CN109784284A - The self study recognition methods of working condition based on rotating machinery - Google Patents
The self study recognition methods of working condition based on rotating machinery Download PDFInfo
- Publication number
- CN109784284A CN109784284A CN201910055613.8A CN201910055613A CN109784284A CN 109784284 A CN109784284 A CN 109784284A CN 201910055613 A CN201910055613 A CN 201910055613A CN 109784284 A CN109784284 A CN 109784284A
- Authority
- CN
- China
- Prior art keywords
- data
- signal
- rotating machinery
- working condition
- wavelet transformation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Abstract
The present invention provides a kind of self study recognition methods of working condition based on rotating machinery, including five steps: data preparation, data prediction, feature extraction, Feature Dimension Reduction, Classification and Identification；Normalized pretreatment is done firstly for signal, secondly feature extraction is done to signal using Deep Scattering Spectrum method, the characteristic value dimension obtained at this time is very high, then Feature Dimension Reduction, Classification and Identification is finally done for signal based on support vector machines, obtains the state of machine.The present invention is the mass data of the long term monitoring of the vibration signal based on rotating machinery, using the algorithm of machine learning as core, assists improving the high-class discrimination of the state and failure to mechanical equipment with digital signal processing algorithm；The fault signature of enough discriminations is unable to get for traditional signal processing method, it can use machine learning algorithm and improve identification, and identification current working status, long term monitoring data is classified, identifies equipment state and failure with reaching high-accuracy.
Description
Technical field
The present invention relates to the fields of machine learning algorithm, and in particular, to the self-study of the working condition based on rotating machinery
Practise recognition methods.
Background technique
The condition monitoring and fault diagnosis of mechanical equipment refers to using modern science and technology and instrument, according to mechanical equipment
The damage status changed to judge working condition or mechanical structure inside machine of (system, structure) external information parameter, really
Property, degree, classification and the position for determining failure, forecast its development trend, and study the mechanism of failure generation.
Mechanical equipment state monitoring and fault diagnosis technology is to ensure one of the basic measures of equipment safety operation, whirler
The status monitoring of tool, first choice install sensor in equipment, mainly identify equipment according to the vibration signal of acceleration acquisition
Operating status, and then identify and be out of order, its essence is understanding and grasping the state of equipment in the process of running, pre- measurement equipment
Reliability determines that its whole or part is normal or abnormal.It can make early prediction to the development of equipment fault, to appearance event
The reason of barrier, position, degree of danger etc. are identified and are evaluated, and forecast that traditional processing of failure is the calculation based on digital signal
Method obtains the characteristic value of signal, recycles statistical analysis and pattern-recognition decision device state, and such method has had centainly
Use value, but in systems in practice, vibration signal change greatly, to establish data model complicated, lead to such processing method
The state of equipment and the accuracy of fault identification be not high.
Summary of the invention
For the defects in the prior art, the object of the present invention is to provide the self studies of the working condition based on rotating machinery
Recognition methods.The present invention provides a set of algorithm based on machine learning, is gone out based on wavelet transformation and the algorithm development of machine learning
The recognizer of a set of operating status and failure for monitoring the equipment with vibration signal；It is the vibration based on rotating machinery
The mass data of the long term monitoring of signal assists improving with digital signal processing algorithm using the algorithm of machine learning as core
The high-class discrimination of state and failure to mechanical equipment.
According to an aspect of the present invention, the self study recognition methods of the working condition based on rotating machinery is provided, including
Following steps:
Step 1, data preparation: acquiring the floor data of monitoring device, and the floor data of acquisition is pressed to the duration of 2s or 3s
Intercept the training sample of the original sample and machine learning algorithm as processing；
Data prediction: step 2 is counted using mean value and standard deviation of the z-score method for normalizing to initial data
According to standardization；Treated data fit standardized normal distribution, mean value 0, standard deviation 1 convert function are as follows:
Wherein, x is expressed as original signal in formula, and μ is expressed as the mean value of initial data, and σ is expressed as the mark of initial data
Quasi- poor, x* is expressed as the signal after normalization；In order to enable normalization is more reasonable, training data and test data needs are separately returned
One changes；
Feature extraction: step 3 obtains the parameter of 6 grades of interative computations using Deep Scattering Spectrum algorithm
Feature connects six groups of parameter attributes to obtain primitive character, and dimension is 7347 dimensions, to remove the time variation of signal, and is obtained
Complete signal；The core of step 3 is the time variation for removing signal, and the vibration signal in equipment running process is time varying signal,
The information of high frequency section is unstable, and Fourier transformation cannot remove the time variation of signal, and Meier frequency spectrum is by carrying out frequency
These coefficients can averagely be stablized, but have information loss；The present invention is in order to stablize time varying signal, and analyzes these loss
Information selects the available equivalent Meier frequency spectrum of the wavelet transformation of constant Q filter group, Meier frequency spectrum using wavelet transformation
The information of high frequency can be lost, but the Meier frequency spectrum that wavelet transformation obtains remains high-frequency information, therefore continues to do small echo to it
Transformation, can recover the information that high frequency section has；
Feature Dimension Reduction: step 4 first carries out dimensionality reduction to primitive character using t-test method, by the spy of conspicuousness p < 0.05
Sign chooses, and dimension is reduced to 3849 dimensions, then uses Feature Generating Machine algorithm dimensionality reduction again, obtains
The new feature value of 140 dimensions；The dimension of every one piece of data is all very high, but the selection of base can obtain lax spy in wavelet transformation
Levy coefficient, that is to say, that most of characteristic quantity is close to the number of 0 very little, this just offers convenience for dimensionality reduction；In addition, from big data
From the point of view of theory, although characteristic value dimension is very high, often most of dimension and output are unrelated, it is therefore desirable to carry out feature
Dimensionality reduction；
Step 5, Classification and Identification: the characteristic value after dimensionality reduction is sent into support vector machines, and Training Support Vector Machines carry out
Classification and Identification finally obtains the algorithm of Classification and Identification.
Preferably, the floor data includes the data of 6 kinds of operating conditions, the data of respectively 3 kinds nominal situations: starting, nothing
The data of load idle running, low speed operation bringing onto load and 3 kinds of unusual service conditions: high speed bringing onto load, rotor unbalance, bearing fault
And lubrication trouble；The original sample sum is 3723, wherein each 521, the sample of 3 kinds of nominal situations, 3 kinds of unusual service conditions
Each 720, sample.Floor data includes that sensor collects the different works of rotary machine (such as engine, bearing, gear-box)
Make the vibration signal of state.
Preferably, the characteristic extraction procedure in the step 3 the following steps are included:
(1) vibration signal of equipment running process is put down using constant Q filter group progress wavelet transformation, modulus and time
Steady three step process obtains equivalent Meier frequency spectrum, to obtain the short-term stationarity characteristic of signal；
(2) high-frequency information compressed in the Meier frequency spectrum data for restoring equivalent by continuing to do wavelet transformation.
By repeatedly based on the interative computation of wavelet transformation, eliminating the time variation of signal, and will not lossing signal height
Frequency information.Wherein first time iteration, i.e. step (1) have obtained equivalent Meier frequency spectrum, and equivalent Meier frequency spectrum can remove letter
Number time variation, be further continued for doing the high-frequency information that wavelet transformation for several times restores the middle compression of Meier frequency spectrum.Calculate equivalent Meier
Modular arithmetic after frequency spectrum, then by hiding high-frequency information by a series of transformation of new small echos restores, this process is defined
For scattering transformation.
Preferably, the step (1) obtain equivalent Meier frequency spectrum calculating process the following steps are included:
A) the small echo ψ for the use of octave resolution ratio being Q1 to sample of signal xλ1Do wavelet transformation, wherein the wavelet transformation
The basic function used is the bandpass filter that one group of Q value is Q1；It is verified by test, Q1 is set as having good effect when 8；
B) modular arithmetic is carried out to the output of the wavelet transformation, replaces square transporting with First-Order Mode " | | " during modular arithmetic
It calculates, substitutes formula (1) by formula (2),
Formula (1) are as follows:
Formula (2) are as follows:
Because biggish wavelet coefficient can be squared operator amplification, in order to avoid being amplified to exceptional value, the operation
Cheng Zhongyong First-Order Mode " | | " replace square operation；
C) result of the modular arithmetic is equalized using the low-pass filter φ (t) that size is T, it is ensured that with time migration
Local invariant, obtain equivalent Meier frequency spectrum, equivalent Meier spectral coefficient, i.e. single order scattering coefficient are as follows:
S1X (t, λ1)=| x* ψ λ1|*φ(t)
Wherein, the x is expressed as the sample of original signal, and λ is expressed as the centre frequency of bandpass filter, when t is expressed as
Between, ψ is the base of Wavelet transformation, and basic function is one group of bandpass filter, and the meaning that x, λ, t, ψ are indicated in following formula is identical.
φ (t) is a normalized low-pass filter, and is met in 0~T time to the result etc. of φ (t) integral
A length of 2s or 3s when in the duration that 1, t is sample, the present invention.
Preferably, continue to do wavelet transformation in the step (2) the following steps are included:
A) obtained single order scattering coefficient is done into second of wavelet transformation, and modulus and carries out obtaining second order after the time is steady
Scattering coefficient, formula are as follows:
This step can be with the high-frequency information of recovered part signal；
B) processing for repeating step a) for several times, calculates the cascade of m wavelet convolution and modulus, obtains m rank scattering coefficient, public
Formula is as follows:
The every wavelet transformation by a step a) of high-frequency information, so that it may it is restore a part more, but its amount restored is gradually
It reduces, the high-frequency information of signal can restore completely substantially after the wavelet transformation of step a) for several times.
Although equivalent Meier frequency spectrum obtains the stationarity of time, high-frequency information is compressed, therefore step a) is obtained
To single order scattering coefficient do second of wavelet transformation, and modulus and do the time it is steady after calculating, the height of signal can be restored
The advantages of information of frequency, this is also the equivalent Meier frequency spectrum that wavelet transformation obtains, normal Meier frequency spectrum operation medium-high frequency information
It can lose, it is irrecoverable.
Preferably, processing 6 times of step a) are repeated in the step b), the m is equal to 6.It is every after 6 wavelet transformations
One section of sample data obtains the characteristic quantity of 7347 dimensions, and high-frequency information can be restored to original 98% at this time, is further continued for carrying out small echo
The high-frequency information amount for converting its recovery is very little, therefore selects 6 grades of wavelet transformations of progress most suitable.
Preferably, the scattering coefficient of every single order is that the First-Order Mode based on upper single order is calculated:
| W | and x=(x* φ (t), | x* ψλ(t)|)T ∈ R, λ ∈ Λ。
Preferably, the wavelet transformation meets the following conditions:
One, it is the constant bandpass filter of Q value；
Two,
Three, to parse small echo, and the centre frequency for parsing small echo normalizes to 1.Parsing small echo has quadrature phase
Phase information meets:The modulus for parsing wavelet coefficient is construed to subband Hilbert envelope solution by wherein ω < 0
It adjusts.Demodulation is used for separate carrier, and obtains envelope.Although small echo modulus operator eliminates complex phase, it will not lose letter
Breath, because remaining the time change of multiple dimensioned envelope.Signal cannot be rebuild from the modulus of its Fourier transform, but can be from
Restore in the modulus of its wavelet transformation.Since time variable t is not by double sampling, so wavelet transformation has more than original signal
More coefficients.When filter has significant frequency overlapping, these coefficients are high redundancies.
Preferably, Feature Generating Machine algorithm is divided into the step 4: worst case analysis and
Sub- problem optimizes two steps, and characteristic, penalty factor, loss function and the number of iterations in algorithm are freely arranged,
It adjusts.
Characteristic B=7 in preferred algorithm, penalty factor=12, loss function loss_type='squred_
Hinge' and the number of iterations max_iter=20.
Preferably, the characteristic value of training data is sent into the step 5 as the input data of algorithm and is used as classifier
Support vector machines after, algorithm model is obtained using the training of 10-fold cross validation method, and excellent by continuously adjusting parameter
Change, after inputting new data again, support vector machines can automatic Classification and Identification.
Compared with prior art, the present invention have it is following the utility model has the advantages that
(1) the self study recognition methods of the working condition of the present invention based on rotating machinery, design principle are as follows: first
Normalized pretreatment is done for signal, feature secondly is done to signal using Deep Scattering Spectrum method and is mentioned
It takes, the characteristic value dimension obtained at this time is very high, then Feature Dimension Reduction, finally does classification for signal based on support vector machines and knows
Not, the state of machine is obtained；
(2) the self study recognition methods of the working condition of the present invention based on rotating machinery is based on rotating machinery
Vibration signal long term monitoring mass data, using the algorithm of machine learning as core, auxiliary with digital signal processing algorithm,
Improve the high-class discrimination of the state and failure to mechanical equipment；
(3) the self study recognition methods of the working condition of the present invention based on rotating machinery, for traditional signal
Processing method is unable to get the fault signature of enough discriminations, can use machine learning algorithm and improves identification, Yi Jishi
Other current working status, is classified long term monitoring data, identifies equipment state and failure with reaching high-accuracy；
(4) the self study recognition methods of the working condition of the present invention based on rotating machinery is based on machine learning side
Method handles the vibration signal of rotating machinery, to identify the working condition of rotating machinery, further promotes monitoring system
The diagnosis capability united to the failure of equipment；
(5) the self study recognition methods of the working condition of the present invention based on rotating machinery, algorithm is simple, and design is skilful
It is wonderful, significant effect；
(6) the self study recognition methods of the working condition of the present invention based on rotating machinery, stable and reliable operation are fitted
Close a wide range of promote and apply.
Detailed description of the invention
Upon reading the detailed description of non-limiting embodiments with reference to the following drawings, other feature of the invention,
Objects and advantages will become more apparent upon:
Fig. 1 is step block diagram；
Fig. 2 is scattering transformation schematic diagram.
Specific embodiment
The present invention is described in detail combined with specific embodiments below.Following embodiment will be helpful to the technology of this field
Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that the ordinary skill of this field
For personnel, without departing from the inventive concept of the premise, various modifications and improvements can be made.These belong to the present invention
Protection scope.
Embodiment
The present embodiment is related to the self study recognition methods of the working condition based on rotating machinery, step block diagram such as attached drawing 1
It is shown, comprising the following steps:
Step 1, data preparation: acquiring the floor data of monitoring device, and the floor data of acquisition is pressed to the duration of 2s or 3s
Intercept the training sample of the original sample and machine learning algorithm as processing；
Data prediction: step 2 is counted using mean value and standard deviation of the z-score method for normalizing to initial data
According to standardization；Treated data fit standardized normal distribution, mean value 0, standard deviation 1 convert function are as follows:
Wherein, x is expressed as original signal in formula, and μ is expressed as the mean value of initial data, and σ is expressed as the mark of initial data
Quasi- poor, x* is expressed as the signal after normalization；In order to enable normalization is more reasonable, training data and test data needs are separately returned
One changes；
Feature extraction: step 3 obtains the parameter of 6 grades of interative computations using Deep Scattering Spectrum algorithm
Feature connects six groups of parameter attributes to obtain primitive character, and dimension is 7347 dimensions, to remove the time variation of signal, and is obtained
Complete signal；The core of step 3 is the time variation for removing signal, and the vibration signal in equipment running process is time varying signal,
The information of high frequency section is unstable, and Fourier transformation cannot remove the time variation of signal, and Meier frequency spectrum is by carrying out frequency
These coefficients can averagely be stablized, but have information loss；The present invention is in order to stablize time varying signal, and analyzes these loss
Information selects the available equivalent Meier frequency spectrum of the wavelet transformation of constant Q filter group, Meier frequency spectrum using wavelet transformation
The information of high frequency can be lost, but the Meier frequency spectrum that wavelet transformation obtains remains high-frequency information, therefore continues to do small echo to it
Transformation, can recover the information that high frequency section has；
Feature Dimension Reduction: step 4 first carries out dimensionality reduction to primitive character using t-test method, by the spy of conspicuousness p < 0.05
Sign chooses, and dimension is reduced to 3849 dimensions, then uses Feature Generating Machine algorithm dimensionality reduction again, obtains
The new feature value of 140 dimensions；The dimension of every one piece of data is all very high, but the selection of base can obtain lax spy in wavelet transformation
Levy coefficient, that is to say, that most of characteristic quantity is close to the number of 0 very little, this just offers convenience for dimensionality reduction；In addition, from big data
From the point of view of theory, although characteristic value dimension is very high, often most of dimension and output are unrelated, it is therefore desirable to carry out feature
Dimensionality reduction；
Step 5, Classification and Identification: the characteristic value after dimensionality reduction is sent into support vector machines, and Training Support Vector Machines carry out
Classification and Identification finally obtains the algorithm of Classification and Identification.
Further, the floor data includes the data of 6 kinds of operating conditions, the data of respectively 3 kinds nominal situations: starting,
The data of tick over, low speed operation bringing onto load and 3 kinds of unusual service conditions: high speed bringing onto load, rotor unbalance, bearing event
Barrier and lubrication trouble；The original sample sum is 3723, wherein each 521, the sample of 3 kinds of nominal situations, 3 kinds of abnormal works
Each 720, the sample of condition.Floor data includes that sensor collects rotary machine (such as engine, bearing, gear-box) difference
The vibration signal of working condition.
Further, the characteristic extraction procedure in the step 3 the following steps are included:
(1) vibration signal of equipment running process is put down using constant Q filter group progress wavelet transformation, modulus and time
Steady three step process obtains equivalent Meier frequency spectrum, to obtain the short-term stationarity characteristic of signal；
(2) high-frequency information compressed in the Meier frequency spectrum data for restoring equivalent by continuing to do wavelet transformation.
By repeatedly based on the interative computation of wavelet transformation, eliminating the time variation of signal, and will not lossing signal height
Frequency information.Wherein first time iteration, i.e. step (1) have obtained equivalent Meier frequency spectrum, and equivalent Meier frequency spectrum can remove letter
Number time variation, be further continued for doing the high-frequency information that wavelet transformation for several times restores the middle compression of Meier frequency spectrum.Calculate equivalent Meier
Modular arithmetic after frequency spectrum, then by hiding high-frequency information by a series of transformation of new small echos restores, this process is defined
For scattering transformation.
Further, the step (1) obtain equivalent Meier frequency spectrum calculating process the following steps are included:
A) the small echo ψ for the use of octave resolution ratio being Q1 to sample of signal xλ1Do wavelet transformation, wherein the wavelet transformation
The basic function used is the bandpass filter that one group of Q value is Q1；It is verified by test, Q1 is set as having good effect when 8；
B) modular arithmetic is carried out to the output of the wavelet transformation, replaces square transporting with First-Order Mode " | | " during modular arithmetic
It calculates, substitutes formula (1) by formula (2),
Formula (1) are as follows:
Formula (2) are as follows:
Mx (t, λ) ≈ ∫ | ∫ x (u) ψλ(v-u)du||φ(v-t)|dv
=| x* ψλ|*|φ|(t)；
Because biggish wavelet coefficient can be squared operator amplification, in order to avoid being amplified to exceptional value, the operation
Cheng Zhongyong First-Order Mode " | | " replace square operation；
C) result of the modular arithmetic is equalized using the low-pass filter φ (t) that size is T, it is ensured that with time migration
Local invariant, obtain equivalent Meier frequency spectrum, equivalent Meier spectral coefficient, i.e. single order scattering coefficient are as follows:
Wherein, the x is expressed as the sample of original signal, and λ is expressed as the centre frequency of bandpass filter, when t is expressed as
Between, ψ is the base of Wavelet transformation, and basic function is one group of bandpass filter, and the meaning that x, λ, t, ψ are indicated in following formula is identical.
φ (t) is a normalized low-pass filter, and is met in 0~T time to the result etc. of φ (t) integral
A length of 2s or 3s when in the duration that 1, t is sample, the present invention.
Further, continue to do wavelet transformation in the step (2) the following steps are included:
A) obtained single order scattering coefficient is done into second of wavelet transformation, and modulus and carries out obtaining second order after the time is steady
Scattering coefficient, formula are as follows:
This step can be with the high-frequency information of recovered part signal；
B) processing for repeating step a) for several times, calculates the cascade of m wavelet convolution and modulus, obtains m rank scattering coefficient, public
Formula is as follows:
The every wavelet transformation by a step a) of high-frequency information, so that it may it is restore a part more, but its amount restored is gradually
It reduces, the high-frequency information of signal can restore completely substantially after the wavelet transformation of step a) for several times.
Although equivalent Meier frequency spectrum obtains the stationarity of time, high-frequency information is compressed, therefore step a) is obtained
To single order scattering coefficient do second of wavelet transformation, and modulus and do the time it is steady after calculating, the height of signal can be restored
The advantages of information of frequency, this is also the equivalent Meier frequency spectrum that wavelet transformation obtains, normal Meier frequency spectrum operation medium-high frequency information
It can lose, it is irrecoverable.
Further, processing 6 times of step a) are repeated in the step b), the m is equal to 6.After 6 wavelet transformations
Each section of sample data obtains the characteristic quantity of 7347 dimensions, and high-frequency information can be restored to original 98% at this time, is further continued for carrying out small
The high-frequency information amount of its recovery of wave conversion is very little, therefore selects 6 grades of wavelet transformations of progress most suitable.
Further, the scattering coefficient of every single order is that the First-Order Mode based on upper single order is calculated:
| W | and x=(x* φ (t), | x* ψλ(t)|)T ∈ R, λ ∈ Λ。
Further, the wavelet transformation meets the following conditions:
One, it is the constant bandpass filter of Q value；
Two,
Three, to parse small echo, and the centre frequency for parsing small echo normalizes to 1.Parsing small echo has quadrature phase
Phase information meets:The modulus for parsing wavelet coefficient is construed to subband Hilbert envelope solution by wherein ω < 0
It adjusts.Demodulation is used for separate carrier, and obtains envelope.Although small echo modulus operator eliminates complex phase, it will not lose letter
Breath, because remaining the time change of multiple dimensioned envelope.Signal cannot be rebuild from the modulus of its Fourier transform, but can be from
Restore in the modulus of its wavelet transformation.Since time variable t is not by double sampling, so wavelet transformation has more than original signal
More coefficients.When filter has significant frequency overlapping, these coefficients are high redundancies.
Further, Feature Generating Machine algorithm is divided into the step 4: worst case analysis
Optimizing two steps with sub- problem, characteristic, penalty factor, loss function and the number of iterations in algorithm are freely arranged,
It adjusts.
Further, the characteristic B=7 in algorithm, penalty factor=12, loss function loss_type='
Squred_hinge ' and the number of iterations max_iter=20.
Further, it is sent into the characteristic value of training data as the input data of algorithm as classification in the step 5
After the support vector machines of device, algorithm model is obtained using the training of 10-fold cross validation method, and by continuously adjusting parameter
Optimization, after inputting new data again, support vector machines can automatic Classification and Identification.
Further, specific step is as follows for Deep Scattering Spectrum algorithm in the step 3:
Firstly, the high-frequency information for obtaining local transitions average on a time constant descriptor x, x passes through small echo mould
Transformation of variables restores, and formula is as follows:
It is calculated with the small echo with octave frequency division resolution Q1.For the signal of acoustical vibration class, Q1=8 is set, it is fixed
Justice has the small echo of identical frequency resolution ratio with frequency filter.Collected sound vibration signal is low on rotating machinery
The lower energy very little of frequency, so S0X (t) is approximately equal to 0.By average small echo modulus coefficient, equivalent Meier spectral coefficient is obtained:
The coefficient is also known as single order scattering coefficient, is applied to each | x* ψλ1| the second small echo modular transformation | W2| meter
It calculates, complementary high-frequency wavelet coefficient is also provided:
Small echo ψλ2With the octave resolution ratio Q2 different from Q1, small echo selection obtains sparse expression, it means that will believe
Number information concentrates on wavelet coefficient as few as possible.These coefficients are equalized by the low-pass filter φ (t) that size is T, really
The local invariant with time migration is protected, as coefficient of first order, defines second order dispersion coefficient:
By to each | | x* ψλ1|*ψλ2| apply the modular transformation of third small echo | W3| to calculate these average values.It passes through volume
New one group small echo ψ of the product with octave resolution ratio Q3λ3To calculate their wavelet coefficient.The iteration process defines any order
The scattering coefficient of m.
For any m >=1, the small echo modulus convolution of iteration is written to:
Wherein m rank small echo ψλmWith octave resolution ratio Qm, and meet stability condition.Average U is indicated with φ (t)mX,
Obtain m rank scattering coefficient:
In UmApplied on x | Wm+1| calculate SmX and Umm+1X:
|Wm+1|UmX=(SmX, Um+1x)
Therefore, by initializing UmX=x and recursive calculation defines the scattering point of maximum order 1 in the section 0≤m≤1
Solution.Scattering transformation is as shown in Fig. 2, scattering transformation iteration small echo modulus operator | Wm| U is stored in calculatemM small echo in x
The cascade of convolution sum modulus, and export average scattering coefficient Smx.Finally Scattering of Vector assembles all scattering systems on 0≤m≤1
Number:
Sx=(Smx)0≤m≤l
The nonlinear scattering cascade of convolution sum can also be construed to convolutional network, wherein UmX is m-th of intranet network layers
Coefficient sets.These networks have been demonstrated highly effective to the classification of sound and vibration class signal.However, with Standard convolution network
Difference, each such layer have output SmX=Um, rather than just the last layer.In addition, all filters are all predefined
Small echo, and do not learn from training data.Scattering transformation, without study, expresses one kind of signal just as MFCC
Invariance.It depends on the previous message of this calculative seed type invariance, and signal is relative to time shift and time herein
There are a kind of invariance for torsional deformation.When not such information is available, or if when the source of variation is much more complex, just
It needs to learn from example, this is being very suitable to deep-neural-network of the task.
Small echo octave resolution ratio is optimized in each layer of m, in next layer of generation Sparse Wavelet coefficient.This is preferably
Stick signal information.Sparsity seems also to play an important role to classification.For the signal x of acoustical vibration class, select Q1=8's
The small echo of octave has shown that collected vibration signal rarefaction representation.It is thin that it nearly corresponds to Meier order frequencies
Point.
In second stage, selects Q2=1 to define the small echo of smaller time support, more suitable for characterization transient state and chase after
Track.Recover the small echo that the high-frequency information of signal may need to have better frequency resolution ratio, therefore Q2 > 1.In higher order
M >=3 are respectively provided with Qm=1, but these coefficients are often ignored.
Scattering cascade more recovers the high-frequency information of signal with aiming at, wherein comprising constant Q filter group
Cascade, is followed by non-linear.First filter group of Q1=8 has obtained the characteristic for being equivalent to Meier spectrum, and second filters
Device group corresponds to the subsequent processing in the model of the filter with Q2=1.
Further, there is multiple classification device for analyzing characteristic value data in the step 5 Classification and Identification, identify sample
Data correspond to different state of runtime machine；
Wherein, each model algorithm comparison is shown in Table 1,
Shallow-layer algorithm is divided into:
SVM: support vector machines, effect is fine, and has to unknown data extensive Generalization Ability well；
ELM: training is quickly, as a result also relatively high, still, relatively high to system resources in computation requirement, meanwhile, as a result less
Stablize；
Deep learning algorithm:
SAE: unsupervised learning, effect is fine, but not as good as SIF feature is directly used, training is than very fast；
DNN: the effect is relatively poor, it may be possible to which data sample is fewer, and model training is not complete enough.
Each model algorithm of table 1 compares (10-fold CV)
After comparing, use SVM as model classifiers, algorithm comparison is mature, and has very strong adaptability herein,
Effect is fine, has to unknown data extensive Generalization Ability well.New collecting sample data are obtained by processing in use
To the characteristic quantity of 140 dimensions, it is sent into trained support vector machines, can be obtained the result of classification.
10 groups of data have been resurveyed, has belonged to 6 kinds of different operating conditions, the results are shown in Table 2 by test of heuristics:
Result of the 26 kinds of different operating conditions of table after test of heuristics
Wherein, accuracy rate: the sample number/total number of samples correctly classified；
TP: concrete class 1, prediction classification are 1；FP: concrete class 1, prediction classification are 0；
FN: concrete class 0, prediction classification are 1；TN: concrete class 0, prediction classification are 0.
It can be seen from the above result that SVM has higher classification accuracy, and also have very to unknown remaining normal data
Good popularization generalization ability.Wherein, real class (TP) and very negative class (TF) are larger, the correct data sample number of presentation class, false
Negative class (FP) is larger, indicates that some failures are predicted to be normal data, this should be reduced as possible, and reason may be due to event
The not comprehensive enough multiplicity of the data sample of barrier, prevents fault data from being correctly predicted.Meanwhile very negative class (FP) is smaller, this is
It is desired, show that model has good Forecasting recognition ability to normal data.
Table 3 is 10 groups of experimental results in trained support vector machines, can automatic Classification and Identification routine interference environment
Under, i.e., in the preferable situation of signal-to-noise ratio, detectivity at least guarantees 92% or more, and false alarm rate is 8% hereinafter, and being in adverse circumstances
Have under electromagnetic interference and strong vibration experiment condition, further decreases detectivity and false alarm rate index.
3 10 groups of the table experimental results in trained support vector machines
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited to above-mentioned
Particular implementation, those skilled in the art can make various deformations or amendments within the scope of the claims, this not shadow
Ring substantive content of the invention.
Claims (10)
1. the self study recognition methods of the working condition based on rotating machinery, which comprises the following steps:
Step 1, data preparation: acquiring the floor data of monitoring device, and the floor data of acquisition is intercepted by the duration of 2s or 3s
Original sample as processing；
Data prediction: step 2 carries out data using mean value and standard deviation of the z-score method for normalizing to initial data
Standardization；Treated data fit standardized normal distribution, mean value 0, standard deviation 1 convert function are as follows:
Wherein, x is expressed as original signal in formula, and μ is expressed as the mean value of initial data, and σ is expressed as the standard deviation of initial data,
X* is expressed as the signal after normalization；
Step 3, feature extraction: the parameter for obtaining 6 grades of interative computations using Deep Scattering Spectrum algorithm is special
Sign connects six groups of parameter attributes to obtain primitive character, and dimension is 7347 dimensions, to remove the time variation of signal, and has been obtained
Whole signal；
Feature Dimension Reduction: step 4 first carries out dimensionality reduction to primitive character using t-test method, by the feature of conspicuousness p < 0.05
It chooses, dimension is reduced to 3849 dimensions, then uses Feature Generating Machine algorithm dimensionality reduction again, obtains 140
The new feature value of dimension；
Classification and Identification: step 5 the characteristic value after dimensionality reduction is sent into support vector machines, Training Support Vector Machines are classified
Identification, finally obtains the algorithm of Classification and Identification.
2. the self study recognition methods of the working condition according to claim 1 based on rotating machinery, which is characterized in that institute
The data that floor data includes 6 kinds of operating conditions are stated, the data of respectively 3 kinds nominal situations: starting, tick over, low speed operation
The data of bringing onto load and 3 kinds of unusual service conditions: high speed bringing onto load, rotor unbalance, bearing fault and lubrication trouble.
3. the self study recognition methods of the working condition according to claim 1 based on rotating machinery, which is characterized in that institute
State the characteristic extraction procedure in step 3 the following steps are included:
(1) vibration signal of equipment running process is subjected to wavelet transformation, modulus and time steady three using constant Q filter group
Step handles to obtain equivalent Meier frequency spectrum, to obtain the short-term stationarity characteristic of signal；
(2) high-frequency information compressed in the Meier frequency spectrum data for restoring equivalent by continuing to do wavelet transformation.
4. the self study recognition methods of the working condition according to claim 3 based on rotating machinery, which is characterized in that institute
State step (1) obtain equivalent Meier frequency spectrum calculating process the following steps are included:
A) the small echo ψ for the use of octave resolution ratio being Q1 to sample of signal xλ1Do wavelet transformation, wherein the wavelet transformation uses
Basic function be one group of Q value be Q1 bandpass filter；
B) modular arithmetic is carried out to the output of the wavelet transformation, replaces square operation with First-Order Mode " | | " during modular arithmetic,
Formula (1) is substituted by formula (2),
Formula (1) are as follows:
Mx (t, λ) ≈ ∫ | ∫ x (u) ψλ(υ-u)du|2|φ(υ-t)|2dυ
=| x* ψλ|2*|φ|2(t)
Formula (2) are as follows:
Mx (t, λ) ≈ ∫ | ∫ x (u) ψλ(υ-u)du||φ(υ-t)|dυ
=| x* ψλ|*|φ|(t)；
C) result of the modular arithmetic is equalized using the low-pass filter φ (t) that size is T, it is ensured that the office with time migration
Portion's invariance obtains equivalent Meier frequency spectrum, equivalent Meier spectral coefficient, i.e. single order scattering coefficient are as follows:
Wherein, the x is expressed as the sample of original signal, and λ is expressed as the centre frequency of bandpass filter, and t is expressed as time, ψ
It is the base of Wavelet transformation, basic function is one group of bandpass filter.
5. the self study recognition methods of the working condition according to claim 3 based on rotating machinery, which is characterized in that institute
State continue to do wavelet transformation in step (2) the following steps are included:
A) obtained single order scattering coefficient is done into second of wavelet transformation, and modulus and carries out obtaining second order dispersion after the time is steady
Coefficient, formula are as follows:
B) processing for repeating step a) for several times, calculates the cascade of m wavelet convolution and modulus, obtains m rank scattering coefficient, formula is such as
Under:
6. the self study recognition methods of the working condition according to claim 5 based on rotating machinery, which is characterized in that institute
Processing 6 times that step a) is repeated in step b) are stated, the m is equal to 6.
7. the working condition recognition methods of the rotating machinery according to claim 5 based on machine learning, which is characterized in that
The scattering coefficient of every single order is that the First-Order Mode based on upper single order is calculated, and the public affairs of modular arithmetic are done to the result of wavelet transformation
Formula is as follows:
| W | and x=(x* φ (t), | x* ψλ(t)|)T ∈ R, λ ∈ Λ。
8. the self study recognition methods of the working condition according to claim 3 based on rotating machinery, which is characterized in that institute
It states wavelet transformation and meets the following conditions:
One, it is the constant bandpass filter of Q value；
Two,
Three, to parse small echo, and the centre frequency for parsing small echo normalizes to 1.
9. the working condition recognition methods of the rotating machinery according to claim 1 based on machine learning, which is characterized in that
Feature Generating Machine algorithm is divided into the step 4: worst case analysis and sub- problem optimize two steps,
Characteristic, penalty factor, loss function and the number of iterations in algorithm are freely arranged.
10. the working condition recognition methods of the rotating machinery according to claim 1 based on machine learning, feature exist
In sending the characteristic value of training data as the input data of algorithm into support vector machines as classifier in the step 5
Afterwards, algorithm model is obtained using the training of 10-fold cross validation method, and by continuously adjusting parameter optimization, inputted again
After new data, support vector machines can automatic Classification and Identification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910055613.8A CN109784284A (en) | 2019-01-22 | 2019-01-22 | The self study recognition methods of working condition based on rotating machinery |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910055613.8A CN109784284A (en) | 2019-01-22 | 2019-01-22 | The self study recognition methods of working condition based on rotating machinery |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109784284A true CN109784284A (en) | 2019-05-21 |
Family
ID=66501839
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910055613.8A Pending CN109784284A (en) | 2019-01-22 | 2019-01-22 | The self study recognition methods of working condition based on rotating machinery |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109784284A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110119789A (en) * | 2019-05-28 | 2019-08-13 | 湃方科技(北京)有限责任公司 | A kind of rotary-type mechanical equipment start and stop judgment method and system |
| CN110134571A (en) * | 2019-05-22 | 2019-08-16 | 湃方科技(北京)有限责任公司 | Rotary-type mechanical equipment health status monitoring method and device |
| CN111222442A (en) * | 2019-12-31 | 2020-06-02 | 中国电子科技集团公司第三十六研究所 | Electromagnetic signal classification method and device |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102519578A (en) * | 2011-12-15 | 2012-06-27 | 广东石油化工学院 | Method for extracting time-frequency domain spectrum of mixed signals of rotating machinery |
| CN103048041A (en) * | 2012-12-20 | 2013-04-17 | 北京信息科技大学 | Fault diagnosis method of electromechanical system based on local tangent space and support vector machine |
| CN104504403A (en) * | 2015-01-16 | 2015-04-08 | 重庆大学 | Rotating machinery failure forecasting method based on scattering conversion |
| US20150317990A1 (en) * | 2014-05-02 | 2015-11-05 | International Business Machines Corporation | Deep scattering spectrum in acoustic modeling for speech recognition |
| CN106844826A (en) * | 2016-12-02 | 2017-06-13 | 上海电机学院 | A kind of method for the diagnosis of gearbox of wind turbine failure predication |
| CN106951669A (en) * | 2017-05-19 | 2017-07-14 | 北京航空航天大学 | A kind of cognitive rolling bearing variable working condition method for diagnosing faults of view-based access control model |
| CN107478418A (en) * | 2017-06-29 | 2017-12-15 | 南京航空航天大学 | A kind of rotating machinery fault characteristic automatic extraction method |
| CN107505133A (en) * | 2017-08-10 | 2017-12-22 | 滁州学院 | The probability intelligent diagnosing method of rolling bearing fault based on adaptive M RVM |
-
2019
- 2019-01-22 CN CN201910055613.8A patent/CN109784284A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102519578A (en) * | 2011-12-15 | 2012-06-27 | 广东石油化工学院 | Method for extracting time-frequency domain spectrum of mixed signals of rotating machinery |
| CN103048041A (en) * | 2012-12-20 | 2013-04-17 | 北京信息科技大学 | Fault diagnosis method of electromechanical system based on local tangent space and support vector machine |
| US20150317990A1 (en) * | 2014-05-02 | 2015-11-05 | International Business Machines Corporation | Deep scattering spectrum in acoustic modeling for speech recognition |
| CN104504403A (en) * | 2015-01-16 | 2015-04-08 | 重庆大学 | Rotating machinery failure forecasting method based on scattering conversion |
| CN106844826A (en) * | 2016-12-02 | 2017-06-13 | 上海电机学院 | A kind of method for the diagnosis of gearbox of wind turbine failure predication |
| CN106951669A (en) * | 2017-05-19 | 2017-07-14 | 北京航空航天大学 | A kind of cognitive rolling bearing variable working condition method for diagnosing faults of view-based access control model |
| CN107478418A (en) * | 2017-06-29 | 2017-12-15 | 南京航空航天大学 | A kind of rotating machinery fault characteristic automatic extraction method |
| CN107505133A (en) * | 2017-08-10 | 2017-12-22 | 滁州学院 | The probability intelligent diagnosing method of rolling bearing fault based on adaptive M RVM |
Non-Patent Citations (1)
| Title |
|---|
| J. ANDÉN AND S. MALLAT: "Deep Scattering Spectrum", 《IEEE TRANSACTIONS ON SIGNAL PROCESSING》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110134571A (en) * | 2019-05-22 | 2019-08-16 | 湃方科技(北京)有限责任公司 | Rotary-type mechanical equipment health status monitoring method and device |
| CN110119789A (en) * | 2019-05-28 | 2019-08-13 | 湃方科技(北京)有限责任公司 | A kind of rotary-type mechanical equipment start and stop judgment method and system |
| CN111222442A (en) * | 2019-12-31 | 2020-06-02 | 中国电子科技集团公司第三十六研究所 | Electromagnetic signal classification method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108241873B (en) | A kind of intelligent failure diagnosis method towards pumping plant main equipment | |
| Youssef et al. | Disturbance classification utilizing dynamic time warping classifier | |
| CN109784284A (en) | The self study recognition methods of working condition based on rotating machinery | |
| CN111238843B (en) | Fan health evaluation method based on rapid spectrum kurtosis analysis | |
| US6574613B1 (en) | System and method for diagnosing jet engine conditions | |
| CN113962259B (en) | Multi-mode double-layer fault diagnosis method for fuel cell system | |
| CN111681132B (en) | Typical power consumption mode extraction method suitable for massive class unbalanced load data | |
| CN112052109B (en) | Cloud service platform event anomaly detection method based on log analysis | |
| CN111189638B (en) | HMM and QPSO optimization algorithm-based bearing fault degree identification method | |
| CN110701087A (en) | Axial flow compressor pneumatic instability detection method based on single-classification overrun learning machine | |
| Sivanantham et al. | Comparing the performance of adaptive boosted classifiers in anomaly based intrusion detection system for networks | |
| CN111783696A (en) | Edge calculation method for low-voltage branch topology real-time analysis based on PV relation | |
| De Lima et al. | On fault classification in rotating machines using fourier domain features and neural networks | |
| CN111458145A (en) | Cable car rolling bearing fault diagnosis method based on road map characteristics | |
| Huang et al. | Research on fan vibration fault diagnosis based on image recognition | |
| CN116625654A (en) | Vibration signal fault diagnosis method, device, equipment and storage medium | |
| CN106709598B (en) | Voltage stability prediction and judgment method based on single-class samples | |
| Tamura et al. | Deep learning based on fine tuning with application to the reliability assessment of similar open source software | |
| CN113639985B (en) | Mechanical fault diagnosis and state monitoring method based on optimized fault characteristic frequency spectrum | |
| CN110631827B (en) | Gearbox fault diagnosis method based on improved collaborative representation classification | |
| Wang et al. | Fisher discriminative dictionary learning for vehicle classification in acoustic sensor networks | |
| CN113505817A (en) | Self-adaptive weighting training method for bearing fault diagnosis model samples under unbalanced data | |
| CN112461546A (en) | Construction method and diagnosis method of pump bearing fault diagnosis model based on improved binary tree support vector machine | |
| Xiang et al. | An Intelligent Fault Diagnosis Method of Rolling Bearing with Wide Convolution Kernel Network | |
| CN114637793B (en) | Equipment fault frequent region positioning method based on big data analysis |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |