CN102841131A - Intelligent steel cord conveyer belt defect identification method and intelligent steel cord conveyer belt defect identification system - Google Patents

Abstract

The invention discloses an intelligent steel cord conveyer belt defect identification method and an intelligent steel cord conveyer belt defect identification system. The identification method includes the following steps: (1) electromagnetic loading; (2) defect signal acquisition; (3) feature extraction; (4) training sample obtainment; (5) class priority determination; (6) multi-class model establishment; (7) multi-class model training; (8) real-time signal acquisition and synchronous class: electromagnetic detection units are adopted for real-time detection, detected signals are synchronously inputted into a data processor, features are extracted and then sent into established multi-class models, and the defect class of a detected conveyer belt is automatically outputted. The identification system comprises an electromagnetic loader, a plurality of electromagnetic detection units, the data processor and an upper computer, the data processor can automatically output the defect class of the detected conveyer belt, and the upper computer bidirectionally communicates with the data processor. The design of the invention is reasonable, the invention is easy to operate and convenient to put into practice, moreover, the using effect is good, the practical value is high, the reliability of conveyer belt defect detection is enhanced, and the efficiency of defect identification is increased.

Description

Steel cable core conveying belt defective intelligent identification Method and system

Technical field

The invention belongs to steel cable core conveying belt defective intelligent identification technology field, especially relate to a kind of steel cable core conveying belt defective intelligent identification Method and system.

Background technology

The Steel cord belt conveyor is the main transportation equipment in present most of collieries, generally will bear the transportation of 65% above coal amount, changes in modern times and is bringing into play important role in the Coal Production.Because long-term, the high loaded process of conveying belt and some unexpected factors; Impact or jam like spoil, coal cinder and other hard objects of reprinting from other coal conveying system; Can cause conveying belt impaired; Cause the inner defectives such as wire rope breaking, fracture of wire, fatigue and joint displacement that produce after impaired, reduce the tensile strength of conveying belt greatly.By analysis, the main cause of the horizontal broken belt of colliery steel cable core conveying belt is that conveying band joint damage causes inner wire rope fracture of wire, disconnected rope and fatigue etc. to cause the belt intension reduction with conveyer belt after impaired, and then causes horizontal broken belt.For a long time, colliery steel cable core conveying belt operational process lacks effective checkout equipment and method, accidents such as conveying belt often damages, fracture; Gently then cause the conveying belt back skating; Coal stops up the tunnel, and big of conveying belt is impaired, and is heavy then cause broken belt; System's damage and casualties bring about great losses for national wealth and people's life.

Steel cable core conveying belt is to be a kind of high-tenacity conveying belt of skeleton through explained hereafter such as sulfurations with the wire rope.At present domestic main with ST series steel cable core conveying belt; Defectives such as disconnected rope, fracture of wire, fatigue can appear in wire rope in the use; On-the-spot sulfuration process of conveying belt and level deficiency also can cause vulcanized joint hidden danger simultaneously; If meet shock loads such as start-stop, lump coal, will directly cause the conveying belt fracture.Analyzing colliery steel cable core conveying belt broken belt the root of the accident, is exactly to lack effective, reliable steel cable core conveying belt defects detection means and evaluation criterion.Nowadays, the steel cable core conveying belt in most of collieries still adopts the method for artificial visually examine and periodic replacement management, has potential safety hazard and serious wasting phenomenon at present.And in the actual application, because the steel cable core conveying belt defect kind is more, the signal characteristic more complicated, the flaw indication pattern-recognition of steel cable core conveying belt belongs to many Classification and Identification.The existing research wretched insufficiency of detection system aspect the flaw indication pattern-recognition, and have problems such as real-time difference and reliability are low.

Along with continuous lifting to conveying belt induction system safe reliability attention rate; The research of colliery steel cable core conveying belt broken belt detection technique also deepens continuously; On-line detecting system will substitute hand inspection gradually; The digitizing of detecting instrument, intellectuality will become development trend, and the Intelligent Recognition of defective will replace people's experience, and improve detection and evaluation algorithms based on knowledge gradually; The reliability, the accuracy that detect will improve constantly, thereby improve the security and the economy of colliery steel cable core conveying belt operation comprehensively.

Summary of the invention

Technical matters to be solved by this invention is to above-mentioned deficiency of the prior art the steel cable core conveying belt defective intelligent identification Method that a kind of recognition methods step is simple, realization is convenient and recognition speed is fast, accuracy of identification is high to be provided.

For solving the problems of the technologies described above, the technical scheme that the present invention adopts is: a kind of steel cable core conveying belt defective intelligent identification Method is characterized in that this method may further comprise the steps:

Step 1, electromagnetism load: adopt the electromagnetism charger that steel cable core conveying belt to be detected is carried out electromagnetism and load;

Step 2, flaw indication collection: the remanent magnetism when adopting electromagnetic detecting unit to multiple different defect state in the steel cable core conveying belt to be detected detects respectively; And with institute's detection signal synchronous driving to data processor; The corresponding N group of the different defect states with the N kind of corresponding acquisition defect state detects information; N organizes and includes electromagnetic detecting unit in the said defect state detection information at detected a plurality of detection signals of different sample period, and wherein N is positive integer and N >=3;

A plurality of said detection signals are said electromagnetic detecting unit detected sample sequence in a sampling period, and comprise in this sample sequence that electromagnetic detecting unit is in a plurality of sampled values that a plurality of sampling instant detected;

Step 3, feature extraction: when the pending data processor receives the detection signal that electromagnetic detecting unit transmits; In each detection signal, extract the stack features parameter that to represent and to distinguish this detection signal respectively; And this stack features parameter comprises M characteristic quantity; And M said characteristic quantity numbered, M said characteristic quantity formed a proper vector, wherein M >=2;

Step 4, training sample obtain: organize in the said defect state detection information at the N after feature extraction respectively, randomly draw m detection signal and form training sample set;

Said training sample is concentrated corresponding l the training sample that comprise, m>=2 wherein, l=m * N; L said training sample belongs to N sample class; M training sample when including steel cable core conveying belt to be detected in each sample class and working in same defect state, N sample class are respectively and the corresponding sample class of the different defect states of N kind of steel cable core conveying belt to be detected 1, sample class 2 ... Sample class N; Each training sample in N sample class is all remembered and is made X _Qs, wherein Q is the category label and the Q=1,2 of sample class ... N, s are the sample sequence number and the s=1,2 of an included m training sample in each sample class ... M; X _QsBe the proper vector of s training sample among the sample class k, X _Qs∈ R ^d, wherein d is X _QsVectorial dimension and d=M;

Step 5, classification priority level confirm that its deterministic process is following:

The class center calculation of step 501, sample class: adopt data processor that the class center of any sample class q in N the said sample class is calculated;

And when the class center of sample class q is calculated, according to formula

Calculate each characteristic quantity average of all training samples among the sample class q; Q=1,2 in the formula ... N, p=1,2 ... D, X _Qs(p) be p characteristic quantity of s training sample among the sample class q,

P characteristic quantity average for all training samples among the sample class q;

Step 502, between class distance are calculated: adopt data processor and according to formula

spacing between any sample class h in the individual said sample class of any sample class q and N described in the step 501 is calculated respectively; Wherein

is p characteristic quantity average of all training samples among the sample class q;

is p characteristic quantity average of all training samples among the sample class h, and h=1,2 ... N;

Step 503, a type spacing sum are calculated: adopt data processor and according to the class spacing sum of formula

to any sample class k described in the step 501;

Step 504, repeatedly repeating step 501 is to step 503, the class spacing sum of all sample classes in calculating N said sample class;

Step 505, according to the descending order of class spacing sum of all sample classes that calculate in the step 504, adopt data processor to determine the classification priority level Y of N said sample class, wherein Y=1,2 ... N; Wherein, the highest and its category level of the classification priority level of the sample class that type spacing sum is maximum is 1, and minimum and its category level of the classification priority level of type sample class that the spacing sum is maximum is N;

Step 6, many disaggregated models are set up: many disaggregated models of being set up comprise N-1 two disaggregated models, and N-1 said two disaggregated models are supporting vector machine model; N-1 said two disaggregated models are according to determined classification priority level in the step 405; N said sample class is concentrated by branching away by class to the back earlier from said training sample, and the method for building up of N-1 said two disaggregated models is all identical and all adopt data processor to set up;

For any the two disaggregated model z in N-1 said two disaggregated models, it is following that it sets up process:

Step 601, kernel function are chosen: select the kernel function of RBF as two disaggregated model z for use;

Step 602, classification function are confirmed: after waiting to punish that the nuclear parameter γ of the RBF of selecting for use in parameters C and the step 601 is definite, obtain the classification function of two disaggregated model z, accomplish the process of setting up of two disaggregated model z; Wherein, 0＜C≤1000,0＜γ≤1000;

The two disaggregated model z that set up be priority level to be classified all sample classes of being higher than z from said training sample concentrate branch away after; With the classification priority level is that the sample class of z is concentrated two disaggregated models that branch away in remaining N-z+1 the sample class, wherein z=1,2 from said training sample ... N-1;

Step 603, two disaggregated models classification priority level is set: concentrate the classification priority level z of the sample class that branches away in remaining N-z+1 the sample class based on two disaggregated model z described in the step 602 from said training sample; Classification priority level R to two disaggregated model z sets, and R=z;

Step 604, repeatedly repeating step 601 is to step 603, until the classification function that obtains N-1 said two disaggregated models, just accomplishes the process of setting up of N-1 said two disaggregated models, obtains to set up many disaggregated models of accomplishing.Many disaggregated models that many disaggregated models of being set up branch away for a plurality of sample classes that said training sample is concentrated one by one;

Step 7, many disaggregated model training: many disaggregated model training that l the training sample that training sample described in the step 4 is concentrated is input in the step 6 to be set up;

Step 8, signal in real time collection and synchronous classification: adopt electromagnetic detecting unit that the remanent magnetism in the steel cable core conveying belt to be detected is detected in real time; And institute's detection signal inputed in many disaggregated models that data processor carries out delivering in the step 6 after the feature extraction and set up synchronously the just defective classification of output steel cable core conveying belt to be detected automatically.

Above-mentioned steel cable core conveying belt defective intelligent identification Method is characterized in that: the quantity of electromagnetic detecting unit described in the step 2 is a plurality of, and a plurality of said electromagnetic detecting unit are evenly laid along the Width of steel cable core conveying belt to be detected; And carry out in the step 3 after the feature extraction, said data processor also need call the fusion processing module, and a plurality of said electromagnetic detecting unit institute detection signal is carried out fusion treatment.

Above-mentioned steel cable core conveying belt defective intelligent identification Method is characterized in that: the electromagnetic detecting unit described in the step 2 comprises that level that the remanent magnetism on the horizontal direction in the steel cable core conveying belt to be detected is detected in real time is to electromagnetic detecting unit and/or remanent magnetism on the vertical direction detects in real time in to steel cable core conveying belt to be detected vertically to electromagnetic detecting unit; Said level is to electromagnetic detecting unit and vertically all be laid on the steel cable core conveying belt to be detected to electromagnetic detecting unit; When said electromagnetic detecting unit comprises level to electromagnetic detecting unit with vertically to electromagnetic detecting unit; Said level to electromagnetic detecting unit with vertically to electromagnetic detecting unit synchronously to steel cable core conveying belt to be detected in the remanent magnetism at same position place detect, and said level to electromagnetic detecting unit with vertical identical to the SF of electromagnetic detecting unit;

The N that is obtained in the step 1 organizes said defect state and detects information and should be N group level mutually and detect information and/or the N group vertically detects information to remanent magnetism to remanent magnetism; Wherein, N organize said level to remanent magnetism detection information include said level to electromagnetic detecting unit at detected a plurality of detection signals of different sample period, and N group said vertically to remanent magnetism detection information include said vertically to electromagnetic detecting unit at detected a plurality of detection signals of different sample period;

When carrying out feature extraction in the step 3; N is organized said level detect to remanent magnetism that information and/or N group are said vertically to be detected information to remanent magnetism and carry out feature extraction respectively, corresponding acquisition is organized said level through the N after the feature extraction and is detected information and/or the group of the N after feature extraction is said vertically detects information to remanent magnetism to remanent magnetism;

When obtaining training sample set in the step 4, corresponding acquisition training sample set one and/or training sample set two; Wherein, said training sample set one is randomly drawed the training sample set that m detection signal formed for to organize said level through the N after the feature extraction in remanent magnetism detection information respectively; Said training sample set two is respectively said vertically in remanent magnetism detection information through the N after feature extraction group, randomly draws a training sample set of m detection signal composition; Said training sample set one is identical with the structure of said training sample set two, and the two includes l training sample, and said training sample set one all belongs to N sample class with the individual said training sample of the l in the said training sample set two;

When the priority level of classifying in the step 5 is confirmed; Confirm method according to step 501 to the classification priority level described in the step 505, respectively the classification priority level of a plurality of sample classes in said training sample set one and/or the said training sample set two is confirmed respectively;

Step 6 is carried out many disaggregated models when setting up, the many disaggregated models one of corresponding acquisition and/or many disaggregated models two; Wherein, the many disaggregated models of said many disaggregated models one for a plurality of sample classes in the said training sample set one are branched away one by one, the many disaggregated models of said many disaggregated models two for a plurality of sample classes in the said training sample set two are branched away one by one;

When carrying out many disaggregated model training in the step 7, tackle said many disaggregated models one and/or many disaggregated models two mutually and train respectively; Wherein, when said many disaggregated models one are trained, l training sample in the said training sample set one is input to said many disaggregated models one trains; When said many disaggregated models two are trained, l training sample in the said training sample set two is input to said many disaggregated models two trains;

When carrying out signal in real time collection and synchronous classification in the step 8, the level of tackling is mutually carried out synchronous classification respectively to electromagnetic detecting unit and/or vertically to the real-time institute of electromagnetic detecting unit detection signal; Wherein, To level when the real-time institute of electromagnetic detecting unit detection signal carries out synchronous classification respectively; Said level to electromagnetic detecting unit to steel cable core conveying belt to be detected in remanent magnetism on the horizontal direction detect in real time; And institute's detection signal is carried out inputing in many disaggregated models one of being set up after the feature extraction, export the defective classification of steel cable core conveying belt to be detected afterwards automatically; To vertically when the real-time institute of electromagnetic detecting unit detection signal carries out synchronous classification respectively; Said vertically to electromagnetic detecting unit to steel cable core conveying belt to be detected in remanent magnetism on the vertical direction detect in real time; And institute's detection signal is carried out inputing in many disaggregated models two of being set up after the feature extraction, export the defective classification of steel cable core conveying belt to be detected afterwards automatically.

Above-mentioned steel cable core conveying belt defective intelligent identification Method; It is characterized in that: carry out in the step 3 after the feature extraction; Said data processor also needs all detection signals that said electromagnetic detecting unit detected are carried out noise reduction process respectively, and the denoise processing method of all detection signals that electromagnetic detecting unit detected is all identical;

For any detection signal X (k) that electromagnetic detecting unit detected; Detection signal X (k) is a sample sequence; K=1,2,3 wherein ... N; N is the sampled point quantity among the sample sequence X (k), and this sample sequence X (k) is an one-dimensional signal, and comprises the sampled value of n sampled point among the one-dimensional signal X (k); When one-dimensional signal X (k) was carried out noise reduction process, its noise reduction process process was following:

Step 201, high-frequency signal extract: adopt data processor that the current one-dimensional signal X (k) that receives is carried out wavelet transformation and extracts high-frequency signal, and its leaching process is following:

Step 2011, wavelet decomposition: call the wavelet transformation module, one-dimensional signal X (k) is carried out wavelet decomposition, and obtain each layer approximation coefficient and each layer detail coefficients after the wavelet decomposition; Wherein, said detail coefficients note is made d _{J, k}, j=1,2 ... J, and J is the number of plies of wavelet decomposition, k=1,2,3 ... The sequence number of n sampled point from front to back among n and its expression one-dimensional signal X (k);

Step 2012, detail coefficients threshold process:

According to formula $d_{j,k}^{\&prime;}=\left\{\begin{array}{c}\mathrm{Sign}\left(d_{j,k}\right)\left[\left(\right|d_{j,k}|-\frac{{\mathrm{\&lambda;}}^2}{\left|d_{j,k}\right|\mathrm{Exp}({\left|d_{j,k}\right|}^2-{\mathrm{\&lambda;}}^2)})\right],\left|d_{j,k}\right|\&GreaterEqual;\mathrm{\&lambda;}\\ 0,\left|d_{j,k}\right|<\mathrm{\&lambda;}\end{array}\right.,$ To obtaining each layer detail coefficients d in the step 2011 _{J, k}Carry out threshold process respectively, and obtain each layer detail coefficients d ' after the threshold process _{J, k}In the formula, λ is the threshold value of confirming according to the signal to noise ratio (S/N ratio) of one-dimensional signal X (k);

Step 2013, detail signal reconstruct: call the wavelet inverse transformation module, and according to each layer detail coefficients d ' after the threshold process in the step 2012 _{J, k}, each layer detail signal after the wavelet decomposition carried out reconstruct, and obtains the high-frequency signal N after the reconstruct ₂(k), k=1,2,3 wherein ... N; Said high-frequency signal N ₂(k) comprise n high-frequency signal sampled value in, and N ₂(k)=[n ₂(1), n ₂(2) ..., n ₂(n)];

Step 202, LMS auto adapted filtering are handled: said data processor 2 calls the LMS sef-adapting filter, to signal N ₂(n) carry out exporting signal y (n) after Minimum Mean Square Error calculating and the acquisition filtering, again according to error signal e (n) and according to formula W (n+1)=W (n)+2/ μ (n) e (n) N ₂(n) W (n) is adjusted, make output signal y (n) be tending towards signal N ₁(n), e (n)=d (n)-y (n) wherein; And after said LMS sef-adapting filter processing finishes, the signal e (n) behind the acquisition noise reduction;

Signal N wherein ₂(n) be input signal vector and N ₂(n)=[n ₂(n), n ₂(n-1) ..., n ₂(n-M+1)] ^T, and n ₂(n), n ₂(n-1) ..., n ₂(n-M+1) correspondence is respectively the N of high-frequency signal described in the step 203 ₂(k) nearest M high-frequency signal sampled value in, M is the length of said LMS sef-adapting filter; D (n) is the desired output signal, and d (n) is the one-dimensional signal X (k) described in the step 1, N ₁(n) noise signal for containing among the X (k); Y (n)=N ₂ ^T(n) W (n), W (n) are the coefficient column matrix of said LMS sef-adapting filter under the current state; μ (n) is a step factor, μ (n)=β (1-exp (a|e (n) |)), and a is the constant and a > of control function shape in the formula; 0; β is the constant and the β > of control function span; 0;

After noise reduction process finishes, carry out feature extraction according to the signal e (n) of the feature extracting method described in the step 3 after again to noise reduction.

Above-mentioned steel cable core conveying belt defective intelligent identification Method; It is characterized in that: when carrying out feature extraction in the step 3; The characteristic parameter that is proposed comprises 12 temporal signatures of detection signal; Be M=12, and 12 temporal signatures are that peak-to-peak value, root-mean-square value, average amplitude, variance, root amplitude, kurtosis, ripple are wide respectively, waveform index, peak value index, pulse index, nargin index and kurtosis index; After carrying out feature extraction in the step 3, also need adopt data processor that the characteristic parameter that is extracted is carried out feature reduction.

Above-mentioned steel cable core conveying belt defective intelligent identification Method is characterized in that: calculate the spacing d between any sample class h in sample class q and N the said sample class in the step 502 _QhAfter, the between class distance data of acquisition sample class q; In the step 504 repeatedly repeating step 501 to step 503, obtain the between class distance data and a type spacing sum of N said sample class; Subsequently, said data processor is formed a between class distance symmetric matrix D with the between class distance data of N said sample class _{N * N}, and the between class distance data of each said sample class are positioned between class distance symmetric matrix D _{N * N}With the same line data in the delegation; The class spacing sum of N said sample class is respectively between class distance symmetric matrix D _{N * N}In each line data sum, and between class distance symmetric matrix D _{N * N}In each line data sum form array (Sumd (1), a Sumd (2) ... Sumd (N)).

Correspondingly, when in the step 505 the classification priority level Y of N said sample class being confirmed, its deterministic process is following:

Step 5051, initial parameter are set: the initial value to classification priority level Y and total sample number n' is set the priority level of wherein classifying Y=0, total sample number n '=N respectively;

Step 5052, comparison array (Sumd (1), Sumd (2) ... Sumd (N)) size of current all data in is therefrom selected maximal value Sumd (L), wherein L=1,2 ... N, and be Y+1 with the classification priority level of sample class L, and this moment Y=Y+1, n'=N-1; Simultaneously, with between class distance symmetric matrix D _{N * N}In the L line data all put 0, with array (Sumd (1), Sumd (2) ... Sumd (N)) Sumd (L) in puts 0;

Step 5053, repeating step 5052 repeatedly are till n '=0.

Above-mentioned steel cable core conveying belt defective intelligent identification Method; It is characterized in that: N-1 said two disaggregated models are the fuzzy support vector machine model in the step 6; And carry out training sample in the step 4 when obtaining, include fuzzy membership μ in each training sample in N sample class _Qs, μ wherein _QsBe X _QsFuzzy membership to sample class Q under it.

Above-mentioned steel cable core conveying belt defective intelligent identification Method; It is characterized in that: when in the step 602 punishment parameters C and nuclear parameter γ being confirmed; Through data processor and adopt improved genetic algorithm that selected punishment parameters C and nuclear parameter γ are optimized, its optimizing process is following:

Step 6021, initialization of population: the value of value and nuclear parameter γ that will punish parameters C is as body one by one; And be a population with a plurality of individual collections, all individualities in the said population all carry out forming the initialization population after the binary coding simultaneously; Wherein, the value of value of punishment parameters C and nuclear parameter γ be from the interval (0,1000] in a numerical value randomly drawing;

Each individual fitness value calculation in step 6022, the initialization population: all individual fitness value calculation methods are all identical in the initialization population; A plurality of said individualities in the initialization population, respectively corresponding a plurality of different disaggregated model z;

For any individuality in the said initialization population; Adopt training sample described in the step 5 to concentrate remaining N-Z+1 sample class; Disaggregated model z to corresponding with this individuality trains, and with the classification accuracy of this disaggregated model z as this individual fitness value;

After treating that all individual fitness values all calculate in the said initialization population, the corresponding again kind group mean fitness value that calculates said initialization population;

Step 6023, selection operation:, select fitness value is high in the said initialization population a plurality of individualities as progeny population according to all individual fitness values in the said initialization population that calculates in the step 6022;

Step 6024, interlace operation and mutation operation: the progeny population to choosing carries out interlace operation and mutation operation, obtains the progeny population of a new generation;

Each individual fitness value calculation in step 6025, the progeny population: all individual fitness value calculation methods are all identical in the progeny population; A plurality of said individualities in the progeny population, respectively corresponding a plurality of different disaggregated model z;

For any individuality in the said progeny population; Adopt training sample described in the step 5 to concentrate remaining N-Z+1 sample class; Disaggregated model z to corresponding with this individuality trains, and with the classification accuracy of this disaggregated model z as this individual fitness value;

After treating that all individual fitness values all calculate in the said progeny population, the corresponding again kind group mean fitness value that calculates said progeny population;

Step 6026, selection operation:, select fitness value is high in the said progeny population a plurality of individualities as progeny population according to all individual fitness values in the said progeny population that calculates in the step 6025;

Step 6027, judge whether to satisfy end condition: when evolutionary generation surpassed maximum adaptation degree value individual in predefined maximum evolutionary generation itmax or the progeny population more than or equal to predefined fitness setting value, genetic algorithm stopped also exporting the current the highest individuality of fitness value in the said progeny population that obtains; Otherwise, return step 6024.

Above-mentioned steel cable core conveying belt defective intelligent identification Method is characterized in that: when carrying out interlace operation and mutation operation in the step 6024, according to crossover probability p _cCarry out interlace operation, and according to the variation Probability p _mCarry out mutation operation; Wherein,

$p_c=\left\{\begin{array}{cc}{p}_{c\mathrm{Max}}-\left(\frac{p_{c\mathrm{Max}}-p_{c\mathrm{Min}}}{\mathrm{It}\mathrm{Max}}\right)\&times;\mathrm{Iter}& ,f^{\&prime;}>f_{\mathrm{Avg}}\\ p_{c\mathrm{Max}}& ,f^{\&prime;}\&le;f_{\mathrm{Avg}}\end{array}\right.,$ $p_m=\left\{\begin{array}{cc}{p}_{m\mathrm{Max}}-\left(\frac{p_{m\mathrm{Max}}-p_{m\mathrm{Min}}}{\mathrm{It}\mathrm{Max}}\right)\&times;\mathrm{Iter}& ,f>f_{\mathrm{Avg}}\\ p_{m\mathrm{Max}}& ,f\&le;f_{\mathrm{Avg}}\end{array}\right.;$ In the formula, p _CmaxBe predefined maximum crossover probability, p _CminBe predefined minimum crossover probability, p _MmaxBe predefined maximum variation probability, p _MminBe predefined minimum variation probability, itmax is predefined maximum evolutionary generation, and iter is current evolutionary generation, f _AvgBe the current kind group mean fitness value that carries out the progeny population of interlace operation and mutation operation, f ' is illustrated in bigger fitness value in two individuals that will intersect, the ideal adaptation degree value that f indicates to make a variation.

Simultaneously; The invention also discloses a kind of circuit design rationally, easy-to-connect, use is easy and simple to handle and result of use is good steel cable core conveying belt defective intelligent identifying system; It is characterized in that: comprise that steel cable core conveying belt to be detected is carried out electromagnetism charger, electromagnetism that electromagnetism loads to load a plurality of electromagnetic detecting unit of the remanent magnetism in the steel cable core conveying belt to be detected being detected the back in real time, a plurality of said electromagnetic detecting unit institute detection signal is carried out analyzing and processing and the data processor of output steel cable core conveying belt defective classification to be detected and the host computer that carries out two-way communication with data processor automatically; A plurality of said electromagnetic detecting unit are all joined with signal conditioning circuit; Said signal conditioning circuit and A/D change-over circuit join, and said A/D change-over circuit and data processor join.

The present invention compared with prior art has the following advantages:

1, the defective intelligent identifying system circuit design that is adopted rationally, easy-to-connect, installation is laid convenient and input cost is lower, result of use is good, and the defective intelligent identification Method step that is adopted is simple, recognition speed is fast and accuracy of identification is high.

2, the denoise processing method that is adopted is reasonable in design, realization is convenient and noise reduction process is effective; Wavelet transformation combined with variable step LMS auto adapted filtering carry out noise reduction process; And proposed a kind of new threshold process method, overcome the shortcoming that has constant deviation between the wavelet coefficient after hard-threshold is handled function handles function in the discontinuous shortcoming in threshold value λ place and soft-threshold wavelet coefficient and quantification; Simultaneously, in variable step LMS auto adapted filtering process, step length regulating method is designed again, make that the signal to noise ratio (S/N ratio) behind speed of convergence, steady-state error and the noise reduction of noise reduction process is all significantly improved.Thereby; The denoise processing method that is adopted has not only merged the advantage of wavelet transformation and auto adapted filtering; And through wavelet threshold being handled the improvement of function and the adjustment of LMS auto adapted filtering step-length; Obtained than small echo and the better anti-acoustic capability of auto adapted filtering, and shown that through comparative analysis denoise processing method that the present invention adopts has good result to the noise reduction of the nonstationary noise in the conveying belt flaw indication, has improved signal to noise ratio (S/N ratio) effectively multiple noise reduction algorithm.

3, the multi-sensor information fusion feature extraction and the feature reduction method that are adopted are reasonable in design, have reduced calculated amount significantly, and can effectively extract the essential characteristic vector.

4, the sorting technique that is adopted is reasonable in design, realization is convenient and classification speed is fast, nicety of grading is high, and classification speed is about 0.015 second, and its nicety of grading is up to 91.5%.

5, the sorting technique that is adopted be the basis with two types of sorting algorithms of FSVM (fuzzy support vector machine); And this sorting technique is a kind of class-based fuzzy binary-tree support vector machine multicategory classification method apart from sum; What it utilized a plurality of sample classes confirms the classification priority level apart from sum; Afterwards according to the classification priority level of confirming with a plurality of sample classes by earlier to after branch away one by one, easy and improve nicety of grading and speed significantly.Simultaneously, the fuzzy support vector machine that is adopted can reduce isolated point and the influence of noise to classifying, more progressive assurance classification speed and precision.Thereby the sorting technique that the present invention adopted effectively extends to the occasions of classifying with this one or two sorting technique of SVMs more, and it is convenient to realize, and can adapt to steel cable core conveying belt defective purpose quick, that accurately discern.

6, the actual branch time-like that carries out; The SVMs parameter is bigger to the nicety of grading influence; And in the practical application, the SVMs selection of parameter is difficulty relatively, and the present invention adopts the genetic algorithm after a kind of improve that the parameter of SVMs is optimized; Genetic algorithm after this improvement has the following advantages: 1. avoided the premature convergence problem of standard genetic algorithm effectively, had good global optimization ability; Avoid the sawtooth problem effectively, had good local optimum ability; 3. the genetic operator operation has clear and definite direction, has good constringency performance.Evidence; Genetic algorithm after this improvement can obtain the SVMs parameter quickly; And the nicety of grading of the disaggregated model that obtains is high; Training speed is fast, and the support vector number is few, and has higher nicety of grading and support vector still less for the branch analogy that the contains noise data many sorting algorithms of SVMs commonly used.

7, result of use is good and practical value is high; Can effectively solve this technical barrier that needs to be resolved hurrily of the online detection of existing Steel cord delivery safety in the present production; Realized the robotization of steel cable core conveying belt defective electromagnetic detection; Improved the reliability and the efficient of conveying belt defects detection, significant to guaranteeing the conveying belt safe and reliable operation.

In sum, the present invention is reasonable in design, use is easy and simple to handle, realization is convenient and result of use is good, practical value is high, has increased substantially the reliability and the defect recognition efficient of conveying belt defects detection.

Through accompanying drawing and embodiment, technical scheme of the present invention is done further detailed description below.

Description of drawings

Fig. 1 by the present invention the method flow block diagram of employing intelligent identification Method.

Fig. 2 by the present invention the schematic block circuit diagram of employing intelligent identifying system.

Fig. 3 by the present invention the theory diagram of employing denoise processing method.

Fig. 4 for institute of the present invention employing level to electromagnetic detecting unit and level installation position synoptic diagram to electromagnetic detecting unit.

Description of reference numerals:

1-1-level is to electromagnetic detecting unit; 1-2-vertically to electromagnetic detecting unit;

2-data processor; 3-data-carrier store; 5-signal conditioning circuit;

6-A/D change-over circuit; 7-host computer; 8-run-length encoding device;

9-ST series steel cable core conveying belt.

Embodiment

A kind of steel cable core conveying belt defective intelligent identification Method as shown in Figure 1 may further comprise the steps:

Step 1, electromagnetism load: adopt the electromagnetism charger that steel cable core conveying belt to be detected is carried out electromagnetism and load.

In the present embodiment, the electromagnetism charger that is adopted is weak magnetic load-on module 4.During actual loaded, specifically adopt the weak magnetic charger of TCK-GMS type, also can adopt the weak magnetic charger of other type.

Step 2, flaw indication collection: the remanent magnetism when adopting electromagnetic detecting unit to multiple different defect state in the steel cable core conveying belt to be detected detects respectively; And with institute's detection signal synchronous driving to data processor 2; The corresponding N group of the different defect states with the N kind of corresponding acquisition defect state detects information; N organizes and includes electromagnetic detecting unit in the said defect state detection information at detected a plurality of detection signals of different sample period, and wherein N is positive integer and N >=3.

A plurality of said detection signals are said electromagnetic detecting unit detected sample sequence in a sampling period, and comprise in this sample sequence that electromagnetic detecting unit is in a plurality of sampled values that a plurality of sampling instant detected.In the present embodiment, the quantity of included sampled value is no less than 3 in the said sample sequence.Wherein, multiple different defect states comprise defect states such as wire rope breaking, fracture of wire, fatigue and joint displacement in the step 2.

During actual the use, said electromagnetic detecting unit comprise level that the remanent magnetism on the horizontal direction in the steel cable core conveying belt to be detected is detected in real time the remanent magnetism on the vertical direction detect in real time to electromagnetic detecting unit 1-1 and/or in to steel cable core conveying belt to be detected vertically to electromagnetic detecting unit 1-2.Said level is to electromagnetic detecting unit 1-1 and vertically all be laid on the steel cable core conveying belt to be detected to electromagnetic detecting unit 1-2.When said electromagnetic detecting unit comprises level to electromagnetic detecting unit 1-1 with vertically to electromagnetic detecting unit 1-2; Said level to electromagnetic detecting unit 1-1 with vertically to electromagnetic detecting unit 1-2 synchronously to steel cable core conveying belt to be detected in the remanent magnetism at same position place detect, and said level to electromagnetic detecting unit 1-1 with vertical identical to the SF of electromagnetic detecting unit 1-2.

The N that is obtained in the step 1 organizes said defect state and detects information and should be N group level mutually and detect information and/or the N group vertically detects information to remanent magnetism to remanent magnetism; Wherein, N organize said level to remanent magnetism detection information include said level to electromagnetic detecting unit 1-1 at detected a plurality of detection signals of different sample period, and N group said vertically to remanent magnetism detection information include said vertically to electromagnetic detecting unit 1-2 at detected a plurality of detection signals of different sample period.

In the present embodiment, said electromagnetic detecting unit comprise level that the remanent magnetism on the horizontal direction in the steel cable core conveying belt to be detected is detected in real time the remanent magnetism on the vertical direction detect in real time to electromagnetic detecting unit 1-1 with in to steel cable core conveying belt to be detected vertically to electromagnetic detecting unit 1-2.During actual the laying; Said level is to electromagnetic detecting unit 1-1 and vertically all be laid on the steel cable core conveying belt to be detected to electromagnetic detecting unit 1-2, and said level to electromagnetic detecting unit 1-1 with vertically to electromagnetic detecting unit 1-2 synchronously to steel cable core conveying belt to be detected in the remanent magnetism at same position place detect.Actual when carrying out signals collecting; Said level is identical to the SF of electromagnetic detecting unit 1-2 with vertically to electromagnetic detecting unit 1-1; Correspondingly; The N that is obtained in the step 1 organizes said defect state detection information and is divided into two types, comprises that N group level vertically detects information to remanent magnetism to remanent magnetism detection information and N group.Wherein, N organize said level to remanent magnetism detection information include said level to electromagnetic detecting unit 1-1 at detected a plurality of detection signals of different sample period, and N group said vertically to remanent magnetism detection information include said vertically to electromagnetic detecting unit 1-2 at detected a plurality of detection signals of different sample period.

Correspondingly, said level is to electromagnetic detecting unit 1-1 with saidly vertically detect detected each the said detection signal of information to remanent magnetism and be a sampling detected sample sequence of period.

In the present embodiment, said level is to electromagnetic detecting unit 1-1 and vertically be weak magnetic sensor to electromagnetic detecting unit 1-2, and is specially the TCK weak magnetic sensor.The weak magnetic of TCK detects and is based on " the space magnetic field vector is synthetic " principle; Adopt width, contactless weak magnetic energy gesture induction installation; Applied weak magnetic energy gesture distributional difference information on the ferrimagnet that magnetic carries through extracting, to have accomplished the electromagnetic nondestructive method of location, the various defectives in quantitative and qualitative identification wire rope inside and outside.The TCK weak magnetic sensor that is adopted is a high sensor, and its by release magnetic cell and magnetic the weighing apparatus element form, wherein release magnetic cell certain low-intensity magnetic field B be provided _x, and go out magnetic field B through the Steel cord volume element residue low-intensity magnetic field B after the planning of weak magnetic is related _y, magnetic weighing apparatus element then can be sensitive also exactly with B _yVariable quantity also converts corresponding electric signal, B into _yRelevant with the stray field that Steel cord defective in the steel cable core conveying belt to be detected produces with the Steel cord internal magnetic field, according to B _yVariable quantity, can reflect the defect condition of Steel cord, thereby realize defects detection steel cable core conveying belt.

Actual when carrying out signals collecting, said level to electromagnetic detecting unit 1-1 with vertically all gather according to the SF of setting to electromagnetic detecting unit 1-2, and SF is 1KHz～8KHz.In the present embodiment; Said level is 4KHz to electromagnetic detecting unit 1-1 with vertical SF to electromagnetic detecting unit 1-2; During actual the use; Can said level be adjusted accordingly in 1KHz～8KHz with vertical SF to electromagnetic detecting unit 1-2 to electromagnetic detecting unit 1-1 according to concrete needs.

In the present embodiment, said level is a sample sequence X (i) to electromagnetic detecting unit 1-1 with the detection signal that is vertically detected to electromagnetic detecting unit 1-2, wherein i=1,2,3 ... N, n are the sampled point quantity among the sample sequence X (i).

Step 3, feature extraction: when pending data processor 2 receives the detection signal that electromagnetic detecting unit transmits; In each detection signal, extract the stack features parameter that to represent and to distinguish this detection signal respectively; And this stack features parameter comprises M characteristic quantity; And M said characteristic quantity numbered, M said characteristic quantity formed a proper vector, wherein M >=2.

In the present embodiment; When carrying out feature extraction; The characteristic parameter that is proposed comprises 12 temporal signatures of detection signal; Be M=12,12 temporal signatures are that peak-to-peak value, root-mean-square value, average amplitude, variance, root amplitude, kurtosis, ripple are wide respectively, waveform index, peak value index, pulse index, nargin index and kurtosis index.

Actual when carrying out feature extraction; According to the feature extracting method described in this step; N is organized said level detect to remanent magnetism that information and/or N group are said vertically to be detected information to remanent magnetism and carry out feature extraction respectively, corresponding acquisition is organized said level through the N after the feature extraction and is detected information and/or the group of the N after feature extraction is said vertically detects information to remanent magnetism to remanent magnetism.

In the present embodiment; When reality is carried out feature extraction to each detection signal that said electromagnetic detecting unit detected, level described in the said electromagnetic detecting unit is carried out feature extraction to electromagnetic detecting unit 1-1 respectively with each detection signal that is vertically detected to electromagnetic detecting unit 1-2.

When carrying out feature extraction, the solution procedure of 12 temporal signatures is following for any detection signal (being sample sequence X (i)) that said level is detected to electromagnetic detecting unit 1-1: according to formula X _P-p=max{x _i}-min{x _i, calculate peak-to-peak value X _P-p, max{x in the formula _iBe the maximal value among the sample sequence X (i), min{x _iIt is the minimum value among the sample sequence X (i); According to formula

Calculate root-mean-square value X _RmsAccording to formula

Calculate average amplitude X _AvAccording to formula

Calculate variances sigma _x ², in the formula

According to formula

Calculate root amplitude X _rAccording to formula

Calculate kurtosis β '; When width W was calculated, when the defect state of steel cable core conveying belt to be detected was the joint displacement, width W was the said vertically spacing between electromagnetic detecting unit 1-2 institute detection signal medium wave peak and trough of same sampling instant; When the defect state of steel cable core conveying belt to be detected is other defect state outside the joint displacement; The starting point of width W be same sampling instant said vertically in electromagnetic detecting unit 1-2 institute detection signal amplitude greater than the rising edge of 0.244V, and its terminating point be same sampling instant said vertically in electromagnetic detecting unit 1-2 institute detection signal amplitude less than the negative edge of 0.244V; According to formula

Calculate the waveform index S, X in the formula _RmsBe root-mean-square value, X _AvBe average amplitude; According to formula

Calculate peak value index C, X in the formula _MaxBe the maximal value among the sample sequence X (i), X _RmsBe root-mean-square value; According to formula

Calculate pulse index I, X in the formula _MaxBe the maximal value among the sample sequence X (i), X _AvBe average amplitude; According to formula

Calculate nargin index L, X in the formula _MaxBe the maximal value among the sample sequence X (i), X _rBe the root amplitude; According to formula

Calculate kurtosis index K, β ' is a kurtosis in the formula, X _RmsBe root-mean-square value.

When carrying out feature extraction for said any detection signal that is vertically detected to electromagnetic detecting unit 1-2, except that width W, all the other features extraction methods are identical to the feature extracting method of electromagnetic detecting unit 1-1 institute detection signal with said level.For width W, when the defect state of steel cable core conveying belt to be detected was the joint displacement, width W was the spacing between this detection signal medium wave peak and the trough; When the defect state of steel cable core conveying belt to be detected is other defect state outside the joint displacement; The starting point of width W be in this detection signal amplitude greater than the rising edge of 0.244V, and its terminating point be in this detection signal amplitude less than the negative edge of 0.244V.

In the present embodiment, be divided into N group level and detect information and N group vertically to remanent magnetism detection information to remanent magnetism because N organizes said defect state detection information.Actual when carrying out feature extraction, N is organized said level detect information to remanent magnetism and carry out feature extraction respectively, obtain N after feature extraction and organize said level and detect information to remanent magnetism; Meanwhile, vertically carry out feature extraction respectively to remanent magnetism detection information to the N group is said, the N group that obtains after feature extraction is said vertically to remanent magnetism detection information.

Step 4, training sample obtain: organize in the said defect state detection information at the N after feature extraction respectively, randomly draw m detection signal and form training sample set.

Said training sample is concentrated corresponding l the training sample that comprise, m>=2 wherein, l=m * N; L said training sample belongs to N sample class; M training sample when including steel cable core conveying belt to be detected in each sample class and working in same defect state, N sample class are respectively and the corresponding sample class of the different defect states of N kind of steel cable core conveying belt to be detected 1, sample class 2 ... Sample class N; Each training sample in N sample class is all remembered and is made X _Qs, wherein Q is the category label and the Q=1,2 of sample class ... N, s are the sample sequence number and the s=1,2 of an included m training sample in each sample class ... M; X _QsBe the proper vector of s training sample among the sample class k, X _Qs∈ R ^d, wherein d is X _QsVectorial dimension and d=M.

And when obtaining said training sample set; Because the classification of N sample class is corresponding with the different defect states of N kind of steel cable core conveying belt to be detected respectively, thereby the classification of N sample class is named respectively according to the N kind difference defect state titles of steel cable core conveying belt to be detected.

In the actual mechanical process, organize when randomly drawing m detection signal in the said defect state detection information, adopt data processor 2 to randomly draw at N.

In the present embodiment, m=50.The actual training sample set that carries out is when obtaining, also can be according to concrete needs, the value of m is adjusted accordingly.

When reality is obtained said training sample set, according to the training sample set acquisition methods described in this step, corresponding acquisition training sample set one and/or training sample set two; Wherein, said training sample set one is randomly drawed the training sample set that m detection signal formed for to organize said level through the N after the feature extraction in remanent magnetism detection information respectively; Said training sample set two is respectively said vertically in remanent magnetism detection information through the N after feature extraction group, randomly draws a training sample set of m detection signal composition; Said training sample set one is identical with the structure of said training sample set two, and the two includes l training sample, and said training sample set one all belongs to N sample class with the individual said training sample of the l in the said training sample set two.

In the present embodiment, be divided into N group level and detect information and N group vertically to remanent magnetism detection information to remanent magnetism because N organizes said defect state detection information.

Thereby; When obtaining training sample set; According to the acquisition methods of training sample set described in this step, organizing said level through the N after the feature extraction in remanent magnetism detection information respectively, randomly draw m detection signal and form a training sample set (being training sample set one); Meanwhile, according to the acquisition methods of training sample set described in the step 3, said vertically in remanent magnetism detection information, randomly draw m detection signal composition another training sample set (being training sample set two) respectively through the N after feature extraction group.Said training sample set one is identical with the structure of said training sample set two, and the two includes l training sample, and said training sample set one all belongs to N sample class with the individual said training sample of the l in the said training sample set two.

Step 5, classification priority level confirm that its deterministic process is following:

The class center calculation of step 501, sample class: adopt the class center of any sample class q in 2 couples of N of data processor said sample class to calculate;

And when the class center of sample class q is calculated, according to formula

Calculate each characteristic quantity average of all training samples among the sample class q; Q=1,2 in the formula ... N, p=1,2 ... D, X _Qs(p) be p characteristic quantity of s training sample among the sample class q, P characteristic quantity average for all training samples among the sample class q;

Step 502, between class distance are calculated: adopt data processor 2 and according to formula

spacing between any sample class h in the individual said sample class of any sample class q and N described in the step 501 is calculated respectively; Wherein

is p characteristic quantity average of all training samples among the sample class q;

is p characteristic quantity average of all training samples among the sample class h, and h=1,2 ... N;

Step 503, a type spacing sum are calculated: adopt data processor 2 and according to the class spacing sum of formula

to any sample class k described in the step 501;

Step 504, repeatedly repeating step 501 is to step 503, the class spacing sum of all sample classes in calculating N said sample class;

Step 505, according to the descending order of class spacing sum of all sample classes that calculate in the step 504, adopt data processor 2 to determine the classification priority level Y of N said sample class, wherein Y=1,2 ... N; Wherein, the highest and its category level of the classification priority level of the sample class that type spacing sum is maximum is 1, and minimum and its category level of the classification priority level of type sample class that the spacing sum is maximum is N.

In the present embodiment, calculate the spacing d between any sample class h in sample class q and N the said sample class in the step 502 _QhAfter, the between class distance data of acquisition sample class q; In the step 504 repeatedly repeating step 501 to step 503, obtain the between class distance data and a type spacing sum of N said sample class; Subsequently, said data processor 2 is formed a between class distance symmetric matrix D with the between class distance data of N said sample class _{N * N}, and the between class distance data of each said sample class are positioned between class distance symmetric matrix D _{N * N}With the same line data in the delegation; The class spacing sum of N said sample class is respectively between class distance symmetric matrix D _{N * N}In each line data sum, and between class distance symmetric matrix D _{N * N}In each line data sum form array (Sumd (1), a Sumd (2) ... Sumd (N)).

Correspondingly, when in the step 505 the classification priority level Y of N said sample class being confirmed, its deterministic process is following:

Step 5051, initial parameter are set: the initial value to classification priority level Y and total sample number n' is set the priority level of wherein classifying Y=0, total sample number n '=N respectively;

Step 5052, comparison array (Sumd (1), Sumd (2) ... Sumd (N)) size of current all data in is therefrom selected maximal value Sumd (L), wherein L=1,2 ... N, and be Y+1 with the classification priority level of sample class L, and this moment Y=Y+1, n'=N-1; Simultaneously, with between class distance symmetric matrix D _{N * N}In the L line data all put 0, with array (Sumd (1), Sumd (2) ... Sumd (N)) Sumd (L) in puts 0;

Step 5053, repeating step 5052 repeatedly are till n '=0.

When the actual priority level of classifying is confirmed, confirm method, respectively the classification priority level of a plurality of sample classes in said training sample set one and/or the said training sample set two is confirmed respectively according to step 501 to the classification priority level described in the step 505.

In the present embodiment, confirm method, respectively the classification priority level of a plurality of sample classes in said training sample set one and the said training sample set two is confirmed respectively according to step 501 to the classification priority level described in the step 505.

Step 6, many disaggregated models are set up: many disaggregated models of being set up comprise N-1 two disaggregated models, and N-1 said two disaggregated models are supporting vector machine model; N-1 said two disaggregated models are according to determined classification priority level in the step 405; N said sample class is concentrated by branching away by class to the back earlier from said training sample, and the method for building up of N-1 said two disaggregated models is all identical and all adopt data processor 2 to set up;

For any the two disaggregated model z in N-1 said two disaggregated models, it is following that it sets up process:

Step 601, kernel function are chosen: select the kernel function of RBF as two disaggregated model z for use;

Step 602, classification function are confirmed: after waiting to punish that the nuclear parameter γ of the RBF of selecting for use in parameters C and the step 601 is definite, obtain the classification function of two disaggregated model z, accomplish the process of setting up of two disaggregated model z; Wherein, 0＜C≤1000,0＜γ≤1000;

The two disaggregated model z that set up be priority level to be classified all sample classes of being higher than z from said training sample concentrate branch away after; With the classification priority level is that the sample class of z is concentrated two disaggregated models that branch away in remaining N-z+1 the sample class, wherein z=1,2 from said training sample ... N-1;

Step 603, two disaggregated models classification priority level is set: concentrate the classification priority level z of the sample class that branches away in remaining N-z+1 the sample class based on two disaggregated model z described in the step 602 from said training sample; Classification priority level R to two disaggregated model z sets, and R=z;

Step 604, repeatedly repeating step 601 is to step 603, until the classification function that obtains N-1 said two disaggregated models, just accomplishes the process of setting up of N-1 said two disaggregated models, obtains to set up many disaggregated models of accomplishing.Many disaggregated models that many disaggregated models of being set up branch away for a plurality of sample classes that said training sample is concentrated one by one.

For example, when z=1, two disaggregated models 1 are that 1 sample class is from concentrated two disaggregated models that branch away of said training sample for the priority level of will classify.

The actual branch time-like that carries out, SVMs is divided into two classes in opposition to each other through optimum lineoid with training sample.Yet in the practical application; Because each sample can not incorporate a certain type fully into; And possibly there are noise or isolated point in the sample; Thereby adopt fuzzy support vector machine (Fuzzy Support Vector Machine FSVM) through increasing a fuzzy membership to sample, make the fuzzy membership of isolated point or noise very little, thereby minimizing isolated point and noise are to the influence of optimal classification lineoid.Thereby fuzzy support vector machine is based on the difference of training sample to the classification effect, give different samples in addition different mistakes divide punishment to overcome isolated point and the adverse effect of noise to classifying, and fuzzy membership is definite extremely important.

Therefore, in the present embodiment, N-1 said two disaggregated models are the fuzzy support vector machine model in the step 6, and carry out training sample in the step 4 when obtaining, and include fuzzy membership μ in each training sample in N sample class _Qs, μ wherein _QsBe X _QsFuzzy membership to sample class Q under it.

Actual in fuzzy membership μ _KsWhen confirming, can adopt fuzzy statistics method, illustration method, expertise method or binary contrast ranking method to confirm.Wherein, the basic thought of fuzzy statistics method is to a definite elements A among the domain U (scope of research refers generally to set of real numbers) ₀Whether belong to a variable clear set A in the domain ^*Make clearly and judging.For different experimenters, clear set A ^*Different Boundary can be arranged, but it corresponds to same fuzzy set A.The calculation procedure of fuzzy statistics method is: in each statistics, and A ₀Confirm A ^*Value be variable, do n time the test, its fuzzy statistics can be calculated according to following formula: A ₀The frequency=A that is subordinate to A ₀The number of times of ∈ A/test total degree n.Along with the increase of n, be subordinate to frequency and also can tend to stable, this is stable to be exactly A ₀Degree of membership value to A.This method has reflected the subjection degree in the fuzzy concept more intuitively, but its calculated amount is very big.

The main thought of illustration method is from known limited μ _AValue estimate the subordinate function of fuzzy subset A in the domain.All human like domain U representative, A is " tall people ", and obviously A is a fuzzy subset.In order to confirm μ _A, confirm a height value h earlier, answer someone and whether calculate " tall person " for one in the selected then language true value (i.e. the really degree of a word).Can be divided into " really " " roughly genuine " " likelihood like false " " roughly false " and " vacation " five kinds of situation like the language true value, and represent these language true value with data 1,0.75,0.5,0.25,0 respectively.To N kind differing heights h1, h2, h3 ... Hn does same inquiry, promptly can obtain the discrete representation of the membership function of A.

The expertise method is that the practical experience according to the expert provides the processing formula of fuzzy message or a kind of method that corresponding weight coefficient value is confirmed membership function.In many cases, normally tentatively confirm rough membership function, and then progressively revise and perfect, and the foundation of membership function is checked and adjusted to actual effect just through " study " and practical experience.

Binary contrast ranking method is a kind of method of practical definite membership function, it through between a plurality of things in twos to recently confirming the order under certain characteristic, decide the general shape of these things thus to the subordinate function of this characteristic.Binary contrast ranking method is estimated difference according to contrast, can be divided into relatively method, the contrast method of average, precedence relationship sequencing method and Similarity Priority pairing comparision etc.

In the present embodiment, to μ _QsWhen confirming; Confirm based on the membership function of linear range through data processor 2 and employing; Wherein based on the function of confirming the degree of membership of sample is regarded as distance between the class center of sample and its place sample class in the feature space of the membership function of linear range; Sample is near more from the distance at type center, and degree of membership is big more, otherwise degree of membership is more little; See the 4th phase of volume in August, 2009 disclosed " Lanzhou University of Science & Technology's journal " the 35th for details, Zhang Qiuyu, exhaust " new method that degree of membership is confirmed in the fuzzy support vector machine " literary composition that deliver in ocean etc.

Because the RBF of being selected for use is the RBF kernel function, then nuclear parameter is the parameter in the RBF kernel function $\mathrm{\&gamma;}=\frac{1}{2{\mathrm{\&sigma;}}^2}.$

In the present embodiment, when in the step 602 punishment parameters C and nuclear parameter γ being confirmed, through data processor 2 and adopt improved genetic algorithm that selected punishment parameters C and nuclear parameter γ are optimized, its optimizing process is following:

Step 6021, initialization of population: the value of value and nuclear parameter γ that will punish parameters C is as body one by one; And be a population with a plurality of individual collections, all individualities in the said population all carry out forming the initialization population after the binary coding simultaneously; Wherein, the value of value of punishment parameters C and nuclear parameter γ be from the interval (0,1000] in a numerical value randomly drawing;

Each individual fitness value calculation in step 6022, the initialization population: all individual fitness value calculation methods are all identical in the initialization population; A plurality of said individualities in the initialization population, respectively corresponding a plurality of different disaggregated model z;

For any individuality in the said initialization population; Adopt training sample described in the step 5 to concentrate remaining N-Z+1 sample class; Disaggregated model z to corresponding with this individuality trains, and with the classification accuracy of this disaggregated model z as this individual fitness value;

After treating that all individual fitness values all calculate in the said initialization population, the corresponding again kind group mean fitness value that calculates said initialization population;

Wherein, with each individual corresponding disaggregated model, be punishment parameters C and the corresponding SVMs disaggregated model in the definite back of nuclear parameter γ in each individuality;

Step 6023, selection operation:, select fitness value is high in the said initialization population a plurality of individualities as progeny population according to all individual fitness values in the said initialization population that calculates in the step 6022;

Step 6024, interlace operation and mutation operation: the progeny population to choosing carries out interlace operation and mutation operation, obtains the progeny population of a new generation;

Each individual fitness value calculation in step 6025, the progeny population: all individual fitness value calculation methods are all identical in the progeny population; A plurality of said individualities in the progeny population, respectively corresponding a plurality of different disaggregated model z;

For any individuality in the said progeny population; Adopt training sample described in the step 5 to concentrate remaining N-Z+1 sample class; Disaggregated model z to corresponding with this individuality trains, and with the classification accuracy of this disaggregated model z as this individual fitness value;

After treating that all individual fitness values all calculate in the said progeny population, the corresponding again kind group mean fitness value that calculates said progeny population;

Step 6026, selection operation:, select fitness value is high in the said progeny population a plurality of individualities as progeny population according to all individual fitness values in the said progeny population that calculates in the step 6025;

Step 6027, judge whether to satisfy end condition: when evolutionary generation surpassed maximum adaptation degree value individual in predefined maximum evolutionary generation itmax or the progeny population more than or equal to predefined fitness setting value, genetic algorithm stopped also exporting the current the highest individuality of fitness value in the said progeny population that obtains; Otherwise, return step 6024.

In the present embodiment, predefined maximum crossover probability p _Cmax=0.6, predefined minimum crossover probability p _Cmin=0.9, predefined maximum variation Probability p _Mmax=0.1, predefined minimum variation Probability p _Mmin=0.0001, predefined maximum evolutionary generation itmax=100.

Carry out in the step 6021 before the initialization of population, the initial value of evolutionary generation iter is set at 1.

In the present embodiment, step 6023 is carried out before the selection operation, according to the roulette back-and-forth method, calculates all individual fitness values in the said initialization population.Carry out in the step 6026 before the selection operation,, calculate all individual fitness values in the said progeny population according to the roulette back-and-forth method.

The actual parameter of carrying out is when confirming, when carrying out interlace operation and mutation operation in the step 6024, interlace operation adopts multiple spot to intersect, and mutation operation adopts real-valued variation.In the present embodiment, when carrying out interlace operation and mutation operation in the step 6024,2 intersections are adopted in interlace operation.

In the present embodiment, when carrying out interlace operation and mutation operation in the step 6024, also need current evolutionary generation iter is added up.

In the present embodiment, when carrying out interlace operation and mutation operation in the step 6024, according to crossover probability p _cCarry out interlace operation, and according to the variation Probability p _mCarry out mutation operation; Wherein,

$p_c=\left\{\begin{array}{cc}{p}_{c\mathrm{Max}}-\left(\frac{p_{c\mathrm{Max}}-p_{c\mathrm{Min}}}{\mathrm{It}\mathrm{Max}}\right)\&times;\mathrm{Iter}& ,f^{\&prime;}>f_{\mathrm{Avg}}\\ p_{c\mathrm{Max}}& ,f^{\&prime;}\&le;f_{\mathrm{Avg}}\end{array}\right.,$ $p_m=\left\{\begin{array}{cc}{p}_{m\mathrm{Max}}-\left(\frac{p_{m\mathrm{Max}}-p_{m\mathrm{Min}}}{\mathrm{It}\mathrm{Max}}\right)\&times;\mathrm{Iter}& ,f>f_{\mathrm{Avg}}\\ p_{m\mathrm{Max}}& ,f\&le;f_{\mathrm{Avg}}\end{array}\right.;$ In the formula, p _CmaxBe predefined maximum crossover probability, p _CminBe predefined minimum crossover probability, p _MmaxBe predefined maximum variation probability, p _MminBe predefined minimum variation probability, itmax is predefined maximum evolutionary generation, and iter is current evolutionary generation, f _AvgBe the current kind group mean fitness value that carries out the progeny population of interlace operation and mutation operation, f ' is illustrated in bigger fitness value in two individuals that will intersect, the ideal adaptation degree value that f indicates to make a variation.

That is to say, regulate individual crossover probability p according to fitness value and evolutionary generation _cWith the variation Probability p _mIf ideal adaptation degree value gives bigger crossover probability and variation probability less than kind of a group mean fitness value to it; If individuality is better; Be that its fitness value is greater than kind of a group mean fitness value; Then give this individual corresponding crossover probability and variation probability according to its iterative state and good degree, the more approaching maximum algebraically itmax that sets of iteration algebraically, individual crossover probability is just more little with the variation probability; This kind crossover probability, variation probability control method have stronger ability of searching optimum and more weak local search ability at the evolution initial stage; Along with the carrying out of evolving, the global optimization ability weakens gradually, and the local optimum ability strengthens gradually.This is improved one's methods and helps protecting defect individual, is convenient to obtain globally optimal solution, can prevent " precocity " phenomenon.

When carrying out selection operation in step 6023 and the step 6026, adopt the optimum individual retention strategy.Because in the operation operator of genetic algorithm, select operator can guarantee that the individuality of selecting all is good, but crossover operator and mutation operator have just been introduced new individuality, these two operation operators can not guarantee that the new individuality that produces is good.Therefore adopt the optimum individual retention strategy to obtain optimum individual.Optimum reserved strategy be intersect and make a variation after the new ideal adaptation degree value that relatively produces be to increase or reduce, just keep should be new individual if the new ideal adaptation degree value that produces increases, otherwise keep former individuality.What this improvement strategy can guarantee genetic manipulation effectively and produced all is good new individuality, has confirmed the direction of evolving, and has avoided individual one degradation phenomena during evolution, has strengthened the convergence of algorithm performance.

When actual nuclear parameter γ to the punishment parameters C and the RBF of selecting for use confirms; Also can adopt the grid search method to confirm; And l the training sample that utilizes training sample described in the step 4 to concentrate, and adopt K folding cross validation method that the two disaggregated model z that set up are verified.

Correspondingly, actually carry out many disaggregated models when setting up, according to the described many disaggregated models method for building up of this step, the many disaggregated models one of corresponding acquisition and/or many disaggregated models two; Wherein, the many disaggregated models of said many disaggregated models one for a plurality of sample classes in the said training sample set one are branched away one by one, the many disaggregated models of said many disaggregated models two for a plurality of sample classes in the said training sample set two are branched away one by one.

In the present embodiment; According to the method for building up of the many disaggregated models described in this step, set up many disaggregated models one that a plurality of sample classes in the said training sample set one are branched away one by one and many disaggregated models two that a plurality of sample classes in the said training sample set two are branched away one by one respectively.

Step 7, many disaggregated model training: many disaggregated model training that l the training sample that training sample described in the step 4 is concentrated is input in the step 6 to be set up.

Correspondingly, when carrying out many disaggregated model training in this step, tackle said many disaggregated models one and/or many disaggregated models two mutually and train respectively; Wherein, when said many disaggregated models one are trained, l training sample in the said training sample set one is input to said many disaggregated models one trains; When said many disaggregated models two are trained, l training sample in the said training sample set two is input to said many disaggregated models two trains.

In the present embodiment, l training sample in the said training sample set one is input to many disaggregated models one trains; Meanwhile, l training sample in the said training sample set two being input to many disaggregated models two trains.

Step 8, signal in real time collection and synchronous classification: adopt electromagnetic detecting unit that the remanent magnetism in the steel cable core conveying belt to be detected is detected in real time; And institute's detection signal inputed in many disaggregated models that data processor 2 carries out delivering in the step 6 after the feature extraction and set up synchronously the just defective classification of output steel cable core conveying belt to be detected automatically.

Correspondingly, the defective classification of steel cable core conveying belt to be detected comprises classifications such as wire rope breaking, fracture of wire, fatigue and joint displacement in this step.

In the present embodiment, adopt 2 pairs of institutes of data processor detection signal to carry out before the feature extraction in the step 8, also need institute's detection signal is carried out noise reduction process; And after the feature extraction, also tackle the characteristic parameter that is extracted and carry out feature reduction, and its feature extracting method is identical with the feature extracting method described in the step 3.

In the actual use; When defective appears in steel cable core conveying belt to be detected; Adopt said electromagnetic detecting unit that the remanent magnetism in the steel cable core conveying belt to be detected is detected in real time; And this moment, the institute detection signal was a flaw indication, and in many disaggregated models that said flaw indication is inputed in the step 6 to be set up, exported the defective classification of steel cable core conveying belt to be detected under the current state automatically.And said flaw indication is a sample sequence.

When carrying out signal in real time collection and synchronous classification in this step, the level of tackling is mutually carried out synchronous classification respectively to electromagnetic detecting unit 1-1 and/or vertically to the real-time institute of electromagnetic detecting unit 1-2 detection signal.Wherein, To level when the real-time institute of electromagnetic detecting unit 1-1 detection signal carries out synchronous classification respectively; Said level to electromagnetic detecting unit 1-1 to steel cable core conveying belt to be detected in remanent magnetism on the horizontal direction detect in real time; And institute's detection signal is carried out inputing in many disaggregated models one of being set up after the feature extraction, export the defective classification of steel cable core conveying belt to be detected afterwards automatically; To vertically when the real-time institute of electromagnetic detecting unit 1-2 detection signal carries out synchronous classification respectively; Said vertically to electromagnetic detecting unit 1-2 to steel cable core conveying belt to be detected in remanent magnetism on the vertical direction detect in real time; And institute's detection signal is carried out inputing in many disaggregated models two of being set up after the feature extraction, export the defective classification of steel cable core conveying belt to be detected afterwards automatically.

In the present embodiment; Said level to electromagnetic detecting unit 1-1 to steel cable core conveying belt to be detected in remanent magnetism on the horizontal direction detect in real time; And institute's detection signal carried out inputing in many disaggregated models one of being set up after the feature extraction defective classification of output steel cable core conveying belt to be detected automatically afterwards.Meanwhile; In the said electromagnetic detecting unit vertically to electromagnetic detecting unit 1-2 to steel cable core conveying belt to be detected in remanent magnetism on the vertical direction detect in real time; And institute's detection signal carried out inputing in many disaggregated models two of being set up after the feature extraction defective classification of output steel cable core conveying belt to be detected automatically afterwards.

In the present embodiment; Carry out after the feature extraction in the step 3; Said data processor 2 also needs all detection signals that said electromagnetic detecting unit detected are carried out noise reduction process respectively, and the denoise processing method of all detection signals that said electromagnetic detecting unit detected is all identical.

In the present embodiment; Because said electromagnetic detecting unit comprises that said level is to electromagnetic detecting unit 1-1 with vertically to electromagnetic detecting unit 1-2; Thereby carry out after the feature extraction in the step 3; Said data processor 2 needs all carry out noise reduction process to electromagnetic detecting unit 1-1 with all detection signals that vertically detected to electromagnetic detecting unit 1-2 to said level, and said level is all identical with the denoise processing method of vertical all detection signals that detected to electromagnetic detecting unit 1-2 to electromagnetic detecting unit 1-1.

Actual to said level when electromagnetic detecting unit 1-1 or the detection signal that vertically detected to electromagnetic detecting unit 1-2 carry out noise reduction process, adopt and carry out noise reduction process based on the signal de-noising method of wavelet transformation and variable step LMS auto adapted filtering.For said level to electromagnetic detecting unit 1-1 and any detection signal (being sample sequence X (k)) that is vertically detected to electromagnetic detecting unit 1-2; K=1,2,3 wherein ... N; N is the sampled point quantity among the sample sequence X (k); This sample sequence X (k) is an one-dimensional signal, and comprises the signal sampling value of n sampled point among the one-dimensional signal X (k).When one-dimensional signal X (k) was carried out noise reduction process, its noise reduction process process was following:

Step 201, high-frequency signal extract: adopt 2 couples of current one-dimensional signal X (k) that receive of data processor to carry out wavelet transformation and extract high-frequency signal, and its leaching process is following:

Step 2011, wavelet decomposition: call the wavelet transformation module, one-dimensional signal X (k) is carried out wavelet decomposition, and obtain each layer approximation coefficient and each layer detail coefficients after the wavelet decomposition; Wherein, said detail coefficients note is made d _{J, k}, j=1,2 ... J, and J is the number of plies of wavelet decomposition, k=1,2,3 ... The sequence number of n sampled point from front to back among n and its expression one-dimensional signal X (k).

Step 2012, detail coefficients threshold process:

According to formula $d_{j,k}^{\&prime;}=\left\{\begin{array}{c}\mathrm{Sign}\left(d_{j,k}\right)\left[\left(\right|d_{j,k}|-\frac{{\mathrm{\&lambda;}}^2}{\left|d_{j,k}\right|\mathrm{Exp}({\left|d_{j,k}\right|}^2-{\mathrm{\&lambda;}}^2)})\right],\left|d_{j,k}\right|\&GreaterEqual;\mathrm{\&lambda;}\\ 0,\left|d_{j,k}\right|<\mathrm{\&lambda;}\end{array}\right.,$ To obtaining each layer detail coefficients d in the step 2011 _{J, k}Carry out threshold process respectively, and obtain each layer detail coefficients d ' after the threshold process _{J, k}In the formula, λ is the threshold value of confirming according to the signal to noise ratio (S/N ratio) of one-dimensional signal X (k).

Wherein, sign (x) is a sign function.

Existing nowadays, confirm threshold value after the standard wavelet transformation following two kinds of threshold process methods are arranged: a kind of is to make that absolute value is zero less than the value of the signaling point of threshold value, is called hard-threshold, and the shortcoming of this method is can produce at some point to be interrupted; Another kind of soft-threshold disposal route is point of discontinuity to occur to be retracted to zero in the basic coboundary of hard-threshold, can effectively avoid being interrupted like this, makes the signal smoother that becomes.Though soft-threshold is widely used in reality with hard noise-reduction method, and has obtained noise reduction preferably, all there are some shortcomings in these two kinds of methods self.Wherein, the detail coefficients w ' after hard-threshold is handled _{J, k}Discontinuous at the λ place, utilization w ' _{J, k}Detail signal after the reconstruct can produce some vibrations.And in the soft-threshold processing, w ' _{J, k}Though continuity is better, work as | w _{J, k}| during>=λ, w ' _{J, k}Handle preceding detail coefficients w with hard-threshold _{J, k}Between have constant deviation, influenced the approximation ratio of de-noising signal and actual signal.In practical application, the de-noising signal smoother that soft-threshold is handled, but distorted signals is bigger; And the noise reduction that hard-threshold is handled is undesirable, especially relatively poor for the time varying signal noise reduction.Therefore, threshold process method described in the step 2012 that the present invention adopted can effectively overcome the shortcoming of soft or hard threshold value, layer detail coefficients d ' of each after the threshold process _{J, k}Value between hard-threshold disposal route and soft-threshold disposal route, make d ' _{J, k}More approach the detail coefficients d before hard-threshold is handled _{J, k}, and d ' _{J, k}| d _{J, k}|=λ place is continuous, along with the increase of wavelet coefficient, d ' _{J, k}With d _{J, k}Between absolute value of the bias reduce gradually, work as d _{J, k}When being tending towards infinity with straight line y=d _{J, k}Be asymptotic line, promptly work as | w _{J, k}| when being tending towards infinity, d ' _{J, k}Level off to d _{J, k}Thereby the threshold process method that is adopted in the step 2012 has overcome the shortcoming that has constant deviation between the wavelet coefficient after hard-threshold is handled function handles function in threshold value λ place discontinuous shortcoming and soft-threshold wavelet coefficient and quantification.

Actual when carrying out noise reduction process, J=8 or 9 in the step 2011.In the present embodiment, J=8 in the step 2011 is promptly to carrying out eight layers of scale-of-two wavelet decomposition to one-dimensional signal X (k).

Actual carrying out in the noise reduction process process before in the step 2012 the detail coefficients threshold process being handled, confirmed threshold value λ earlier; Said threshold value λ is default threshold value, penalty threshold value or adopts the Birge-Massart strategy, carries out rigrsure rule, the sqtMolog rule that adopts fixing threshold value form, the heursure rule of taking heuristic threshold value selection mode that adaptive threshold selects or the threshold value of confirming based on the minimaxi rule of minimax principle based on the no partial likelihood estimation principle of Stein.

The threshold value of the λ of threshold value described in the present embodiment for adopting the Birge-Massart strategy to confirm.

Step 2013, detail signal reconstruct: call the wavelet inverse transformation module, and according to each layer detail coefficients d ' after the threshold process in the step 2012 _{J, k}, each layer detail signal after the wavelet decomposition carried out reconstruct, and obtains the high-frequency signal N after the reconstruct ₂(k), k=1,2,3 wherein ... N; Said high-frequency signal N ₂(k) comprise n high-frequency signal sampled value in, and N ₂(k)=[n ₂(1), n ₂(2) ..., n ₂(n)].

Step 202, LMS auto adapted filtering are handled: said data processor 2 calls the LMS sef-adapting filter, to signal N ₂(n) carry out exporting signal y (n) after Minimum Mean Square Error calculating and the acquisition filtering, again according to error signal e (n) and according to formula W (n+1)=W (n)+2 μ (n) e (n) N ₂(n) W (n) is adjusted, make output signal y (n) be tending towards signal N ₁(n), e (n)=d (n)-y (n) wherein; And after said LMS sef-adapting filter processing finishes, the signal e (n) behind the acquisition noise reduction;

Signal N wherein ₂(n) be input signal vector and N ₂(n)=[n ₂(n), n ₂(n-1) ..., n ₂(n-M+1)] ^T, and n ₂(n), n ₂(n-1) ..., n ₂(n-M+1) correspondence is respectively the N of high-frequency signal described in the step 203 ₂(k) nearest M high-frequency signal sampled value in, M is the length of said LMS sef-adapting filter; D (n) is the desired output signal, and d (n) is the one-dimensional signal X (k) described in the step 1, N ₁(n) noise signal for containing among the X (k); Y (n)=N ₂ ^T(n) W (n), W (n) are the coefficient column matrix of said LMS sef-adapting filter under the current state; μ (n) is a step factor, μ (n)=β (1-exp (α | e (n) |)), α is the constant and the α > of control function shape in the formula; 0; β is the constant and the β > of control function span; 0.

Noise reduction process carries out the signal e (n) behind the noise reduction is carried out feature extraction after finishing.

In the present embodiment, after said data processor 2 receives electromagnetic detecting unit institute detection signal, also need in received signal stores synchronized to the data-carrier store 3.Simultaneously, said data processor 2 also needs record is carried out in the pairing sampling instant of each sampled point in institute's detection signal synchronously.

In the present embodiment, when in the step 2011 one-dimensional signal X (k) being carried out wavelet decomposition, adopt the MALLAT algorithm and according to formula $\left\{\begin{array}{c}{c}_{j,k}=\underset{n}{\mathrm{\&Sigma;}}{h}_{k-2n}{c}_{j-1,n}\\ d_{j,k}=\underset{n}{\mathrm{\&Sigma;}}{g}_{k-2n}{d}_{j-1,n}\end{array}\right.,$ To one-dimensional signal X (k) decomposition of dispersing, k=1,2,3 in the formula ... N, j=1,2,3 ... J, n are the sampled point quantity among the sample sequence X (k), and J is the number of plies of wavelet decomposition, c _{J, k}For obtaining each layer approximation coefficient, h after the wavelet decomposition _K-2nBe the impulse response of the low-pass filter relevant with scaling function, and d _{J, k}Be each layer detail signal that obtains after the wavelet decomposition, g _K-2nImpulse response for the BPF. relevant with wavelet function;

When in the step 2013 each layer detail signal after the wavelet decomposition being carried out reconstruct; Carry out reconstruct according to formula

; J=J wherein, J-1 ... 1.

In the present embodiment, in the step 202 0<β<1/> λ<sub >Max</sub>, λ wherein<sub >Max</sub>Be input signal vector N<sub >2</sub>The eigenvalue of maximum of autocorrelation matrix (n).

Actual carrying out in the noise reduction process process, when the α value was big more, the speed of convergence of said LMS sef-adapting filter and tracking velocity were fast more in the step 202, and the steady-state error of said LMS sef-adapting filter is big more; Otherwise, when the α value more hour, the speed of convergence of said LMS sef-adapting filter and tracking velocity are slow more, and the steady-state error of said LMS sef-adapting filter is more little;

When the β value was big more, the speed of convergence of said LMS sef-adapting filter and tracking velocity were fast more; Otherwise, when the β value more hour, the speed of convergence of said LMS sef-adapting filter and tracking velocity are slow more.

During actual treatment, α=10～10000, β=0.0001～0.2.In the present embodiment, α=3000, β=0.0015.When specifically carrying out noise reduction process, can be according to actual needs, the value of α and β is adjusted accordingly in above-mentioned scope.

Correspondingly; In the present embodiment; Adopt said electromagnetic detecting unit that the remanent magnetism in the steel cable core conveying belt to be detected is detected in real time in the step 8; And before many disaggregated models that institute's detection signal is inputed in the step 6 to be set up synchronously, adopt data processor 2 earlier and, the real-time institute of said electromagnetic detecting unit detection signal is carried out noise reduction process according to the described method of step 201 to step 202.Specifically; To level to electromagnetic detecting unit 1-1 and/or vertically before the real-time institute of electromagnetic detecting unit 1-2 detection signal carries out synchronous classification respectively; Earlier according to the described method of step 201 to step 202, respectively to level to electromagnetic detecting unit 1-1 and/or vertically carry out noise reduction process to the real-time institute of electromagnetic detecting unit 1-2 detection signal.

In the present embodiment; With said level before the real-time institute of electromagnetic detecting unit 1-1 detection signal inputs to many disaggregated models one of being set up; Adopt data processor 2 earlier and, said level is carried out noise reduction process to the real-time institute of electromagnetic detecting unit 1-1 detection signal according to the described method of step 201 to step 202; Meanwhile; With said vertical before the real-time institute of electromagnetic detecting unit 1-2 detection signal inputs to many disaggregated models two of being set up; Adopt data processor 2 earlier and, vertically carry out noise reduction process to the real-time institute of electromagnetic detecting unit 1-2 detection signal to said according to the described method of step 201 to step 202.

Because sef-adapting filter (specifically referring to the LMS sef-adapting filter) is a kind of special S filter that can adjust self parameter automatically; If the statistical property of input signal changes; It can follow the tracks of this variation, adjusts parameter automatically, makes performance of filter reach best again.

Variable Step Size LMS Adaptive Filtering Algorithm mainly is to improve step factor μ (n), is improved to the step factor of fixing can change.Reduce step factor μ (n) and can improve convergence of algorithm precision and the stable state imbalance noise that reduces algorithm, but the minimizing of step factor μ (n) causes the tracking velocity of algorithm and speed of convergence to reduce.Therefore, the LMS adaptive filter algorithm of fixed step size is conflicting aspect tracking velocity, speed of convergence and the convergence precision algorithm adjustment step factor being required.Step-length adjustment formula μ (the n)=β that is adopted in the step 202 of the present invention (1-exp (α | e (n) |)); It is when initial convergence phase or unknown system parameter change; Step-length is bigger, thereby has faster tracking velocity and speed of convergence to time-varying system; And after convergence, no matter the much undesired signals of input all keep less adjustment step-length to reach very little stable state imbalance noise.Draw through test; Noise reduction process comparison with fixed step size; After adopting the adjustment of step-length described in the step 202 formula μ (n)=β (1-exp (α | e (n) |)) to carry out noise reduction process, the signal to noise ratio (S/N ratio) behind speed of convergence, steady-state error and the noise reduction is all significantly improved.

To sum up, most critical is how to select appropriate threshold and how to carry out threshold process in the wavelet de-noising process, and it directly influences the noise reduction quality of wavelet transformation to signal.When adopting the described method of step 201 to step 202 to carry out noise reduction process, signal to noise ratio (S/N ratio) is greatly improved.Noise reduction process result through to the threshold process method that adopted in soft-threshold disposal route, hard-threshold disposal route and the step 2012 compares and can know, adopt the soft-threshold disposal route to carry out noise reduction process after, the smoothness of signal is better, but distortion is bigger; And the smoothness of hard-threshold disposal route is relatively poor, but distorted signals is less, and the hard-threshold disposal route that is adopted in the step 2012 not only noise reduction is best, and smoothness is higher, excellent noise reduction effect.

In addition, in the noise reduction process process, the noise reduction process effect was best when the wavelet decomposition number of plies was 8 layers or 9 layers.

In conjunction with Fig. 3, when denoise processing method of the present invention was handled, input signal X (k) comprised useful signal s (n) and noise signal N ₁(n), and it is desired output signal d (n), the high frequency detail signal N after wavelet transformation decomposes X (k) ₂(n) as the input signal of LMS sef-adapting filter, so N ₂(n) and N ₁(n) relevant, but uncorrelated with s (n).Afterwards, the LMS sef-adapting filter adjustment of utilization variable step self parameter is so that its output signal y (n) ≈ N ₁(n), then error e (n) promptly is called the optimum estimate to useful signal s (n).Thereby; The denoise processing method that the present invention adopted utilizes the multiple dimensioned LMS of the being decomposed into sef-adapting filter of wavelet transformation to provide and imports undesired signal preferably, and the advantage of dynamic adjustment step-length has improved filter effect and tracking velocity effectively in LMS sef-adapting filter good adaptive property and the convergence process.

To sum up, for nonstationary noises such as weak magnetic detection signals, be difficult to realize optimal filtering with Wiener filtering or Kalman filtering; And auto adapted filtering can provide filter effect preferably; But because the requirement of LMS algorithm between rate of convergence, following rate and steady-state error of fixed step size is contradiction, thereby in the denoise processing method that the present invention adopted, undesired signal is relevant in the input undesired signal of LMS sef-adapting filter and desired output signal; With useful signal when uncorrelated; It has filter effect preferably, and the method that adopts wavelet transformation to combine with variable step LMS auto adapted filtering realizes the noise reduction process of flaw indication, and the signal to noise ratio (S/N ratio) of signal is high; Root-mean-square error is little, and processing speed is fast.

In the present embodiment, when obtaining training sample in the step 4, also need to organize at N respectively to randomly draw b detection signal composition test sample book collection in the said defect state detection information; Said test sample book is concentrated corresponding F the test sample book that comprise, b >=2 wherein, F=b * N; F said test sample book belongs to N said sample class; To after many disaggregated model training of being set up, also need import F said test sample book in the step 7, the classification accuracy rate of the many disaggregated models of set up is tested.

In the present embodiment, b=150.The actual training sample set that carries out is when obtaining, also can be according to concrete needs, the value of b is adjusted accordingly.

In the present embodiment; When obtaining training sample in the step 4; According to the acquisition methods of training sample set described in this step, organize said level in remanent magnetism detection information at N, randomly draw b detection signal and form a test sample book collection (being test sample book collection one); Meanwhile, according to the acquisition methods of training sample set described in the step 4, said vertically in remanent magnetism detection information, randomly draw b detection signal composition another test sample book collection (being test sample book collection two) respectively through the N after feature extraction group.Said test sample book collection one is identical with the structure of said test sample book collection two, and the two includes F training sample, and said test sample book collection one all belongs to N sample class with the individual said training sample of the l in the said test sample book collection two.

Correspondingly, after in the step 7 many disaggregated models of being set up one and many disaggregated models two being trained, also need import F said test sample book in the said test sample book collection one, the classification accuracy rate of the many disaggregated models one of set up is tested; Meanwhile, also need import F said test sample book in the said test sample book collection two, the classification accuracy rate of the many disaggregated models two of set up is tested.

In the actual use, because cable wire inside cable wire core conveying belt to be detected is more, the magnetic field coverage is wide, needs a plurality of sensors of the same type to detect the conveying belt defective simultaneously, therefore must adopt a plurality of sensor informations to merge and obtain defect information.

In the present embodiment, the quantity of electromagnetic detecting unit described in the step 2 is a plurality of, and a plurality of said electromagnetic detecting unit are evenly laid along the Width of steel cable core conveying belt to be detected.

And carry out in the step 3 after the feature extraction, said data processor 2 also need call the fusion processing module, and a plurality of said electromagnetic detecting unit institute detection signal is carried out fusion treatment.In the present embodiment, when a plurality of said electromagnetic detecting unit institute detection signal is carried out fusion treatment, adopt method of weighted mean, and specifically be characteristic level weighting fusion.

In addition, because the characteristic quantity that is extracted in the step 3 is more, calculated amount is big, thereby need from a plurality of characteristic quantities, extract the principal character amount, promptly characteristic is carried out yojan.That is to say, carry out feature extraction in the step 3 after, also need adopt 2 pairs of characteristic parameters that extracted of data processor to carry out feature reduction.Rough set theory is a kind of good feature reduction method, has in the feature reduction field widely to use, and extracts the principal character of flaw indication, reduces the characteristic quantity of defect recognition, with the speed of effective raising defect recognition.

Actual when carrying out feature reduction, can adopt following several kinds of rough set attribute reduction methods commonly used: Pawlak old attribute reduction algorithms, difference matrix old attribute reduction algorithms, attribute importance degree Algorithm for Reduction, information entropy old attribute reduction algorithms and neighborhood rough set Algorithm for Reduction.After the characteristic parameter that extracts carries out feature reduction in the step 3, not only reduce the data dimension in a large number, and improved nicety of grading.

In the present embodiment, in conjunction with Fig. 4, said level to electromagnetic detecting unit 1-1 with said vertically to the quantity of electromagnetic detecting unit 1-2 be a plurality of and the two quantity identical.A plurality of said levels to electromagnetic detecting unit 1-1 be laid in the perpendicular straight line of the center line of steel cable core conveying belt to be detected on, and a plurality of said vertically to electromagnetic detecting unit 1-2 be laid in the perpendicular straight line of the center line of steel cable core conveying belt to be detected on.

Carry out after the feature extraction in the step 3; Said data processor 2 calls the fusion processing module; A plurality of said levels are carried out fusion treatment to electromagnetic detecting unit 1-1 at same sampling instant institute detection signal, again the detection signal one that after fusion treatment, is obtained is carried out feature extraction subsequently; Meanwhile; Said data processor 2 calls the fusion processing module; Saidly vertically carry out fusion treatment at same sampling instant institute detection signal to a plurality of, again the detection signal two that after fusion treatment, is obtained is carried out feature extraction subsequently to electromagnetic detecting unit 1-2.

A kind of steel cable core conveying belt defective intelligent identifying system as shown in Figure 2; Comprise and carry out electromagnetism charger, electromagnetism that electromagnetism loads and load a plurality of electromagnetic detecting unit of the remanent magnetism in the steel cable core conveying belt to be detected being detected the back in real time, a plurality of said electromagnetic detecting unit institute detection signal is carried out analyzing and processing and the data processor 2 of output steel cable core conveying belt defective classification to be detected and the host computer 7 that carries out two-way communication with data processor 2 automatically to be detected; A plurality of said electromagnetic detecting unit are all joined with signal conditioning circuit 5; Said signal conditioning circuit 5 joins with A/D change-over circuit 6, and said A/D change-over circuit 6 joins with data processor 2.

In the present embodiment; Said electromagnetic detecting unit comprise level that the remanent magnetism on the horizontal direction in the steel cable core conveying belt to be detected is detected in real time the remanent magnetism on the vertical direction detect in real time to electromagnetic detecting unit 1-1 with in to steel cable core conveying belt to be detected vertically to electromagnetic detecting unit 1-2, and said level is to electromagnetic detecting unit 1-1 with vertically all be laid on the steel cable core conveying belt to be detected to electromagnetic detecting unit 1-2.

During actual the use, the level in the said electromagnetic detecting unit to electromagnetic detecting unit 1-1 with vertically to electromagnetic detecting unit 1-2 synchronously to steel cable core conveying belt to be detected in the remanent magnetism at same position place detect.Actual when carrying out signals collecting, said level is identical to the SF of electromagnetic detecting unit 1-2 with vertically to electromagnetic detecting unit 1-1.

In the present embodiment, carry out two-way communication through the tcp/ip communication module between said data processor 2 and the host computer 7.

Simultaneously, steel cable core conveying belt defective intelligent identifying system of the present invention also comprises the run-length encoding device 8 that joins with data processor 2 and the stroke of steel cable core conveying belt to be detected is detected in real time.

In the present embodiment, steel cable core conveying belt to be detected is a ST series steel cable core conveying belt 9, during actual the use, also can carry out Intelligent Recognition to the steel cable core conveying belt defective of other type.

In the present embodiment, the quantity of said electromagnetic detecting unit is a plurality of.

During actual the laying, a plurality of said electromagnetic detecting unit are evenly laid along the Width of ST series steel cable core conveying belt 9.Correspondingly; Said level to electromagnetic detecting unit 1-1 with said vertically to the quantity of electromagnetic detecting unit 1-2 corresponding be a plurality of and the two quantity identical; A plurality of said levels to electromagnetic detecting unit 1-1 be laid in the perpendicular straight line of the center line of steel cable core conveying belt to be detected on; And a plurality of said vertically to electromagnetic detecting unit 1-2 be laid in the perpendicular straight line of the center line of steel cable core conveying belt to be detected on, see Fig. 4 for details.

In the present embodiment, said data processor 2 is an arm processor.

In addition; In the actual use; Electromagnetic detecting unit described in the step 2 also can be merely level that the remanent magnetism on the horizontal direction in the steel cable core conveying belt to be detected is detected in real time to electromagnetic detecting unit 1-1, perhaps be merely to the remanent magnetism on the vertical direction in the steel cable core conveying belt to be detected detect in real time vertically to electromagnetic detecting unit 1-2.During actual the use, to electromagnetic detecting unit 1-1 and vertical defect recognition accuracy rate to electromagnetic detecting unit 1-2 institute detection signal, the selection level is to electromagnetic detecting unit 1-1 or vertically to electromagnetic detecting unit 1-2 according to level.

Like this, the N that is obtained in the step 1 organizes said defect state and detects information and should be N group level mutually and detect information or the N group vertically detects information to remanent magnetism to remanent magnetism.When carrying out feature extraction in the step 3; Corresponding need are organized said level to N and are detected information or the N group is said vertically carries out feature extraction to remanent magnetism detection information to remanent magnetism, and corresponding acquisition is organized said level through the N after the feature extraction and detected information or the group of the N after feature extraction is said vertically detects information to remanent magnetism to remanent magnetism.When obtaining training sample set in the step 4, corresponding acquisition training sample set one or training sample set two.When the priority level of classifying in the step 5 is confirmed; Confirm method according to step 501 to the classification priority level described in the step 505, respectively the classification priority level of a plurality of sample classes in said training sample set one or the said training sample set two is confirmed respectively.Step 6 is carried out many disaggregated models when setting up, the many disaggregated models one of corresponding acquisition or many disaggregated models two.When carrying out many disaggregated model training in the step 7, tackle said many disaggregated models one or many disaggregated models two mutually and train respectively; When carrying out signal in real time collection and synchronous classification in the step 8, reply level mutually is to electromagnetic detecting unit 1-1 or vertically carry out synchronous classification to the real-time institute of electromagnetic detecting unit 1-2 detection signal.

The above; It only is preferred embodiment of the present invention; Be not that the present invention is done any restriction, every technical spirit changes any simple modification, change and the equivalent structure that above embodiment did according to the present invention, all still belongs in the protection domain of technical scheme of the present invention.

Claims (10)

1. steel cable core conveying belt defective intelligent identification Method is characterized in that this method may further comprise the steps:

Step 1, electromagnetism load: adopt the electromagnetism charger that steel cable core conveying belt to be detected is carried out electromagnetism and load;

Step 2, flaw indication collection: the remanent magnetism when adopting electromagnetic detecting unit to multiple different defect state in the steel cable core conveying belt to be detected detects respectively; And with institute's detection signal synchronous driving to data processor (2); The corresponding N group of the different defect states with the N kind of corresponding acquisition defect state detects information; N organizes and includes electromagnetic detecting unit in the said defect state detection information at detected a plurality of detection signals of different sample period, and wherein N is positive integer and N >=3;

A plurality of said detection signals are said electromagnetic detecting unit detected sample sequence in a sampling period, and comprise in this sample sequence that electromagnetic detecting unit is in a plurality of sampled values that a plurality of sampling instant detected;

Step 3, feature extraction: when pending data processor (2) receives the detection signal that electromagnetic detecting unit transmits; In each detection signal, extract the stack features parameter that to represent and to distinguish this detection signal respectively; And this stack features parameter comprises M characteristic quantity; And M said characteristic quantity numbered, M said characteristic quantity formed a proper vector, wherein M >=2;

Step 4, training sample obtain: organize in the said defect state detection information at the N after feature extraction respectively, randomly draw m detection signal and form training sample set;

Said training sample is concentrated corresponding l the training sample that comprise, m>=2 wherein, l=m * N; L said training sample belongs to N sample class; M training sample when including steel cable core conveying belt to be detected in each sample class and working in same defect state, N sample class are respectively and the corresponding sample class of the different defect states of N kind of steel cable core conveying belt to be detected 1, sample class 2 ... Sample class N; Each training sample in N sample class is all remembered and is made X _Qs, wherein Q is the category label and the Q=1,2 of sample class ... N, s are the sample sequence number and the s=1,2 of an included m training sample in each sample class ... M; X _QsBe the proper vector of s training sample among the sample class k, X _Qs∈ R ^d, wherein d is X _QsVectorial dimension and d=M;

Step 5, classification priority level confirm that its deterministic process is following:

The class center calculation of step 501, sample class: adopt data processor (2) that the class center of any sample class q in N the said sample class is calculated;

And when the class center of sample class q is calculated, according to formula

Calculate each characteristic quantity average of all training samples among the sample class q; Q=1,2 in the formula ... N, p=1,2 ... D, X _Qs(p) be p characteristic quantity of s training sample among the sample class q,

P characteristic quantity average for all training samples among the sample class q;

Step 502, between class distance are calculated: adopt data processor (2) and according to formula

spacing between any sample class h in the individual said sample class of any sample class q and N described in the step 501 is calculated respectively; Wherein

is p characteristic quantity average of all training samples among the sample class q;

is p characteristic quantity average of all training samples among the sample class h, and h=1,2 ... N;

Step 503, a type spacing sum are calculated: adopt data processor (2) and according to the class spacing sum of formula

to any sample class k described in the step 501;

Step 504, repeatedly repeating step 501 is to step 503, the class spacing sum of all sample classes in calculating N said sample class;

Step 505, according to the descending order of class spacing sum of all sample classes that calculate in the step 504, adopt data processor (2) to determine the classification priority level Y of N said sample class, wherein Y=1,2 ... N; Wherein, the highest and its category level of the classification priority level of the sample class that type spacing sum is maximum is 1, and minimum and its category level of the classification priority level of type sample class that the spacing sum is maximum is N;

Step 6, many disaggregated models are set up: many disaggregated models of being set up comprise N-1 two disaggregated models, and N-1 said two disaggregated models are supporting vector machine model; N-1 said two disaggregated models are according to determined classification priority level in the step 405; N said sample class is concentrated by branching away by class to the back earlier from said training sample, and the method for building up of N-1 said two disaggregated models is all identical and all adopt data processor (2) to set up;

For any the two disaggregated model z in N-1 said two disaggregated models, it is following that it sets up process:

Step 601, kernel function are chosen: select the kernel function of RBF as two disaggregated model z for use;

Step 602, classification function are confirmed: after waiting to punish that the nuclear parameter γ of the RBF of selecting for use in parameters C and the step 601 is definite, obtain the classification function of two disaggregated model z, accomplish the process of setting up of two disaggregated model z; Wherein, 0＜C≤1000,0＜γ≤1000;

The two disaggregated model z that set up be priority level to be classified all sample classes of being higher than z from said training sample concentrate branch away after; With the classification priority level is that the sample class of z is concentrated two disaggregated models that branch away in remaining N-z+1 the sample class, wherein z=1,2 from said training sample ... N-1;

Step 603, two disaggregated models classification priority level is set: concentrate the classification priority level z of the sample class that branches away in remaining N-z+1 the sample class based on two disaggregated model z described in the step 602 from said training sample; Classification priority level R to two disaggregated model z sets, and R=z;

Step 604, repeatedly repeating step 601 is to step 603, until the classification function that obtains N-1 said two disaggregated models, just accomplishes the process of setting up of N-1 said two disaggregated models, obtains to set up many disaggregated models of accomplishing; Many disaggregated models that many disaggregated models of being set up branch away for a plurality of sample classes that said training sample is concentrated one by one;

Step 7, many disaggregated model training: many disaggregated model training that l the training sample that training sample described in the step 4 is concentrated is input in the step 6 to be set up;

Step 8, signal in real time collection and synchronous classification: adopt electromagnetic detecting unit that the remanent magnetism in the steel cable core conveying belt to be detected is detected in real time; And institute's detection signal inputed in many disaggregated models that data processor (2) carries out delivering in the step 6 after the feature extraction and set up synchronously the just defective classification of output steel cable core conveying belt to be detected automatically.

2. according to the described steel cable core conveying belt defective of claim 1 intelligent identification Method; It is characterized in that: the quantity of electromagnetic detecting unit described in the step 2 is a plurality of, and a plurality of said electromagnetic detecting unit are evenly laid along the Width of steel cable core conveying belt to be detected; And carry out in the step 3 after the feature extraction, said data processor (2) also need call the fusion processing module, and a plurality of said electromagnetic detecting unit institute detection signal is carried out fusion treatment.

3. according to claim 1 or 2 described steel cable core conveying belt defective intelligent identification Methods, it is characterized in that: the electromagnetic detecting unit described in the step 2 comprise level that the remanent magnetism on the horizontal direction in the steel cable core conveying belt to be detected is detected in real time the remanent magnetism on the vertical direction detect in real time to electromagnetic detecting unit (1-1) and/or in to steel cable core conveying belt to be detected vertically to electromagnetic detecting unit (1-2); Said level is to electromagnetic detecting unit (1-1) and vertically all be laid on the steel cable core conveying belt to be detected to electromagnetic detecting unit (1-2); When said electromagnetic detecting unit comprises that level is to electromagnetic detecting unit (1-1) with vertically when electromagnetic detecting unit (1-2); Said level to electromagnetic detecting unit (1-1) and vertically to electromagnetic detecting unit (1-2) synchronously to steel cable core conveying belt to be detected in the remanent magnetism at same position place detect, and said level to electromagnetic detecting unit (1-1) with vertically the SF to electromagnetic detecting unit (1-2) is identical;

The N that is obtained in the step 1 organizes said defect state and detects information and should be N group level mutually and detect information and/or the N group vertically detects information to remanent magnetism to remanent magnetism; Wherein, N organize said level to remanent magnetism detection information include said level to electromagnetic detecting unit (1-1) at detected a plurality of detection signals of different sample period, and N group said vertically to remanent magnetism detection information include said vertically to electromagnetic detecting unit (1-2) at detected a plurality of detection signals of different sample period;

When carrying out feature extraction in the step 3; N is organized said level detect to remanent magnetism that information and/or N group are said vertically to be detected information to remanent magnetism and carry out feature extraction respectively, corresponding acquisition is organized said level through the N after the feature extraction and is detected information and/or the group of the N after feature extraction is said vertically detects information to remanent magnetism to remanent magnetism;

When obtaining training sample set in the step 4, corresponding acquisition training sample set one and/or training sample set two; Wherein, said training sample set one is randomly drawed the training sample set that m detection signal formed for to organize said level through the N after the feature extraction in remanent magnetism detection information respectively; Said training sample set two is respectively said vertically in remanent magnetism detection information through the N after feature extraction group, randomly draws a training sample set of m detection signal composition; Said training sample set one is identical with the structure of said training sample set two, and the two includes l training sample, and said training sample set one all belongs to N sample class with the individual said training sample of the l in the said training sample set two;

When the priority level of classifying in the step 5 is confirmed; Confirm method according to step 501 to the classification priority level described in the step 505, respectively the classification priority level of a plurality of sample classes in said training sample set one and/or the said training sample set two is confirmed respectively;

Step 6 is carried out many disaggregated models when setting up, the many disaggregated models one of corresponding acquisition and/or many disaggregated models two; Wherein, the many disaggregated models of said many disaggregated models one for a plurality of sample classes in the said training sample set one are branched away one by one, the many disaggregated models of said many disaggregated models two for a plurality of sample classes in the said training sample set two are branched away one by one;

When carrying out many disaggregated model training in the step 7, tackle said many disaggregated models one and/or many disaggregated models two mutually and train respectively; Wherein, when said many disaggregated models one are trained, l training sample in the said training sample set one is input to said many disaggregated models one trains; When said many disaggregated models two are trained, l training sample in the said training sample set two is input to said many disaggregated models two trains;

When carrying out signal in real time collection and synchronous classification in the step 8, the level of tackling is mutually carried out synchronous classification respectively to electromagnetic detecting unit (1-1) and/or vertically to the real-time institute of electromagnetic detecting unit (1-2) detection signal; Wherein, To level to electromagnetic detecting unit (1-1) when in real time institute's detection signal carries out synchronous classification respectively; Said level to electromagnetic detecting unit (1-1) to steel cable core conveying belt to be detected in remanent magnetism on the horizontal direction detect in real time; And institute's detection signal is carried out inputing in many disaggregated models one of being set up after the feature extraction, export the defective classification of steel cable core conveying belt to be detected afterwards automatically; To vertically when the real-time institute of electromagnetic detecting unit (1-2) detection signal carries out synchronous classification respectively; Said vertically to electromagnetic detecting unit (1-2) to steel cable core conveying belt to be detected in remanent magnetism on the vertical direction detect in real time; And institute's detection signal is carried out inputing in many disaggregated models two of being set up after the feature extraction, export the defective classification of steel cable core conveying belt to be detected afterwards automatically.

4. according to claim 1 or 2 described steel cable core conveying belt defective intelligent identification Methods; It is characterized in that: carry out in the step 3 after the feature extraction; Said data processor (2) also needs all detection signals that said electromagnetic detecting unit detected are carried out noise reduction process respectively, and the denoise processing method of all detection signals that electromagnetic detecting unit detected is all identical;

For any detection signal X (k) that electromagnetic detecting unit detected; Detection signal X (k) is a sample sequence; K=1,2,3 wherein ... N; N is the sampled point quantity among the sample sequence X (k), and this sample sequence X (k) is an one-dimensional signal, and comprises the sampled value of n sampled point among the one-dimensional signal X (k); When one-dimensional signal X (k) was carried out noise reduction process, its noise reduction process process was following:

Step 201, high-frequency signal extract: adopt data processor (2) that the current one-dimensional signal X (k) that receives is carried out wavelet transformation and extracts high-frequency signal, and its leaching process is following:

Step 2011, wavelet decomposition: call the wavelet transformation module, one-dimensional signal X (k) is carried out wavelet decomposition, and obtain each layer approximation coefficient and each layer detail coefficients after the wavelet decomposition; Wherein, said detail coefficients note is made d _{J, k}, j=1,2 ... J, and J is the number of plies of wavelet decomposition, k=1,2,3 ... The sequence number of n sampled point from front to back among n and its expression one-dimensional signal X (k);

Step 2012, detail coefficients threshold process:

According to formula $d_{j,k}^{\&prime;}=\left\{\begin{array}{c}\mathrm{Sign}\left(d_{j,k}\right)\left[\left(\right|d_{j,k}|-\frac{{\mathrm{\&lambda;}}^2}{\left|d_{j,k}\right|\mathrm{Exp}({\left|d_{j,k}\right|}^2-{\mathrm{\&lambda;}}^2)})\right],\left|d_{j,k}\right|\&GreaterEqual;\mathrm{\&lambda;}\\ 0,\left|d_{j,k}\right|<\mathrm{\&lambda;}\end{array}\right.,$ To obtaining each layer detail coefficients d in the step 2011 _{J, k}Carry out threshold process respectively, and obtain each layer detail coefficients d ' after the threshold process _{J, k}In the formula, λ is the threshold value of confirming according to the signal to noise ratio (S/N ratio) of one-dimensional signal X (k);

Step 2013, detail signal reconstruct: call the wavelet inverse transformation module, and according to each layer detail coefficients d ' after the threshold process in the step 2012 _{J, k}, each layer detail signal after the wavelet decomposition carried out reconstruct, and obtains the high-frequency signal N after the reconstruct ₂(k), k=1,2,3 wherein ... N; Said high-frequency signal N ₂(k) comprise n high-frequency signal sampled value in, and N ₂(k)=[n ₂(1), n ₂(2) ..., n ₂(n)];

Step 202, LMS auto adapted filtering are handled: said data processor (2) calls the LMS sef-adapting filter, to signal N ₂(n) carry out exporting signal y (n) after Minimum Mean Square Error calculating and the acquisition filtering, again according to error signal e (n) and according to formula W (n+1)=W (n)+2 μ (n) e (n) N ₂(n) W (n) is adjusted, make output signal y (n) be tending towards signal N ₁(n), e (n)=d (n)-y (n) wherein; And after said LMS sef-adapting filter processing finishes, the signal e (n) behind the acquisition noise reduction;

Signal N wherein ₂(n) be input signal vector and N ₂(n)=[n ₂(n), n ₂(n-1) ..., n ₂(n-M+1)] ^T, and n ₂(n), n ₂(n-1) ..., n ₂(n-M+1) correspondence is respectively the N of high-frequency signal described in the step 203 ₂(k) nearest M high-frequency signal sampled value in, M is the length of said LMS sef-adapting filter; D (n) is the desired output signal, and d (n) is the one-dimensional signal X (k) described in the step 1, N ₁(n) noise signal for containing among the X (k); Y (n)=N ₂ ^T(n) W (n), W (n) are the coefficient column matrix of said LMS sef-adapting filter under the current state; μ (n) is a step factor, μ (n)=β (1-exp (α | e (n) |)), α is the constant and the α > of control function shape in the formula; 0; β is the constant and the β > of control function span; 0;

After noise reduction process finishes, carry out feature extraction according to the signal e (n) of the feature extracting method described in the step 3 after again to noise reduction.

5. according to claim 1 or 2 described steel cable core conveying belt defective intelligent identification Methods; It is characterized in that: when carrying out feature extraction in the step 3; The characteristic parameter that is proposed comprises 12 temporal signatures of detection signal; Be M=12, and 12 temporal signatures are that peak-to-peak value, root-mean-square value, average amplitude, variance, root amplitude, kurtosis, ripple are wide respectively, waveform index, peak value index, pulse index, nargin index and kurtosis index; After carrying out feature extraction in the step 3, also need adopt data processor (2) that the characteristic parameter that is extracted is carried out feature reduction.

6. according to claim 1 or 2 described steel cable core conveying belt defective intelligent identification Methods, it is characterized in that: calculate the spacing d between any sample class h in sample class q and N the said sample class in the step 502 _QhAfter, the between class distance data of acquisition sample class q; In the step 504 repeatedly repeating step 501 to step 503, obtain the between class distance data and a type spacing sum of N said sample class; Subsequently, said data processor (2) is formed a between class distance symmetric matrix D with the between class distance data of N said sample class _{N * N}, and the between class distance data of each said sample class are positioned between class distance symmetric matrix D _{N * N}With the same line data in the delegation; The class spacing sum of N said sample class is respectively between class distance symmetric matrix D _{N * N}In each line data sum, and between class distance symmetric matrix D _{N * N}In each line data sum form array (Sumd (1), a Sumd (2) ... Sumd (N)).

Correspondingly, when in the step 505 the classification priority level Y of N said sample class being confirmed, its deterministic process is following:

Step 5051, initial parameter are set: the initial value to classification priority level Y and total sample number n ' is set the priority level of wherein classifying Y=0, total sample number n '=N respectively;

Step 5052, comparison array (Sumd (1), Sumd (2) ... Sumd (N)) size of current all data in is therefrom selected maximal value Sumd (L), wherein L=1,2 ... N, and be Y+1 with the classification priority level of sample class L, and this moment Y=Y+1, n'=N-1; Simultaneously, with between class distance symmetric matrix D _{N * N}In the L line data all put 0, with array (Sumd (1), Sumd (2) ... Sumd (N)) Sumd (L) in puts 0;

Step 5053, repeating step 5052 repeatedly are till n'=0.

7. according to claim 1 or 2 described steel cable core conveying belt defective intelligent identification Methods; It is characterized in that: N-1 said two disaggregated models are the fuzzy support vector machine model in the step 6; And carry out training sample in the step 4 when obtaining, include fuzzy membership μ in each training sample in N sample class _Qs, μ wherein _QsBe X _QsFuzzy membership to sample class Q under it.

8. according to claim 1 or 2 described steel cable core conveying belt defective intelligent identification Methods; It is characterized in that: when in the step 602 punishment parameters C and nuclear parameter γ being confirmed; Through data processor (2) and adopt improved genetic algorithm that selected punishment parameters C and nuclear parameter γ are optimized, its optimizing process is following:

Step 6021, initialization of population: the value of value and nuclear parameter γ that will punish parameters C is as body one by one; And be a population with a plurality of individual collections, all individualities in the said population all carry out forming the initialization population after the binary coding simultaneously; Wherein, the value of value of punishment parameters C and nuclear parameter γ be from the interval (0,1000] in a numerical value randomly drawing;

Each individual fitness value calculation in step 6022, the initialization population: all individual fitness value calculation methods are all identical in the initialization population; A plurality of said individualities in the initialization population, respectively corresponding a plurality of different disaggregated model z;

For any individuality in the said initialization population; Adopt training sample described in the step 5 to concentrate remaining N-Z+1 sample class; Disaggregated model z to corresponding with this individuality trains, and with the classification accuracy of this disaggregated model z as this individual fitness value;

After treating that all individual fitness values all calculate in the said initialization population, the corresponding again kind group mean fitness value that calculates said initialization population;

Step 6023, selection operation:, select fitness value is high in the said initialization population a plurality of individualities as progeny population according to all individual fitness values in the said initialization population that calculates in the step 6022;

Step 6024, interlace operation and mutation operation: the progeny population to choosing carries out interlace operation and mutation operation, obtains the progeny population of a new generation;

Each individual fitness value calculation in step 6025, the progeny population: all individual fitness value calculation methods are all identical in the progeny population; A plurality of said individualities in the progeny population, respectively corresponding a plurality of different disaggregated model z;

For any individuality in the said progeny population; Adopt training sample described in the step 5 to concentrate remaining N-Z+1 sample class; Disaggregated model z to corresponding with this individuality trains, and with the classification accuracy of this disaggregated model z as this individual fitness value;

After treating that all individual fitness values all calculate in the said progeny population, the corresponding again kind group mean fitness value that calculates said progeny population;

Step 6026, selection operation:, select fitness value is high in the said progeny population a plurality of individualities as progeny population according to all individual fitness values in the said progeny population that calculates in the step 6025;

Step 6027, judge whether to satisfy end condition: when evolutionary generation surpassed maximum adaptation degree value individual in predefined maximum evolutionary generation itmax or the progeny population more than or equal to predefined fitness setting value, genetic algorithm stopped also exporting the current the highest individuality of fitness value in the said progeny population that obtains; Otherwise, return step 6024.

9. according to the described steel cable core conveying belt defective of claim 8 intelligent identification Method, it is characterized in that: when carrying out interlace operation and mutation operation in the step 6024, according to crossover probability p _cCarry out interlace operation, and according to the variation Probability p _mCarry out mutation operation; Wherein,

$p_c=\left\{\begin{array}{cc}{p}_{c\mathrm{Max}}-\left(\frac{p_{c\mathrm{Max}}-p_{c\mathrm{Min}}}{\mathrm{It}\mathrm{Max}}\right)\&times;\mathrm{Iter}& ,f^{\&prime;}>f_{\mathrm{Avg}}\\ p_{c\mathrm{Max}}& ,f^{\&prime;}\&le;f_{\mathrm{Avg}}\end{array}\right.,$ $p_m=\left\{\begin{array}{cc}{p}_{m\mathrm{Max}}-\left(\frac{p_{m\mathrm{Max}}-p_{m\mathrm{Min}}}{\mathrm{It}\mathrm{Max}}\right)\&times;\mathrm{Iter}& ,f>f_{\mathrm{Avg}}\\ p_{m\mathrm{Max}}& ,f\&le;f_{\mathrm{Avg}}\end{array}\right.;$ In the formula, p _CmaxBe predefined maximum crossover probability, p _CminBe predefined minimum crossover probability, p _MmaxBe predefined maximum variation probability, p _MminBe predefined minimum variation probability, itmax is predefined maximum evolutionary generation, and iter is current evolutionary generation, f _AvgBe the current kind group mean fitness value that carries out the progeny population of interlace operation and mutation operation, f' is illustrated in bigger fitness value in two individuals that will intersect, the ideal adaptation degree value that f indicates to make a variation.

10. realize the intelligent identifying system of intelligent identification Method according to claim 1 for one kind; It is characterized in that: comprise and carry out electromagnetism charger, electromagnetism that electromagnetism loads and load a plurality of electromagnetic detecting unit of the remanent magnetism in the steel cable core conveying belt to be detected being detected the back in real time, a plurality of said electromagnetic detecting unit institute detection signal is carried out analyzing and processing and the data processor (2) of output steel cable core conveying belt defective classification to be detected and the host computer (7) that carries out two-way communication with data processor (2) automatically to be detected; A plurality of said electromagnetic detecting unit are all joined with signal conditioning circuit (5); Said signal conditioning circuit (5) joins with A/D change-over circuit (6), and said A/D change-over circuit (6) joins with data processor (2).

Images