CN107798283A - A kind of neural network failure multi classifier based on the acyclic figure of decision-directed - Google Patents
A kind of neural network failure multi classifier based on the acyclic figure of decision-directed Download PDFInfo
- Publication number
- CN107798283A CN107798283A CN201610819565.1A CN201610819565A CN107798283A CN 107798283 A CN107798283 A CN 107798283A CN 201610819565 A CN201610819565 A CN 201610819565A CN 107798283 A CN107798283 A CN 107798283A
- Authority
- CN
- China
- Prior art keywords
- feature
- analysis
- frequency
- acyclic
- kinds
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
The present invention relates to a kind of neural network failure multi classifier based on the acyclic figure of decision-directed.The grader is divided into characteristic extraction part and identification classified part, on-line operation status monitoring and mechanical fault diagnosis available for high voltage electric equipments such as gas insulated metal enclosed swit chgears.Characteristic extraction part uses three kinds of different feature analysis als (i.e. industrial frequency harmonic structure characteristic analysis algorithm, third-octave spectrum sigtral response algorithm and cepstrum feature analysis al), respectively from the minutia and gross feature of voice signal frequency spectrum, and some three aspects of hiding information feature carry out individual features extraction, comprehensively describe the characteristic information of high-voltage electrical equipment operation voice signal.Meanwhile identify that classified part uses neural network classifier --- a K class grader is decomposed into by k two classification device based on the acyclic figure method of decision-directed, while it is trained and optimized based on three kinds of feature analysis al comprehensive analysis.
Description
Technical field：
The present invention proposes a kind of neural network failure multi classifier based on the acyclic figure method of decision-directed.This method is based on
Various features extraction algorithm comprehensive analysis carries out nicety of grading to each root node of the acyclic figure neural network classifier of decision-directed
Optimization, to improve classifier performance.This method can be used for caused by the operation of the high voltage electric equipments such as gas insulated metal enclosed swit chgear
Voice signal carries out feature extraction, and carries out mechanical breakdown identification automatically.
Background technology：
Closed gas combined electrical apparatus (Gas Insulated Switchgear, GIS) is due to high reliable
Property is used more and more widely.GIS has good insulating properties, is widely used in high voltage power transmisson system.With
China's power industry develops rapidly, and GIS tends to Large Copacity, high voltage development.As the visual plant run in power system,
GIS once breaks down, it will influences the normal power supply of power system, causes huge economic loss and bad social shadow
Ring.Therefore, GIS operational reliability is extremely important.
The GIS device of domestic early stage puts into operation in late 1980s and the beginning of the nineties more, has transported at present
Go nearly 20 years, arrived the stage occurred frequently of accident, had constantly, just during declaration of an accident of the recent years on GIS
It is explainable.Therefore method, the developing direction of accurate judgement defect type and defect and the serious journey of GIS Insulation monitorings are furtherd investigate
Degree, it is of great significance for ensureing GIS safe and reliable operation and instructing GIS maintenance undoubtedly to have.
With carrying out in a deep going way for repair based on condition of component work, the work of the state-detection such as live detection, on-line monitoring has become electricity
The evaluation of power equipment state, the important support means work of hidden troubles removing, and it is often main currently for GIS state-detection means
The insulating properties defect of GIS bodies and the mechanical defect of mechanism are diagnosed, and the mechanical defect of GIS bodies can not also detect at present.
In GIS actual motion, in addition to discharging fault, mechanical faults are also that one caused the accident is big
Main cause.So-called mechanical breakdown refers to when some defects in GIS be present, as switch contact contact is abnormal, housing docking
When imbalance, loosened fastener etc., discharging fault will not now occur, but due to alternation electricity caused by alternating current in conductor
Electromagnetic force etc. caused by power, transformer iron core can caused by GIS produce mechanicalness motion, due to the presence of mechanical defect, its
Abnormal transient vibration signal can be produced outside normal vibration.The abnormal vibrations of GIS bodies have very big harm to GIS bodies, for a long time
Vibration may make bolt looseness, cause gas leakage, and pressure reduces, and causes insulation fault；Insulator and insulated column can be caused
Infringement；The firm of earthing of casing point can be influenceed.
Discharging fault is concentrated mainly on for GIS device fault diagnosis both at home and abroad at present, is diagnosed for mechanical faults
Rarely has research.It is even more so far there are no research papers to carry out technology for mechanical fault diagnosis based on GIS device noise, and the present invention carries
Go out the present invention and propose a kind of neural network failure multi classifier based on the acyclic figure method of decision-directed.This method can be used for leading to
Failure caused by crossing the voice signal identification equipment mechanical defect sent during detection GIS equipment operations, this method can pass through analysis
Several different characteristic vectors of audible signal caused by GIS device operation, realize and carry out non-intruding, not damaged to GIS device
Formula fault detect and early warning.
The content of the invention：
The voice signal feature sent during it is an object of the invention to be run for GIS device chooses three kinds of different sound
Signature analysis algorithm carries out feature extraction to its voice signal respectively, and is based on the acyclic figure neural network classification of decision-directed
Device carries out Fault Identification.Caused voice signal is after wavelet algorithm filtering environmental noise under GIS equipment operational condition, respectively
Signature analysis, extraction are carried out by three kinds of industrial frequency harmonic feature, third-octave energy spectrum and cepstral coefficients feature algorithms.
The present invention is designed using the acyclic figure of decision-directed (Decision Directed Acyclic Graph, DDAG) method
Neutral net multi classifier.As shown in figure 12, k class classification problems are divided into multiple two classes subclassification machines by this method, that is, are schemed
In each root node.The error in classification of the subclassification machine of each root node can slowly add up from top to bottom in figure, so that influenceing
The performance of whole multi-categorizer.The present invention is entered to the subclassification machine at each root node respectively based on three kinds of different characteristic vectors
Row training and cross validation, the characteristic vector design subclassification machine with highest accuracy rate is chosen, to ensure whole multicategory classification
Device performance is optimized.
The present invention is that technical scheme is used by solving its technical problem：
It is proposed a kind of DDAG neutral net multiclass that comprehensive analysis optimization is carried out based on three kinds of different feature extraction algorithms
Grader method for diagnosing faults.This method includes the design and optimization two of feature extraction algorithm and DDAG neural network classifiers
Point.Selected by the voice signal feature that three kinds of feature extraction algorithms of the present invention are sent when being and being run based on GIS device
, concrete analysis is as follows：
Three kinds of feature analysis als used by this method --- industrial frequency harmonic structure characteristic analysis algorithm, third-octave
Feature analysis al and cepstrum feature analysis al, it is respectively provided with respective advantage and specific aim.Closed gas combined electrical apparatus
Belong to Large-scale High-Pressure transmission facility, its mechanical breakdown common when running has case vibration caused by debris, electromagnetic force, mangneto to stretch
Vibrated caused by the caused vibration of contracting, switching manipulation, vibrated caused by GIS contacts overheat causes loose contact, be interior
Vibrated caused by portion's element ground connection, abnormal sound vibration, the failure such as vibration caused by grid switching operation accident caused by outside noise.
GIS case vibrations signal is electricapparatus vibration signal, and frequency range is essentially 10~2000Hz, and amplitude is between 0.5~50.GIS
Acoustic signals frequency spectrum caused by equipment operation is wider, and due to sound wave caused by operation under malfunction caused by mechanical defect
Signal energy focuses primarily upon low frequency part (0~3000Hz or so).
Three kinds of feature analysis als of the present invention can independently be selected needed for extract signal frequency range.Work
Frequency harmonic characteristic parser is mainly used in characterizing in high-pressure machinery electric equipment operation audible signal by 50Hz power frequency electric signals
Caused harmonic components Energy distribution architectural feature and harmonic components and the feature of nonharmonic component energy ratio.Third-octave
Spectrum analysis is to be used to analyze the common method that plant equipment produces the operation sound in engineering, and octave spectrum can be easily
From macroscopically signal Analysis spectral characteristic, and good real-time can be reached by FFT.For GIS device etc.
Such plant equipment with labyrinth, often it is difficult to find that some features are believed using the simple Spectral Analysis Method of above two
Breath, and cepstrum has stronger recognition capability compared to lower.To acoustic signals extraction scramble spectrum signature caused by GIS device operation
The characteristic information of signal can be more fully described, prevents from producing missing inspection to the characteristic information of some fault types.
The design of DDAG neutral nets multi classifier and optimization method concrete analysis of the present invention are as follows：
Neural network classification algorithm has very strong nonlinear fitting ability, can map arbitrarily complicated non-linear relation,
And with very strong robustness, memory capability and powerful self-learning capability etc..Especially for having under large sample training condition
Unique advantage.Therefore, if the fault diagnosis system designed by inventive algorithm is applied to GIS device for a long time in real time exist
Line monitors, then can obtain a large amount of GIS devices caused acoustic signals sample under different faults type.This sets for increase GIS
Acoustic characteristic database caused by standby various fault types and improve neural network failure classification accuracy have considerable advantage and
Actual meaning.Be advantageous to establish sound difference caused by operation under GIS device normal condition and various fault type situations
Property data base and the fault diagnosis confidence level for being advantageous to greatly improve neutral net.
As shown in figure 12, neutral net multi classifier can be decomposed into one group by the present invention using the acyclic figure method of decision-directed
Two classification device, each sub-classifier are served only for distinguishing two class samples.After a sample to be sorted enters grader, pass through
The Classification and Identification of one group of sub-classifier, the sample are eventually divided into certain one kind.As shown in figure 12, classify for three classes
Device, it is only necessary to can draw classification results by two sub-classifiers.For K classification problems, for a kind of feature analysis al
The characteristic vector of extraction, need to train (k-1) k/2 two class subclassification machines altogether.Although the quantity of classifier is with sample type
Number increased dramatically, but be greatly reduced for the training time of the neutral net subclassification machine of two classification problem, and
Its classification accuracy also improves a lot with respect to multi classifier.And DDAG methods are based on, for K class classification problems, each
Sample to be sorted, which need to only pass through K-1 sub- classifiers, can draw classification results, therefore its speed of decision is also considerable
's.
To sum up learn, for K class graders, it is decomposed into (k-1) k/2 two class subclassifications as shown in figure 12 first
Machine, then choose corresponding sample for each sub-classifier and be trained and test.
Concrete technical scheme of the present invention is as follows：
1. feature extraction algorithm implementation
(1) industrial frequency harmonic architectural feature extraction algorithm
Present invention is generally directed to the diagnosis of potential faults caused by GIS device mechanical defect, it is investigated that related domestic and foreign literature can
Know, GIS device run under mechanical breakdown state caused by vibration signal show low-frequency range (0~3000Hz or so), so
And for the failure modes of small sample, feature vector dimension is too high to influence nicety of grading on the contrary.Therefore the main extraction of the present invention
GIS device runs the harmonic structure feature of sound low-frequency band, the frequency range of extraction required for user can independently select.Such as figure
1_A show industrial frequency harmonic architectural feature extraction algorithm flow chart.Exemplified by extracting voice signal 2kHz harmonic structure features,
The technical solution adopted in the present invention is as follows：
Acoustic signals are intercepted using the rectangular window function of fundamental frequency 50Hz signal integral multiple n length, if sample rate is fs, then n
It is taken as meeting the smallest positive integral value such as lower inequality.
LPF is carried out to above-mentioned windowing signal, cut-off frequecy of passband f1 is slightly larger than 2kHz, then upper by m1 times
Sampling and m2 times of down-sampling, it is f to obtain sample rates′.Accurately to carry out harmonic characteristic extraction, f must be causeds'=L=2*
F1, i.e. spectral resolution are just 1Hz；So m1 and m2 meet below equation：
M1, m2 are positive integer
Adding window is carried out to signal using rectangular window function, there is maximum main lobe, therefore video stretching effect is minimum；It is but right
Signal interception is carried out in aperiodic integral multiple, signal corresponding to window function head and the tail has compared with macromutation, therefore can produce serious
Spectral leakage problem.For solve the problems, such as technical scheme that rectangular window spectral leakage uses for：Choose a length of power frequency letter of rectangular window window
The integral multiple length in number cycle, the signal for being slightly longer than window length is then chosen, then cause window function to enter line slip by the signal beginning
Choose；When the amplitude square of signal value corresponding to rectangular window head and the tail and for it is minimum when, that is, represent that the mutation of this segment signal is minimum, and
And now adding window is periodically blocked equivalent to signal progress, is had minimum spectral leakage, is illustrated in figure 2 for same
Signal, compared with using rectangular window periodically block and blocking obtained spectrum leakage effect with aperiodicity.It is it is clear that non-
Serious spectrum leakage problem can be caused by periodically blocking, and increased window length in addition and caused it to be much larger than the signal period, enable to
Main lobe narrows, and secondary lobe diminishes, and energy is concentrated, and resolution ratio improves.
40 dimension industrial frequency harmonic characteristic vector within 2kHz is obtained based on the above method, is then added by power frequency within 2kH
Harmonic wave gross energy is used for wherein one-dimensional characteristic with anharmonic wave gross energy, and 41 Wei Gongpinxiebote are obtained after being normalized
Sign vector.
(2) third-octave energy spectrum feature extraction
As Fig. 1 _ B show 1/3 octave component energy spectrum feature extraction FB(flow block).Third-octave frequency spectrum is substantially by one group
The frequency spectrum obtained after the filtering of constant percentage bandwidth filter, it is that the poverty alleviation of constant percentage bandwidth filter is special as shown in Figure 3
Linearity curve.Wherein, f0 is filter section centre frequency, and it is lower-cut-off frequency f1 and the upper limit that amplitude, which declines corresponding to 3dB, at f0
Cut-off frequency f2, filter bandwidht B=f2-f1.
The ratio between two neighboring filtering window centre frequency is：(f0)i+1/(f0)i=2n, during n=1/3, as 1/3 octave component
Band filter；And two neighboring wave filter is end to end, (f1)i+1=(f2)i, therefore have：
The 1/3 octave band frequency meter (Hz) of table 1
Such as filtering one by one is carried out using each bandpass filter and obtain 1/3 octave component frequency spectrum, then realize less efficient.Engineering
On, the third-octave frequency spectrum of voice signal can be obtained first with Fast Fourier Transform (FFT).In noise testing, using 1/3 times
Sound interval spectrum analysis can reflect the spectral characteristic of noise source in detail.
The mechanical noise frequency distribution that GIS device is sent when running is broad, in the spectrum analysis of reality, it is not necessary to
Each frequency content can not possibly be made a concrete analysis of, for the ease of caused mechanical sounds under observation GIS equipment operational condition
The Energy distribution of signal macroscopically, ignore signal frequency or phase information minor variations to observing the influence of result, 20~
20kHz audio range is divided into several paragraphs, and each paragraph is referred to as frequency band or sound interval, its centre frequency and bound cutoff frequency
Rate meets above-mentioned relation, as table 1 lists voice signal 1/3 octave band frequency meter.
Its specific third-octave spectrum signature extracting method is as follows：
It is illustrated in figure 4 third-octave feature extraction flow chart.Backstage fault diagnosis platform receiving front-end is sent doubtful
Fault-signal, pretreatment stage is consistent with industrial frequency harmonic feature extraction algorithm, and calculates its power spectrum using FFT；By power spectrum
Decibel amplitude is converted in units of pa, and conversion formula is as follows：
Pa=20*10(-5+dB/20)
Finally, the frequency range of third-octave divides to power spectrum as shown in table 1, in 14.1Hz to 3600Hz
Form 24 frequency ranges, at the same will in each frequency range all power spectral density be added, it is average after, that is, obtain this section of voice signal
1*24 dimension third-octave spectrum signature vectors.
(3) cepstral coefficients feature extraction
As Fig. 1 _ C show cepstral coefficients feature extraction FB(flow block).For the machinery and power equipment of complexity, its machine
Tool structure harmony source forcing model is sufficiently complex, in fact, being even more such as mechanical sound mechanism caused by GIS device
This.GIS device not only has complicated mechanical structure, and it is accompanied by various sound source excitations under operation, and
Various electromagnetic coupled mechanism.Therefore, it is often difficult to recognize with general frequency spectrum analysis method, and uses cepstrum, then can strengthens
Identification capability.Cepstrum can provide the information that FFT spectrograms are difficult to catch, such as the cyclical component in FFT spectrums, system periphery ring
The interference in border and sideband signals, when side frequency is modulated by more races that being difficult to occurs in the spectrogram of mechanical fault signals, cepstrum
It can decompose and identify failure-frequency, analysis and tracing trouble Producing reason.Cepstrum is divided into power cepstrum and multiple scramble
Spectrum.Cepstral coefficients feature extraction algorithm of the present invention is to be based on power cepstrum, and its feature extracting method is as follows：
Power cepstrum is defined as the frequency spectrum of log power spectrum.If the unilateral power S of signal x (t)x(f), then：
cp(q)=| F { logSx(f)}|2 (1)
Wherein, q- inverted frequency, the big person of q values are high inverted frequency, rapid fluctuations and intensive harmonics on cepstrum are represented, conversely, q
It is low inverted frequency to be worth small person, represents slowly to fluctuate and sparse harmonics on cepstrum.
The suspected malfunctions signal that backstage centralized control center fault diagnosis platform receiving front-end is sent, i.e., be L Hammings using length
Window intercept signal, make L point FFT, calculate signal and compose S from power magnitudex(f), and take the logarithm；Then produced as inverse Fourier transform
To the cepstrum of signal.Nearly 20 range values are as signal cepstral coefficients characteristic vector before taking cepstrum.
2. tagsort
As described above, obtain each different dimensional of GIS device operation voice signal by three kinds of different feature extraction algorithms
Number characteristic vector.It is illustrated in figure 4 neural network classifier.If inputoutput pair (Xp, Tp), p=1,2 ..., P, wherein：P is
Number of training, XpFor p-th of sample input vector：Xp=(xp1..., xpM), M is input vector dimension；TpFor p-th of sample
This expectation output vector, Tp=(tp1..., tpN), N is output vector dimension, and the reality output vector of network is Yp=
(yp1..., ypN).Neutral net uses single hidden layer configuration, and the nodes of hidden layer are K, connection weight between input layer and hidden layer
Use αikRepresent, the connection weight β between hidden layer and output layerkjRepresent, αikRepresent i-th of node of input layer to k-th of hidden layer
Connection weight between node.The hidden layer activation primitive Ψ of neutral netkUsing Sigmoid type functions, Ψ (x)=1/ (1+e-x), error functionη is learning efficiency.
Three layers of BP neural network program realize that algorithm steps are as follows：
The output of hidden node：
Wherein, xiFor the input of i-th of input layer, hkFor the output of k-th of hidden node.
Export node layer yjFor：
Individualized training sample enters neutral net, and obtained overall error is：
Back-propagation algorithm：
Define downward gradient δj：
Therefore,
The BP neural network connection weight updated value that formula (2) (3) is gone out by back-propagation algorithmic derivation, right value update
Formula is：
Fig. 6 is fault diagnosis system frame diagram.The above-mentioned BP neural network grader as in Fig. 5 constitutes overall multiclass point
Two classification machine in class device (as shown in figure 12).
DDAG neutral net multi classifiers are that K class classification problems are resolved into (K-1) K/2 two class subclassification machines, and
The algorithm classified by acyclic figure is oriented to.It is illustrated in figure 7 DDAG algorithm flow charts.
(1) training stage：Subclassification machine Classifier { i, j } for distinguishing the i-th class and jth class, it is trained and excellent
It is as follows to change selection step：
(i) data sample of the i-th class and jth class is chosen, above-mentioned three kinds of feature analysis als is based respectively on and extracts identical number
Six groups of characteristic vector A of amountij, Bij, Cij；
(ii) cross-validation method is based on to every group of characteristic vector obtained above and trains and test neutral net subclassification machine.
Such as one group of characteristic vector A that the i-th class sample is obtained based on industrial frequency harmonic feature extraction algorithmijFirst it is divided into M
Group, i.e. Aij={ a1, a2..., aM}.A is utilized successivelymGroup as training sample training sub-classifier, remaining subcombinations into
Test sample tests the classification accuracy of the classifier；
(iii) it is based on cross-validation method in (ii) and draws the comprehensive of the neural network classification machine that each feature vectors train
Close discrimination PA, PB, PC, it is up to principle according to discrimination, selected characteristic vector trains subclassification machine Classifier { i, j }.
And utilize the subclassification machine on all root nodes of above method design alternative.
(2) decision phase：As shown in Fig. 6 and Figure 12, an input sample is given, is based respectively on above-mentioned three kinds to it first
Feature extraction algorithm extracts corresponding characteristic vector.Since root node, the required characteristic vector that is marked according to two classifiers
Type is selected characteristic vector and classified for subclassification machine, and the output valve of the two classification machine of the node determines that it walks left side or right side
Path, so until untill leaf node, obtain the classification belonging to input sample.
The advantages of DDAG methods is：Speed of decision is significantly than one kind to remaining class method (One Versus Rest, OVR) and one
Class is fast to a kind of method (One Versus One, OVO) method, and in the absence of can not subregion.Its shortcoming is：The selection of root node is straight
Connecing influences the result of this classification, and two different classifiers may be different as root node, its classification results, so as to produce classification
As a result uncertainty.
The characteristic vector pair that classification innovatory algorithm designed by the present invention is obtained based on three kinds of different feature extraction algorithms
Two classification machine on each root node carries out combined training, and chooses accuracy highest classifier.
Therefore, corresponding two class samples are belonged to for each root node classifier in the acyclic figure of decision-directed as shown in figure 12
This optimal classification machine, the nicety of grading of each root node can be reduced, so as to solve the critical defect of DDAG methods.
Brief description of the drawings：
Fig. 1 system features extract block diagram
Compared with Fig. 2 rectangular window cycles leak with aperiodic truncated spectrum
Fig. 3 constant percentage bandwidth filter amplitude-versus-frequency curves
Fig. 4 third-octave feature extraction flow charts
Fig. 5 neural network classifiers
Fig. 6 .DDAG neural network classifier fault diagnosis principle block diagrams
Fig. 7 are based on various features parser optimization DDAG root node sub-classifiers
The normalization one-sided power spectrum figure of the class sample datas (A, B, C) of Fig. 8 tri-
The normalization harmonic characteristic of the class sample datas (A, B.C) of Fig. 9 tri-
The normalization third-octave spectrum signature of the class sample datas (A, B.C) of Figure 10 tri-
The normalization scramble spectrum signature of the class sample datas (A, B.C) of Figure 11 tri-
The acyclic drawing method of Figure 12 decision-directeds
Embodiment：
(1) industrial frequency harmonic signature analysis --- according to signal sampling rate, it is 50Hz signals week to determine rectangular window length L1, L1
The integral multiple of phase.And searching is slided on the signal of selection using the rectangular window and make it that the corresponding signal location mutation of window head and the tail is minimum
Position intercepted, by the signal of interception after FIR low pass filter filters, first pass through m1 times and up-sample, then by m2
The signal length obtained after down-sampling again is N1, and N is equal to sample rate f s1 now；Then it is calculated by N1 points FFT
The auto-power spectrum of signal, signal power frequency within certain frequency band range and its amplitude on harmonic wave frequency are extracted on amplitude spectrum as humorous
Wave component characteristic vector, and be normalized, it is eventually adding the ratio between signal harmonic component gross energy and non-harmonic signals gross energy
Value tag amount, a 1*num industrial frequency harmonic structural eigenvector is obtained；
(2) third-octave energy Spectral characteristics analysis --- acoustic signals pretreatment mode (including adding window and variable sampling rate
Deng) with above-mentioned industrial frequency harmonic feature extraction algorithm.The signal after adding window and variable sampling rate is obtained to calculate by N2 points FFT
Be converted to pa (Pa) to its logarithm auto-power spectrum, and by power spectral amplitude ratio dB, then as shown in table 1 frequency range to power spectrum domain
Divided, choose num frequency range of formation within frequency band scope, while by all power spectrums in each frequency range
Degree is added, average to obtain 1*num dimensional feature vectors；
(3) cepstral coefficients signature analysis --- acoustic signals pretreatment mode (including adding window and variable sampling rate etc.) is ibid
State industrial frequency harmonic feature extraction algorithm.Obtaining the signal after adding window and variable sampling rate, by N3 points FFT its to be calculated right
Number is composed from power magnitude, is chosen unilateral log power amplitude spectrum progress IFFT and is transformed to cepstral domains.Finally, num before selection
Scramble spectral amplitude ratio is as cepstrum coefficient characteristic vector.
According to (1) (2) (3), that is, obtain three kinds of different characteristic vectors of system input signal.By three features
Vector is trained to each root node neutral net subclassification machine as shown in figure 12 respectively, obtains three seed classification machines, so
The synthesis identification probability for the grader that every kind of features training obtains is obtained based on modes such as cross validations afterwards, and chooses correct identification
Probability highest characteristic vector trains neural network classifier.
The DDAG neutral net multi classifiers after various features parser optimizes, the grader can to sum up be obtained
In each root node not only in store node subclassification machine data also have this node classifier used by feature extraction calculate
Method.As shown in fig. 6, when receiving certain sample to be sorted, the sample is first extracted into three kinds of different characteristic vectors respectively, so
Afterwards since being oriented to acyclic figure root node, corresponding characteristic vector is adaptively chosen according to the feature required for each root node and used
Classify in the node.Think that it should turn left or turn right for the classification output result obtained by this node to continue to classify, such as
This at leaf node until obtained the final recognition result of the sample signal.
Implementation result：
In order to show the inventive method classifying quality, the present invention is sent when being run using three kinds of different high-voltage electrical equipments
Voice signal as data sample, classified using this method, verify overall Classification and Identification probability.Choose as shown in figure 11
The normalized power spectrogram of voice data when three class High-Voltage Electrical Appliances are run.
The cross validation classification results of table 2
As shown in table 2, (taken for all synthesis discriminations for distinguishing grader two-by-two of the class classification problems of A/B/C tri-
Value) be respectively：A/B --- 91.453%, 98.718% and 99.145%；A/C --- 97.436%, 97.436% He
91.880%；B/C --- 97.863%, 50%, 96.154%.So as shown in figure 12, the selection of 1v3 root nodes is based on power frequency
The two classification machine of harmonic characteristic vector or third-octave characteristic vector；The selection of 2v3 root nodes is based on industrial frequency harmonic characteristic vector
Two classification device；1v2 root nodes select the two classification machine based on cepstrum characteristic vector.
Simulation result and analysis：
The total final discrimination of three class testing samples is 99.145%.Analysis is understood, based on the more of the acyclic figure of decision-directed
Sorting algorithm has very high feature recognition probability, while its computation complexity is linear with class number, improves mould
Formula recognition rate.In addition, the method used in the present invention based on various features analysis, is selected most preferably on each root node
The corresponding two classification machine of feature extracting method, greatly overcomes itself drawback of the acyclic nomography of decision-directed, improves
The stability of sorting algorithm.
Claims (3)
- A kind of 1. neural network failure multi classifier based on the acyclic figure of decision-directed.The grader can be used for closed composition The on-line operation status monitoring and mechanical fault diagnosis of the high voltage electric equipments such as electrical equipment.Method for diagnosing faults designed by the present invention By three kinds of different feature analysis als, (i.e. industrial frequency harmonic structure characteristic analysis algorithm, third-octave spectrum sigtral response are calculated Method and cepstrum feature analysis al) analyze the sound letter sent during the operation of the high voltage electric equipments such as gas insulated metal enclosed swit chgear Number, extract corresponding characteristic vector and be used to examine failure caused by the on-line monitoring and mechanical defect of the equipment running status It is disconnected.
- 2. feature is carried out respectively to high-voltage electrical equipment according to three kinds of acoustic feature signal extraction algorithms described in claim 1 Analysis.Three kinds of feature analysis al features are：(1) industrial frequency harmonic character extraction algorithm is mainly used in extracting voice signal Each harmonic energy-distributing feature caused by 50Hz power-frequency voltages and harmonic component gross energy and non-harmonic component total energy in frequency spectrum Ratio feature between amount.(2) third-octave spectrum signature extraction algorithm is to be used to analyze plant equipment running noises in engineering Principal character extraction algorithm.High-voltage electrical equipment typically runs that audio frequency distribution is broad, and third-octave spectrum signature is easy to The Energy distribution of mechanical movement voice signal macroscopically is observed, ignores signal frequency or phase information minor variations to observing result Influence.(3) cepstrum feature extraction algorithm can extract the information that general frequency spectrum analysis method is difficult to identify, such as periodically divide Amount, the interference of system periphery environment and sideband signals etc..In summary, three kinds of feature extraction algorithms are respectively from voice signal frequency spectrum Minutia and gross feature, and the aspect of some hiding information features three carries out individual features extractions, comprehensively describes The characteristic information of high-voltage electrical equipment operation voice signal.
- 3. according to the neutral net multi classifier based on the acyclic figure method of decision-directed described in claim 1 and 2, the present invention Each node subclassification machine in the acyclic figure of guiding is trained and optimized using three kinds of different feature analysis als.It optimizes Feature is：In the training stage, for each node subclassification machine, three feature vectors groups are extracted respectively, and test using intersection Card method trains to obtain three groups of sub-classifiers and its corresponding test discrimination.Then, the comprehensive son based on three feature vectors The discrimination of classifier, choose characteristic vector training subclassification machine optimal at the node.In sorting phase, category signal is treated Three feature vectors are extracted, then the characteristic vector since root node according to needed for each subclassification is classified, until arriving Acyclic figure leaf node is oriented to, that is, completes classification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610819565.1A CN107798283A (en) | 2016-08-31 | 2016-08-31 | A kind of neural network failure multi classifier based on the acyclic figure of decision-directed |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610819565.1A CN107798283A (en) | 2016-08-31 | 2016-08-31 | A kind of neural network failure multi classifier based on the acyclic figure of decision-directed |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107798283A true CN107798283A (en) | 2018-03-13 |
Family
ID=61529569
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610819565.1A Withdrawn CN107798283A (en) | 2016-08-31 | 2016-08-31 | A kind of neural network failure multi classifier based on the acyclic figure of decision-directed |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107798283A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109325475A (en) * | 2018-11-19 | 2019-02-12 | 国网河北省电力有限公司电力科学研究院 | Transformer vibration signal separation method and terminal device |
| CN110427835A (en) * | 2019-07-11 | 2019-11-08 | 清华大学 | The electromagnet signal recognition method and device of figure convolutional network and transfer learning |
| CN111141503A (en) * | 2019-12-20 | 2020-05-12 | 国网浙江海盐县供电有限公司 | Portable GIS isolating switch operating mechanism mechanical characteristic tester and testing method |
| CN111191723A (en) * | 2019-12-30 | 2020-05-22 | 创新奇智(北京)科技有限公司 | Few-sample commodity classification system and method based on cascade classifier |
| CN111401399A (en) * | 2019-12-24 | 2020-07-10 | 中国国家铁路集团有限公司 | Accident early warning and classifying method and device for railway freight |
| CN112082639A (en) * | 2019-06-14 | 2020-12-15 | 现代自动车株式会社 | Engine state diagnosis method and diagnosis modeling method thereof |
| CN113933590A (en) * | 2020-07-14 | 2022-01-14 | 森兰信息科技（上海）有限公司 | Method, system, medium, and apparatus for calculating frequency spectrum of wave |
| CN115290759A (en) * | 2022-08-05 | 2022-11-04 | 河北变检电力科技有限公司 | Intelligent porcelain insulator nondestructive testing analysis method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1584433A (en) * | 2004-06-04 | 2005-02-23 | 广东科龙电器股份有限公司 | Noise source identifying method for air-conditioner based on nervous network |
| CN101533068A (en) * | 2009-04-08 | 2009-09-16 | 南京航空航天大学 | Analog-circuit fault diagnosis method based on DAGSVC |
| CN103149029A (en) * | 2013-01-16 | 2013-06-12 | 南京航空航天大学 | Aero-engine rotor-stator rubbing position identification method based on application of cepstrum |
| CN103245524A (en) * | 2013-05-24 | 2013-08-14 | 南京大学 | Acoustic fault diagnosis method based on neural network |
| CN104697767A (en) * | 2014-12-17 | 2015-06-10 | 天津大学 | Rotor system fault diagnosis method and device based on vibration analysis |
| CN104751182A (en) * | 2015-04-02 | 2015-07-01 | 中国人民解放军空军工程大学 | DDAG-based SVM multi-class classification active learning algorithm |
-
2016
- 2016-08-31 CN CN201610819565.1A patent/CN107798283A/en not_active Withdrawn
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1584433A (en) * | 2004-06-04 | 2005-02-23 | 广东科龙电器股份有限公司 | Noise source identifying method for air-conditioner based on nervous network |
| CN101533068A (en) * | 2009-04-08 | 2009-09-16 | 南京航空航天大学 | Analog-circuit fault diagnosis method based on DAGSVC |
| CN103149029A (en) * | 2013-01-16 | 2013-06-12 | 南京航空航天大学 | Aero-engine rotor-stator rubbing position identification method based on application of cepstrum |
| CN103245524A (en) * | 2013-05-24 | 2013-08-14 | 南京大学 | Acoustic fault diagnosis method based on neural network |
| CN104697767A (en) * | 2014-12-17 | 2015-06-10 | 天津大学 | Rotor system fault diagnosis method and device based on vibration analysis |
| CN104751182A (en) * | 2015-04-02 | 2015-07-01 | 中国人民解放军空军工程大学 | DDAG-based SVM multi-class classification active learning algorithm |
Non-Patent Citations (4)
| Title |
|---|
| ANNA WANG ET AL.: "Researches on Soft Fault Diagnosis Algorithm of Analogy Circuits Based on DDAGSVMs", 《PROCEEDINGS OF THE 2007 IEEE INTERNATIONAL CONFERENCE ON INTEGRATION TECHNOLOGY》 * |
| 易辉 等: "基于相关向量机与决策导向无环图的故障定位方法", 《2013 25TH CHINESE CONTROL AND DECISION CONFERENCE (CCDC)》 * |
| 曹文明 等: "基于音频识别的无人值守变电站设备在线监测", 《湖南大学学报（自然科学版）》 * |
| 黄晶晶: "噪声分析系统在航空发动机试车台的应用研究", 《中国优秀博硕士学位论文全文数据库（硕士） 工程科技II辑》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109325475A (en) * | 2018-11-19 | 2019-02-12 | 国网河北省电力有限公司电力科学研究院 | Transformer vibration signal separation method and terminal device |
| CN112082639A (en) * | 2019-06-14 | 2020-12-15 | 现代自动车株式会社 | Engine state diagnosis method and diagnosis modeling method thereof |
| CN110427835A (en) * | 2019-07-11 | 2019-11-08 | 清华大学 | The electromagnet signal recognition method and device of figure convolutional network and transfer learning |
| CN110427835B (en) * | 2019-07-11 | 2022-03-11 | 清华大学 | Electromagnetic signal identification method and device for graph convolution network and transfer learning |
| CN111141503A (en) * | 2019-12-20 | 2020-05-12 | 国网浙江海盐县供电有限公司 | Portable GIS isolating switch operating mechanism mechanical characteristic tester and testing method |
| CN111401399A (en) * | 2019-12-24 | 2020-07-10 | 中国国家铁路集团有限公司 | Accident early warning and classifying method and device for railway freight |
| CN111401399B (en) * | 2019-12-24 | 2023-11-10 | 中国国家铁路集团有限公司 | Accident early warning and classifying method and device for railway freight |
| CN111191723A (en) * | 2019-12-30 | 2020-05-22 | 创新奇智(北京)科技有限公司 | Few-sample commodity classification system and method based on cascade classifier |
| CN111191723B (en) * | 2019-12-30 | 2023-06-20 | 创新奇智(北京)科技有限公司 | Cascade classifier-based few-sample commodity classification system and classification method |
| CN113933590A (en) * | 2020-07-14 | 2022-01-14 | 森兰信息科技（上海）有限公司 | Method, system, medium, and apparatus for calculating frequency spectrum of wave |
| CN115290759A (en) * | 2022-08-05 | 2022-11-04 | 河北变检电力科技有限公司 | Intelligent porcelain insulator nondestructive testing analysis method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107798283A (en) | A kind of neural network failure multi classifier based on the acyclic figure of decision-directed | |
| Zhu et al. | Wavelet-based fuzzy reasoning approach to power-quality disturbance recognition | |
| Gargoom et al. | Automatic classification and characterization of power quality events | |
| CN109033612A (en) | A kind of Diagnosis Method of Transformer Faults based on vibration noise and BP neural network | |
| CN105447502A (en) | Transient power disturbance identification method based on S conversion and improved SVM algorithm | |
| CN109633368A (en) | The method of duration power quality disturbances containing distributed power distribution network based on VMD and DFA | |
| CN103398769A (en) | Transformer on-line fault detecting method based on sampling integrated SVM (support vector machine) under wavelet GGD (general Gaussian distribution) feather and unbalanced K-mean value | |
| CN110044623A (en) | The rolling bearing fault intelligent identification Method of empirical mode decomposition residual signal feature | |
| Shi et al. | Online event detection in synchrophasor data with graph signal processing | |
| CN111308260B (en) | Electric energy quality monitoring and electric appliance fault analysis system based on wavelet neural network and working method thereof | |
| CN109444667A (en) | Power distribution network initial failure classification method and device based on convolutional neural networks | |
| CN104089699B (en) | A kind of substation equipment speech reconstruction algorithm | |
| Beg et al. | Cyber-physical anomaly detection in microgrids using time-frequency logic formalism | |
| Izadi et al. | A synchronized lissajous-based method to detect and classify events in synchro-waveform measurements in power distribution networks | |
| CN110516742A (en) | A kind of distribution terminal fault distinguishing method and system based on Combination neural network model | |
| CN110020637A (en) | A kind of analog circuit intermittent fault diagnostic method based on more granularities cascade forest | |
| CN108181059A (en) | Multiphase flow pipeline leakage acoustic signals recognition methods based on small echo signal | |
| Saikia et al. | Detection and classification of power quality disturbances using wavelet transform, fuzzy logic and neural network | |
| Jarrahi et al. | An EMD based fault type identification scheme in transmission line | |
| Christy et al. | Wavelet based detection of power quality disturbance-A case study | |
| CN103400021A (en) | KDSNE hydroelectric generating set noise source detection method based on heavy tailed distribution | |
| Kezunovic | Automated analysis of voltage sags, their causes and impacts | |
| Song et al. | Research on the fault diagnosis method of mine fan based on sound signal analysis | |
| Shicheng et al. | An effective S-transform feature extraction method for classification of power quality disturbance Signals | |
| CN113610020A (en) | Method for identifying transformer vibration reason based on FFT and wavelet packet algorithm |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20180313 |