to a signal simultaneously using various features amount, the process no longer office for allowing blind source to extract
It is limited to a kind of judgement of single characteristic quantity, the present invention is directed to traditional blind source extracting method, right on the basis of blind source separating
Judge and calculating process is improved, it is proposed that a kind of blind source extracting method based on multi-characteristicquantity quantity constraint, for different spies
Sign amount without changing in an iterative process, it is only necessary to add Rule of judgment, using simple, common signal characteristic quantity is equal
It can be added in constraints, wide usage is stronger, optimizes blind source extracting method.
Description of the drawings
Fig. 1 is the flow chart of the present invention
Fig. 2 is the source signal figure of the present invention
Fig. 3 is source signal partial enlarged view of the present invention
Fig. 4 is source signal of the present invention signal graph after mixing at random
Fig. 5 is mixed signal partial enlarged view of the present invention
Fig. 6 is that result figure is extracted in the blind source of the present invention
Fig. 7 is result figure partial enlarged view of the present invention
Specific implementation mode
Illustrate the specific implementation mode of the present invention with reference to embodiment and attached drawing:The data of the verification present invention come from
Collected break signal in matlab data simulations and actual experiment.
Execute step 1:In order to prove the present invention under any circumstance, or it is applicable in most cases, in the present invention
The middle signal used is four groups of signals, respectively three analog signals and one group of actual signal, is using the purpose of analog signal
It is added in actual signal as noise, and final purpose is then that true letter is desirably extracted in one group of mixed signal
Number or noise signal.The process of realization is that this four groups of signals are multiplied with the square formation of a random 4*4, obtained result
For mixed signal, the process that processing separates and extracts echo signal is carried out to this four groups of mixed signals.
For the value of different n and k, four vector values are finally obtained, respectively
W1=[100 11 1] '
W2=[1 100 1 1] '
W3=[1 1 100 1] '
W4=[1 11 100] '
Three noise signals that selection is tested done in the present invention are respectively square-wave signal, sinusoidal signal and white noise letter
Number, the actual signal of selection is collected break signal in actual experiment, and four signals are right as shown in Fig. 2, viewing for convenience
Square wave, sine wave and white noise signal carry out partial enlargement respectively, and gained image is as shown in Figure 3.
It is once tested at random, the random hybrid matrix of the source signal used is
Image after mixing respectively carries out partial enlargement, such as Fig. 5 institutes as shown in figure 4, for convenience of watching per signal all the way
Show.
It can be seen from the figure that the case where can not intuitively finding out actual signal after mixing.
Calculate separately the kurtosis of four source signals first, respectively square-wave signal kurtosis is -2, sine wave signal kurtosis is -
1.7~-1.2, the kurtosis of white noise signal is -1.5~-0.5, and the kurtosis of actual signal is -1.5~9.H=500 is set, if
It is 1.5 to determine Ring-down count threshold value, sets and judges length as m=250 nodes, square wave Ring-down count number is 0 after being detached, just
String wave is 0, and white noise is 10~30, and actual signal is indefinite due to signal distinctive ringing count value.
Since all kinds of characteristic quantities of actual signal are uncertain, so herein for convenience's sake first by each noise point
From extraction, the signal i.e. actual signal finally obtained in this way, so initial weight vector also corresponds to these three noises letter by calculating
Number, by calculating, WpkCorresponding characteristic quantity is respectively Wp1Kurtosis be -1.14016734287159, Ring-down count numerical value is
13, Wp2Kurtosis is -1.48616720964444, and Ring-down count numerical value is 0, Wp3Kurtosis is -2, and Ring-down count numerical value is 0, Wp4It is high and steep
Degree is 0.520844115596123, and Ring-down count numerical value is 0, compares the various features figureofmerit of previous three obtained noise,
Correspond to initial weight vector W1Corresponding white noise signal, initial weight vector W2Corresponding sine wave signal, initial weight vector W3Counterparty
Wave signal.
Execute step 2:The signal extracted at first is set in an experiment as square wave, is secondly white noise, is then sine
Wave finally extracts actual signal again, and selection initial weight vector is respectively W in sequence in this way3,W1,W2,W4, should note herein
Meaning is about W4Selection because having selected corresponding initial power according to the various characteristic values of each signal before this
Vector, and after detaching each time in remaining signal then be not present this noise like, therefore ought artificially add noise (square wave, just
String wave, white noise) detached after when removing, actual signal is theoretically only existed in signal at this time, therefore no matter what is chosen
The finally obtained result of kind initial weight vector all only has actual signal, so choosing W herein for convenience's sake4As initial
Weight vector is further processed.
Execute step 3:After choosing four initial weight vectors, into iterative process, the results are shown in Figure 6 after iteration,
In order to facilitate observation, each noise signal isolated is subjected to partial enlargement respectively, obtains that the results are shown in Figure 7.From figure
In it can be seen that after treatment respectively it is anticipated that sequence extracted square-wave signal, white noise signal, sine wave signal
And actual signal, and the result extracted is very intuitive apparent, and this also demonstrates the present invention and being based on various features amount as a kind of
Blind source extracting method have very strong theoretical and practical meaning in engineering.
Claims (4)
1. a kind of blind source extracting method of rail cracks signal based on the constraint of various features amount, it is characterised in that it includes following step
Suddenly:
Step 1:The blind source extraction algorithm of tradition is improved using variable or various features value addition constraint;
Step 2:Initial weight vector is selected using scheduled each characteristic quantity;
Step 3:It is iterated using initial weight vector and extraction result is calculated.
2. a kind of blind source extracting method of Rail crack detection based on the constraint of various features amount according to claim 1,
It is characterized in that the step one is:
1) first against j group source signals, the column vector that initial weight vector is j as one group of length is set, one group selected herein
Vector isWherein α=sgn | | sgn (n-k) | -1 |, k is vectorial sequence herein
Number, n is each vectorial interior element serial number, n, k=1,2,3 ..., j, this j vector corresponding one group of source signal respectively, and subsequent
Work is exactly to correspond this j groups weights with every group of signal respectively.
2) signal that may possibly be separated out is calculated, it is S to enable mixed signal, is utilized respectively j kind vectors as initial power
Vector is iterated calculating, stops after iteration h times, calculates separately its characteristic value.Traditional blind source extraction iterative process utilizes
Formula be
Wherein W is previous weight vector,For the weight vector after iteration, β=E { WTSg(WT), g is nonlinear function, is led to
Often take g (y)=y3, g (y)=tanh (y) etc., should carry out further deriving on the basis of this makes it meet wanting for blind source extraction
It asks.
G (y)=tanh (y) is taken herein, and mainly W is selected and classified before the iteration.Therefore first is first taken to adopt
Sampling pointWherein α=sgn | | sgn (n-k) | -1 |, stop after calculating h times
Only, each characteristic quantity is stored at this time spare.Therefore, as follows to the calculation formula that characteristic quantity is preliminary herein:
Wherein p is sampled point serial number, and k is vectorial serial number, and n is each vectorial interior element serial number, n, k=1,2,3 ..., j, α=
Sgn | | sgn (n-k) | -1 |,The each characteristic quantity t that will be obtainedl, l=1,2,3 ... storage it is spare.
It when needing to take out related initial weight vector, is illustrated by taking two predetermined characteristic values as an example, if there are two range of characteristic values
Respectively a1<T1<b1,a2<T2<b2, at this point for WkSelection follow the principles
a1<tk1<b1,a2<tk2<b2
At this point, iterative formula becomes
The characteristic value used in the present invention is Ring-down count and kurtosis, and Ring-down count is one fixed threshold value of setting, for super
The threshold value sampled point crossed is counted, and formula is then utilized for kurtosisIt calculates, y is to appoint herein
Meaning echo signal.If iteration all recalculates the obvious waste of resource of entire signal and operation time each time, therefore for letter
Change to calculate and introduces constant f, f herein<1, it enables Wherein k counts for present sample,
m4(y (k+1))=f × m4(y(k))+(1-f)×y(k)4
m2(y (k+1))=f × m2(y(k))+(1-f)×y(k)2
Finally obtained data and expection are compared, obtain result.
3. a kind of Blind Signal Extraction method based on the constraint of various features amount according to claim 1, it is characterised in that institute
The step of stating two be:
Each characteristic quantity of each vector and its corresponding initial gross separation that are obtained in previous step is stored, is being related to
The characteristic quantity that these have been stored is taken out when to predetermined characteristic amount to be compared, finding qualified characteristic quantity can look for
It is that initial weight vector is iterated calculating to its corresponding vector, and with this vector.
4. a kind of Blind Signal Extraction method based on the constraint of various features amount according to claim 1, it is characterised in that institute
The step of stating three be:
For the initial weight vector obtained in previous step, it is iterated calculating, the formula of calculating is
Formula is used identical with step 1, and wherein p is sampled point serial number, and k is vectorial serial number, and n is each vectorial interior element sequence
Number, n, k=1,2,3 ..., j, m are preset judgement point, and it acts as initial power is reset when deviation occurs in judging result
Vector, α=sgn | | sgn (n-k) | -1 |,WkFor previous weight vector,For the power after iteration
Vector.In order to avoid obtained result is wrong, it can be judged to when fixed node in calculating process, the mode of judgement is
The characteristic quantity for calculating each separating resulting at this time, continues interative computation, if characteristic quantity if characteristic quantity is met the requirements
Be unsatisfactory for it is expected require then to convert initial weight vector re-start iteration.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810309626.9A CN108520234B (en) | 2018-04-08 | 2018-04-08 | Steel rail crack signal blind source extraction method based on multiple characteristic quantity constraints |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810309626.9A CN108520234B (en) | 2018-04-08 | 2018-04-08 | Steel rail crack signal blind source extraction method based on multiple characteristic quantity constraints |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108520234A true CN108520234A (en) | 2018-09-11 |
CN108520234B CN108520234B (en) | 2021-04-06 |
Family
ID=63430747
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810309626.9A Active CN108520234B (en) | 2018-04-08 | 2018-04-08 | Steel rail crack signal blind source extraction method based on multiple characteristic quantity constraints |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108520234B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20040260522A1 (en) * | 2003-04-01 | 2004-12-23 | Laurent Albera | Method for the higher-order blind identification of mixtures of sources |
CN102721545A (en) * | 2012-05-25 | 2012-10-10 | 北京交通大学 | Rolling bearing failure diagnostic method based on multi-characteristic parameter |
CN103124245A (en) * | 2012-12-26 | 2013-05-29 | 燕山大学 | Kurtosis-based variable-step-size self-adaptive blind source separation method |
CN104198187A (en) * | 2014-09-04 | 2014-12-10 | 昆明理工大学 | Mechanical vibration fault characteristic time domain blind extraction method |
-
2018
- 2018-04-08 CN CN201810309626.9A patent/CN108520234B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20040260522A1 (en) * | 2003-04-01 | 2004-12-23 | Laurent Albera | Method for the higher-order blind identification of mixtures of sources |
CN102721545A (en) * | 2012-05-25 | 2012-10-10 | 北京交通大学 | Rolling bearing failure diagnostic method based on multi-characteristic parameter |
CN103124245A (en) * | 2012-12-26 | 2013-05-29 | 燕山大学 | Kurtosis-based variable-step-size self-adaptive blind source separation method |
CN104198187A (en) * | 2014-09-04 | 2014-12-10 | 昆明理工大学 | Mechanical vibration fault characteristic time domain blind extraction method |
Also Published As
Publication number | Publication date |
---|---|
CN108520234B (en) | 2021-04-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110826059B (en) | Method and device for defending black box attack facing malicious software image format detection model | |
Cho et al. | Measuring superparticle masses at hadron collider using the transverse mass kink | |
CN108052911A (en) | Multi-modal remote sensing image high-level characteristic integrated classification method based on deep learning | |
CN109934222A (en) | A kind of insulator chain self-destruction recognition methods based on transfer learning | |
CN111123048A (en) | Series fault arc detection device and method based on convolutional neural network | |
CN106326926B (en) | Hyperspectral image target spectrum learning method | |
CN115424145A (en) | Planetary remote sensing image surface terrain change detection method based on deep learning | |
CN109753987A (en) | File identification method and feature extracting method | |
CN115296919B (en) | Method and system for calculating special traffic packet by edge gateway | |
CN115937774A (en) | Security inspection contraband detection method based on feature fusion and semantic interaction | |
Kronmueller et al. | Application of deep neural networks to event type classification in IceCube | |
CN108520234A (en) | A kind of blind source extracting method of rail cracks signal based on the constraint of various features amount | |
Pudney et al. | Automatic parameterization for magnetometer zero offset determination | |
CN106770007A (en) | A kind of characteristic wavelength of near-infrared spectrum system of selection for least square method supporting vector machine model | |
CN116502091A (en) | Network intrusion detection method based on LSTM and attention mechanism | |
CN115565082A (en) | Method, system and device for removing cloud noise of satellite remote sensing image | |
CN115169735A (en) | Electroceramic quality prediction method and system based on sample expansion | |
Luo et al. | Robust fault detection of Markovian jump systems with different system modes | |
Ma et al. | Artificial intelligence model for gravitational wave search based on the waveform envelope | |
CN112348042B (en) | Improved YOLOv 3-based infrared target detection method | |
CN114282585A (en) | Astronomical spectrum data-based outlier celestial body classification method | |
CN113177521A (en) | Intelligent radiation source identification method based on combined twin network | |
CN109286619A (en) | Binary sequence recognition methods and circuit | |
Ravi et al. | Intuitionistic fuzzy automaton for approximate string matching | |
Wang et al. | CFAR-guided EfficientDet with improved loss function for SAR target detection |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |