CN105784340A - Air valve fault diagnosis method based on hybrid intelligent technology - Google Patents

Abstract

The invention discloses an air valve fault diagnosis method based on a hybrid intelligent technology. The method comprises the following steps of: S1, utilizing an EMD to extract an IMF component of a vibration signal and extract energy and sample entropy, containing fault information, of the IMF component, and forming a characteristic vector; S2, constructing sensitive characteristic subsets; S21, utilizing a CV-SVM to calculate the classification correct rate of each characteristic to a single fault mode; S22, selecting the characteristic highest in the classification correct rate as a most sensitive characteristic; S3, determining an optimal sensitive characteristic subset; S31, arranging fault modes according to the priority from high to low, and obtaining a groups of characteristic sequences; S32, removing the characteristics low in classification correct rate in the characteristic sequences; and S33, according to the classification correct rates of the characteristic subsets, selecting the subset highest in classification correct rate as the optimal characteristic subset. According to the invention, the method is combined as EMD and CV-SVM intelligent technologies and the like, three kinds of fault types of an air valve can be effectively identified, the diagnosis precision is high, and fault of a complex type can be accurately diagnosed.

Description

Valve fault diagnosis method based on Hybrid Intelligent Technology

Technical field

The invention belongs to air valve detection technique field, particularly to a kind of valve fault diagnosis method based on Hybrid Intelligent Technology.

Background technology

Reciprocating compressor, as a kind of universal machine, has purposes the most widely in fields such as oil, food, refrigeration. Air valve is one of most important parts of reciprocating compressor, and air valve once breaks down, and compressor will be made to lose ventilatory, Cisco unity malfunction.According to statistics, the compressor fault that air valve reason causes accounts for the 36% of fault sum, therefore, to air valve Do fault diagnosis to be extremely important.The fault diagnosis of reciprocating compressor air-valve is substantially belonged to pattern recognition ask Topic, needs to extract the characteristic parameter of valve fault signal, then fault type is carried out Classification and Identification.

Compressor valve fault-signal has non-stationary and nonlinear feature, and traditional time-frequency domain method is difficult to reflect comprehensively Signal fault feature.Gao Jingbo of Harbin Institute of Technology et al., uses Time-Frequency Analysis Method diagnosis valve fault, achieves Good effect, but the diagnosis effect for spring for valve fault is the most undesirable.Huang etc. propose empirical mode decomposition, Signal decomposition can be become limited intrinsic mode functions (Intrinsic Mode Function, IMF), these IMF smoothly by the method Highlight primary signal fault message in different frequency range；Fault message is primarily present in high frequency band, and noise information is main It is present in low-frequency band, the IMF of low-frequency band can be in by rejecting and reach the purpose of denoising.In terms of Classification and Identification, air valve It is big that fault data obtains difficulty, and has certain risk, causes sample size the most very limited, therefore neutral net side Method is inapplicable.The support vector machine that Vapnik proposes improves the generalization ability of learning machine by seeking structural risk minimization, It is particularly suitable for processing the classification and identification of small sample.Xing Xiuyu of Northcentral University et al. is by combined energy feature and Sample Entropy Feature, improves the accuracy of classification, but assemblage characteristic dimension is big, and interference component dramatically increases, and needs to be selected by feature Select and filter out sensitive features.

The fault characteristic information extracted merely with a certain method due to the fault diagnosis algorithm of the most a lot of air valves diagnoses, this Exist for fault signature representativeness not enough, thus cause diagnostic accuracy fault not enough, to complicated type to be difficult to Precise Diagnosis etc. Problem.

Summary of the invention

It is an object of the invention to overcome that diagnostic accuracy in existing valve fault diagnosis is low, diagnostic-type is difficult to the deficiency that confirms, There is provided one to combine EMD, CV-SVM, Feature Fusion and feature selection scheduling algorithm, can effectively identify air valve Three kinds of fault types, make diagnostic result valve fault diagnosis method based on Hybrid Intelligent Technology more accurately.

It is an object of the invention to be achieved through the following technical solutions: valve fault diagnosis method based on Hybrid Intelligent Technology, It is characterized in that, comprise the following steps:

S1, feature extraction, extract the IMF component of vibration signal with EMD, retains the IMF component comprising fault message, Reject the IMF component containing noise jamming；Extract energy and Sample Entropy, the composition characteristic vector of the IMF component retained；

S2, structure sensitive features subset, including following sub-step:

S21, calculated the classification to single failure pattern of each feature in the characteristic vector that obtains of S1 by CV-SVM algorithm Accuracy；

The highest feature of S22, selection sort accuracy is as the most sensitive feature of each fault mode；If classification accuracy rate is the highest Feature have multiple, then the feature that multiple classification accuracy rates are the highest is carried out Feature Fusion and feature compares, find out each fault The most sensitive feature of pattern；

S23, the most sensitive feature that S22 is obtained composition sensitive features subset；

S3, determine optimal feature subset, including following sub-step:

S31, compare fault mode priority, compare each most sensitive feature classification accuracy rate to corresponding fault mode, point The fault mode priority corresponding to most sensitive feature that class accuracy is the lowest is the highest, by fault mode the most from high in the end Arrangement, obtains a stack features sequence；

In S32, rejecting characteristic sequence, classification accuracy rate is less than the feature of average correct classification rate；

S33, select from characteristic sequence before k feature and sensitive features subset composition characteristic subset, by CV-SVM algorithm When checking k takes different values, the classification accuracy rate of the character subset formed, the subset conduct that selection sort accuracy is the highest Optimal feature subset；Wherein, k=1 ..., N, N are characterized the number of feature in sequence.

Further, step S1 concrete methods of realizing is: c_iT () represents the IMF component comprising fault message, i is for comprising event The number of the IMF component of barrier information；Construction feature vector includes following sub-step:

S11, calculating c_iThe energy value of (t):

$E_i=\underset{-\mathrm{\∞}}{\overset{+\mathrm{\∞}}{\∫}}{\left|c_i\left(t\right)\right|}^{2}dt---\left(1\right);$

S12, energy value to the IMF component comprising fault message are normalized, and construct energy feature vector T₁:

T₁=[x₁ x₂ … x_i] (2)；

S13, Sample Entropy are carried out gauge signal with seasonal effect in time series complexity and are produced the size of new model probability, and new model produces Probability the biggest explanation seasonal effect in time series complexity is the biggest, and corresponding Sample Entropy is the biggest；Calculate the IMF comprising fault message The Sample Entropy of component:

$SampEn(m,r,N)=-ln\[\frac{B^{m+1}\left(r\right)}{B^m\left(r\right)}\]---\left(3\right)$

Wherein, m is the dimension of new makeup time sequence, and m takes 1 or 2；B is ratio；R is distance threshold values, and the value of r is positioned at Between 1.5sd and 2.5sd, sd is the standard deviation of initial data；N is that time series data is counted；

S14, Sample Entropy to the IMF component comprising fault message are normalized, and obtain Sample Entropy characteristic vector T₂:

T₂=[x_i+1 x_i+2 … x_2i] (4)；

S15, by energy feature vector T₁With Sample Entropy characteristic vector T₂It is combined into union feature vector T=[T₁ T₂]。

Further, the Feature Fusion concrete methods of realizing of step S22 is: carry out the feature that multiple classification accuracy rates are the highest Weighted feature fusion calculation, formula is as follows:

$r=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{p}_i{x}_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{p}_i}---\left(5\right)$

In formula, n is Characteristic Number before merging, and r is fusion feature, p_iIt is characterized x_iClassification accuracy rate sum to other fault mode.

After weighted feature merges, needs assessment merge after feature with merge before feature compared with which is more preferable, step S22 Character select concrete methods of realizing be: compare the feature after fusion with merge before feature to such fault mode Classification accuracy rate, the feature that selection sort accuracy is the highest is output characteristic；If the feature that classification accuracy rate is the highest has multiple, Then comparing they classification accuracy rates to all fault modes, the feature that corresponding classification accuracy rate is the highest is output characteristic；

If output characteristic is the feature after merging, in order to avoid the increase of intrinsic dimensionality, and reduces redundancy, pass through phase relation The calculating of number, the feature before merging, rejecting one is maximally related with fusion feature, and the computing formula of correlation coefficient is as follows:

$P_{rx}=\frac{COV(r,x)}{\sqrt{D\left(r\right)\×D\left(x\right)}}---\left(6\right)$

Wherein, x is the feature before merging, P_rxFor the correlation coefficient of x Yu r, COV is covariance, and D is variance；

Feature after rejecting, is not involved in follow-up feature selection.

The invention has the beneficial effects as follows: the valve fault diagnosis method of the present invention combines EMD, CV-SVM, Feature Fusion And the intelligent algorithm such as feature selection, have the advantage that 1. utilize EMD method to process impact is strong, noise jamming is serious Non-stationary signal, eliminates the noise impact on diagnostic result, improves diagnostic accuracy；2. CV-SVM algorithm is used to count Calculate each feature classification accuracy rate to single failure pattern, use classification accuracy rate to weigh the contribution degree of fault signature, root Selecting most sensitive feature according to classification accuracy rate, algorithm is the most effective；3. Feature Fusion and feature selection, and phase relation are used Most sensitive feature is chosen in the calculating of number, it is to avoid the increase of intrinsic dimensionality, decreases redundancy；4. use based on CV-SVM The appraisal procedure of classification accuracy rate, is ranked up feature, eliminates interference characteristic, improve character subset coverage rate and Classification accuracy rate so that the optimal feature subset extracted can fully reflect the correlation behavior type of air valve.The present invention can be non- Often efficiently identifying three kinds of fault types of air valve, diagnostic accuracy is high, also can make the fault of complicated type and judging accurately, Can be widely used in all kinds of valve fault diagnosis.

Accompanying drawing explanation

Fig. 1 is the diagnostic method flow chart of the present invention；

Fig. 2 is the vibration signal of the air valve that valve block fracture occurs in the present embodiment；

Fig. 3 is the EMD decomposed signal figure of the air valve vibration signal that valve block fracture occurs in the present embodiment.

Detailed description of the invention

Further illustrate technical scheme below in conjunction with the accompanying drawings.

As it is shown in figure 1, the valve fault diagnosis method based on Hybrid Intelligent Technology of the present invention, comprise the following steps:

S1, feature extraction, extract vibration letter with EMD (Empirical Mode Decomposition, empirical mode decomposition) Number IMF component, non-stationary, nonlinear properties can be resolved into limited IMF component by EMD method, retain comprise therefore The IMF component of barrier information, rejects the IMF component containing noise jamming；Extract energy and the sample of the IMF component retained Entropy, composition characteristic vector；Its concrete methods of realizing is: c_iT () represents the IMF component comprising fault message, i is for comprising fault The number of the IMF component of information；Construction feature vector includes following sub-step:

S11, calculating c_iThe energy value of (t):

$E_i=\underset{-\mathrm{\∞}}{\overset{+\mathrm{\∞}}{\∫}}{\left|c_i\left(t\right)\right|}^{2}dt---\left(1\right);$

S12, energy value to the IMF component comprising fault message are normalized, and construct energy feature vector T₁:

T₁=[x₁ x₂ … x_i] (2)

The cycle of operation of compressor valve comprises expansion, air-breathing, compression, aerofluxus Four processes, when air valve breaks down, In sealed volume, pressure reduces, and the vibratory impulse power suffered by air valve reduces, and energy leakage occurs the most in a duty cycle, because of Energy variation situation in this each frequency range can relatively significantly characterize the fault type that air valve is occurred；

S13, Sample Entropy are carried out gauge signal with seasonal effect in time series complexity and are produced the size of new model probability, and new model produces Probability the biggest explanation seasonal effect in time series complexity is the biggest, and corresponding Sample Entropy is the biggest；Calculate the IMF comprising fault message The Sample Entropy of component:

$SampEn(m,r,N)=-ln\[\frac{B^{m+1}\left(r\right)}{B^m\left(r\right)}\]---\left(3\right)$

Wherein, m is the dimension of new makeup time sequence, and m takes 1 or 2；B is ratio；R is distance threshold values, and the value of r is positioned at Between 1.5sd and 2.5sd, sd is the standard deviation of initial data；N is that time series data is counted；The present embodiment takes m=2, R=1.5sd.

S14, Sample Entropy to the IMF component comprising fault message are normalized, and obtain Sample Entropy characteristic vector T₂:

T₂=[x_i+1 x_i+2 … x_2i] (4)；

S15, by energy feature vector T₁With Sample Entropy characteristic vector T₂It is combined into union feature vector T=[T₁ T₂]。

S2, structure sensitive features subset, including following sub-step:

S21, the fault diagnosis of gas valves of reciprocating compressors belong to many classification problems, and first this be accomplished by constructing multi-categorizer；This Bright employing support vector machine based on cross validation (CV-SVM) is as grader, with it to the classification of single failure pattern just Really rate weighs the contribution degree of feature；The each feature in the characteristic vector that S1 obtains is calculated to single by CV-SVM algorithm The classification accuracy rate of fault mode；

The highest feature of S22, selection sort accuracy is as the most sensitive feature of each fault mode；But due to the sample in reality This quantity is often limited, it is easy to causing multiple feature to have identical classification accuracy rate, feature selection is just caused by this Difficulty.If the feature that classification accuracy rate is the highest has multiple, then the feature that multiple classification accuracy rates are the highest is carried out Feature Fusion and Feature compares, and finds out the most sensitive feature of each fault mode；Wherein, the concrete methods of realizing of Feature Fusion is: to multiple The feature that classification accuracy rate is the highest is weighted Feature Fusion and calculates, and formula is as follows:

$r=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{p}_i{x}_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{p}_i}---\left(5\right)$

In formula, n is Characteristic Number before merging, and r is fusion feature, p_iIt is characterized x_iClassification accuracy rate sum to other fault mode；

After weighted feature merges, needs assessment merge after feature with merge before feature compared with which is more preferable, the present invention The feature comparative approach used is as follows: the feature before comparing the feature after fusion and merging is correct to the classification of such fault mode Rate, the feature that selection sort accuracy is the highest is output characteristic；If the feature that classification accuracy rate is the highest has multiple, then compare it Classification accuracy rate to all fault modes, the feature that corresponding classification accuracy rate is the highest is output characteristic；

If output characteristic is the feature after merging, in order to avoid the increase of intrinsic dimensionality, and reduces redundancy, pass through phase relation The calculating of number, the feature before merging, rejecting one is maximally related with fusion feature, and the computing formula of correlation coefficient is as follows:

$P_{rx}=\frac{COV(r,x)}{\sqrt{D\left(r\right)\×D\left(x\right)}}---\left(6\right)$

Wherein, x is the feature before merging, P_rxFor the correlation coefficient of x Yu r, COV is covariance, and D is variance；

Feature after rejecting, is not involved in follow-up feature selection；

S23, the most sensitive feature that S22 is obtained composition sensitive features subset；

S3, determine optimal feature subset, including following sub-step:

S31, compare fault mode priority, compare each most sensitive feature classification accuracy rate to corresponding fault mode, point The fault mode priority corresponding to most sensitive feature that class accuracy is the lowest is the highest, by fault mode the most from high in the end Arrangement, obtains a stack features sequence；

S32, calculate the average correct classification rate of characteristic sequence that S31 obtains, and reject classification accuracy rate in characteristic sequence and be less than The feature of average correct classification rate；

S33, select from characteristic sequence before k feature and sensitive features subset composition characteristic subset, by CV-SVM algorithm When checking k takes different values, the classification accuracy rate of the character subset formed, the subset conduct that selection sort accuracy is the highest Optimal feature subset；Wherein, k=1 ..., N, N are characterized the number of feature in sequence.

Technical scheme is further illustrated below in conjunction with specific embodiment.

Verify as a example by several malfunctions of certain type reciprocating compressor air-valve.The running status of air valve includes: valve block Fracture, valve block have 2 breach, few 2 spring these three fault types and four kinds of states of normal work.Original vibration signal Sample frequency 20000HZ, sampling number 65000.Fig. 2 is that compressor valve plate occurs acceleration vibration during fracture to believe Number, owing to being mingled with random noise signal, so the feature that causes of fault inconspicuous.

First the vibration signal to valve block fracture does EMD decomposition, and raw 9 the IMF components of common property, front 8 IMF components are Intrinsic mode function under different time scales, IMF9 is survival function, as shown in Figure 3.IMF1 and IMF2 reflection be Original vibration signal local feature in high band, here vibrational energy concentrate, time series complicated and changeable, contain master The fault message wanted.Frequency from IMF3 to IMF8 reduces successively as can be seen from Figure, and vibrational energy distribution is uneven, time Between the complexity of sequence be gradually lowered, though the fault message not having IMF1 and IMF2 to comprise enriches, but still suffer from some therefore Barrier information.IMF9 lines trend almost flat, vibrational energy ignores substantially, and time series is dull, is practically free of fault message. Therefore, in order to further fault signature extracts, only retain front 8 components of IMF.

By sampled point in units of 2500, altogether it is divided into 26 groups, 16 groups of training samples as CV-SVM, 10 groups of works For test sample.First, after being decomposed by EMD, carry out energy feature and the calculating of Sample Entropy feature.In Tables 1 and 2 Listing the Partial Feature vector of three kinds of fault modes and normal condition, totally 16 dimension, table 1 is energy feature, and table 2 is sample Entropy feature.

The energy feature vector of 1 four kinds of running statuses of table

The Sample Entropy characteristic vector of 2 four kinds of running statuses of table

The classification accuracy rate of the energy feature of 3 four kinds of running statuses of table

The Sample Entropy tagsort accuracy of 4 four kinds of running statuses of table

Utilize the characteristic vector in Tables 1 and 2, by each feature in CV-SVM calculating characteristic vector to single failure mould The classification accuracy rate of formula, result of calculation is as shown in Table 3 and Table 4.As shown in Table 3, x₃Classification accuracy rate to valve block fracture The highest, as the most sensitive feature characterizing valve block fracture.x₅The classification accuracy rate having 2 breach to valve block is the highest, will It has the most sensitive feature of 2 breach as characterizing valve block.x₁And x₁₄Identical to the classification accuracy rate of few 2 springs, according to Feature Fusion and system of selection are to x₁And x₁₄Process.First, weighted feature fusion formula is utilized to calculate fusion feature r₁。 Then, r is calculated₁The classification accuracy rate of two piece springs few to valve block is 87.5%, this and x₁And x₁₄Classification accuracy rate identical. According to Feature Fusion and system of selection, draw r₁Classification accuracy rate to all fault modes is 79.69%, more than x₁70.31% And x₁₄42.19%, therefore r₁Most sensitive feature for few 2 springs.Finally, r is calculated₁With x₁Correlation coefficient be 0.98, r₁With x₁₄Correlation coefficient be 0.30, it can thus be appreciated that r₁With x₁There is the strongest dependency, reject x₁。x₄、x₁₀And x₁₂Right The classification accuracy rate of normal condition is all 87.5%, in like manner, can be calculated x₄、x₁₀And x₁₂Fusion feature r₂To normally The classification accuracy rate of state is 93.75%, more than before merging the 87.5% of feature, so, fusion feature r₂For normal condition Most sensitive feature.Finally calculate r₂With x₄、x₁₀、x₁₂Correlation coefficient be respectively-0.54 ,-0.83 ,-0.7, reject feature x₁₀。 Therefore, four most sensitive features x₃、x₅、r₁And r₂Composition sensitive features subset.After obtaining sensitive features subset, reject surplus Feature relatively low to the classification accuracy rate of all fault modes in remaining feature.x₈、x₁₂、x₁₃、x₁₄、x₁₅、x₁₆To all events The classification accuracy rate of barrier pattern is not higher than 50%, rejects them.

In sensitive features subset, x₃Classification accuracy rate to valve block fracture is 100%, x₅Valve block is had the classification of 2 breach Accuracy is 100%, r₁Classification accuracy rate to few 2 springs is 87.5%, r₂To the classification accuracy rate of normal condition it is 93.75%.By contrast, r₁Minimum to the classification accuracy rate of few 2 springs, therefore, in order to improve to few 2 springs this The discrimination of one fault type, arrange its priority for the highest, needs according to fault mode classification accuracy from high to low Residue character is ranked up by order.X is can be seen that from table 3 and table 4₆And x₇, x₉And x₁₁Classification to few 2 springs Accuracy is identical, and therefore according to Feature Fusion and system of selection, we respectively obtain x₆And x₇Fusion feature r₃, x₉And x₁₁ Fusion feature r₄。r₃Classification accuracy rate to few 2 springs is 87.5%, r₄To the classification accuracy rate of few 2 springs it is 81.25%.It is ordered as according to the classification accuracy rate lacking 2 springs: r₃、r₄、x₂、x₄。

K feature and sensitive features subset composition characteristic subset before selecting, but k value is not to be the bigger the better, and k value is big, then say The feature of bright selection is many, causes amount of calculation to increase, and also easily produces high correlated characteristic；And k value is little, it is likely to result in again excellent The loss of elegant feature.Here, we are respectively provided with k is 1 and 2, compares its quality, and choice accuracy is big and dimension is little Character subset is optimal feature subset.Utilize CV-SVM as shown in table 5 to the training result of character subset.

The training result of table 5 character subset

As can be seen from the table, the precision of ratio of precision k=2 during k=1 is higher.Therefore as k=1, character subset Z=[x₃ x₅ r₁ r₂ r₃] it is optimal feature subset, utilize CV-SVM that three kinds of malfunctions of air valve are tested, its discrimination Being 100%, effect is the best.Therefore, the optimal feature subset that the Hybrid Intelligent Technology proposed by the present invention is extracted can fill Divide the correlation behavior type of reflection air valve, can effectively identify three kinds of fault types of air valve with it；This most fully says Understand institute of the present invention extracting method effectiveness in terms of detection valve fault.

Those of ordinary skill in the art is it will be appreciated that embodiment described here is to aid in the reader understanding present invention's Principle, it should be understood that protection scope of the present invention is not limited to such special statement and embodiment.This area common It is various specifically that technical staff can make various other without departing from essence of the present invention according to these technology disclosed by the invention enlightenment Deformation and combination, these deformation and combination are the most within the scope of the present invention.

Claims (4)

1. valve fault diagnosis method based on Hybrid Intelligent Technology, it is characterised in that comprise the following steps:

S1, feature extraction, extract the IMF component of vibration signal with EMD, retains the IMF component comprising fault message, Reject the IMF component containing noise jamming；Extract energy and Sample Entropy, the composition characteristic vector of the IMF component retained；

S2, structure sensitive features subset, including following sub-step:

S21, calculated the classification to single failure pattern of each feature in the characteristic vector that obtains of S1 by CV-SVM algorithm Accuracy；

The highest feature of S22, selection sort accuracy is as the most sensitive feature of each fault mode；If classification accuracy rate is the highest Feature have multiple, then the feature that multiple classification accuracy rates are the highest is carried out Feature Fusion and feature compares, find out each fault The most sensitive feature of pattern；

S23, the most sensitive feature that S22 is obtained composition sensitive features subset；

S3, determine optimal feature subset, including following sub-step:

S31, compare fault mode priority, compare each most sensitive feature classification accuracy rate to corresponding fault mode, point The fault mode priority corresponding to most sensitive feature that class accuracy is the lowest is the highest, by fault mode the most from high in the end Arrangement, obtains a stack features sequence；

In S32, rejecting characteristic sequence, classification accuracy rate is less than the feature of average correct classification rate；

S33, select from characteristic sequence before k feature and sensitive features subset composition characteristic subset, by CV-SVM algorithm When checking k takes different values, the classification accuracy rate of the character subset formed, the subset conduct that selection sort accuracy is the highest Optimal feature subset；Wherein, k=1 ..., N, N are characterized the number of feature in sequence.

Valve fault diagnosis method based on Hybrid Intelligent Technology the most according to claim 1, it is characterised in that described Step S1 concrete methods of realizing be: c_iT () represents the IMF component comprising fault message, i is the IMF comprising fault message The number of component；Construction feature vector includes following sub-step:

S11, calculating c_iThe energy value of (t):

$E_i=\underset{-\mathrm{\∞}}{\overset{+\mathrm{\∞}}{\∫}}|c_i\left(t\right){|}^{2}dt---\left(1\right);$

S12, energy value to the IMF component comprising fault message are normalized, and construct energy feature vector T₁:

T₁=[x₁ x₂ … x_i] (2)；

S13, calculate the Sample Entropy of IMF component comprising fault message:

$SampEn(m,r,N)=-ln\[\frac{B^{m+1}\left(r\right)}{B^m\left(r\right)}\]---\left(3\right)$

Wherein, m is the dimension of new makeup time sequence, and m takes 1 or 2；B is ratio；R is distance threshold values, and the value of r is positioned at 1.5sd and between 2.5sd, sd is the standard deviation of initial data；N is that time series data is counted；

S14, Sample Entropy to the IMF component comprising fault message are normalized, and obtain Sample Entropy characteristic vector T₂:

T₂=[x_i+1 x_i+2 … x_2i] (4)；

S15, by energy feature vector T₁With Sample Entropy characteristic vector T₂It is combined into union feature vector T=[T₁ T₂]。

Valve fault diagnosis method based on Hybrid Intelligent Technology the most according to claim 1, it is characterised in that described The Feature Fusion concrete methods of realizing of step S22 be: the feature that multiple classification accuracy rates are the highest is weighted Feature Fusion Calculating, formula is as follows:

$r=\frac{\underset{i=1}{\overset{n}{\Σ}}{p}_i{x}_i}{\underset{i=1}{\overset{n}{\Σ}}{p}_i}---\left(5\right)$

In formula, n is Characteristic Number before merging, and r is fusion feature, p_iIt is characterized x_iClassification accuracy rate sum to other fault mode.

Valve fault diagnosis method based on Hybrid Intelligent Technology the most according to claim 3, it is characterised in that described Step S22 character select concrete methods of realizing be: compare the feature after fusion with merge before feature to such The classification accuracy rate of fault mode, the feature that selection sort accuracy is the highest is output characteristic；If the spy that classification accuracy rate is the highest Having levied multiple, then compared they classification accuracy rates to all fault modes, the feature that corresponding classification accuracy rate is the highest is defeated Go out feature；

If output characteristic is the feature after merging, by the calculating of correlation coefficient, the feature before merging is rejected one with Fusion feature is maximally related, and the computing formula of correlation coefficient is as follows:

$P_{rx}=\frac{COV(r,x)}{\sqrt{D\left(r\right)\×D\left(x\right)}}---\left(6\right)$

Wherein, x is the feature before merging, P_rxFor the correlation coefficient of x Yu r, COV is covariance, and D is variance；

Feature after rejecting, is not involved in follow-up feature selection.