CN110376497A  Lowvoltage distribution system series fault arc method of identification based on all phase deep learning  Google Patents
Lowvoltage distribution system series fault arc method of identification based on all phase deep learning Download PDFInfo
 Publication number
 CN110376497A CN110376497A CN201910739837.0A CN201910739837A CN110376497A CN 110376497 A CN110376497 A CN 110376497A CN 201910739837 A CN201910739837 A CN 201910739837A CN 110376497 A CN110376497 A CN 110376497A
 Authority
 CN
 China
 Prior art keywords
 phase
 deep learning
 arc
 low
 current
 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
 238000001228 spectrum Methods 0.000 claims abstract description 81
 238000003062 neural network model Methods 0.000 claims abstract description 12
 238000010891 electric arc Methods 0.000 claims description 43
 230000001537 neural Effects 0.000 claims description 24
 230000003595 spectral Effects 0.000 claims description 20
 230000000875 corresponding Effects 0.000 claims description 15
 238000004458 analytical method Methods 0.000 claims description 13
 238000005070 sampling Methods 0.000 claims description 11
 239000000284 extract Substances 0.000 claims description 10
 238000005259 measurement Methods 0.000 claims description 9
 230000005611 electricity Effects 0.000 claims description 8
 238000000605 extraction Methods 0.000 claims description 7
 238000004611 spectroscopical analysis Methods 0.000 claims description 7
 238000010183 spectrum analysis Methods 0.000 claims description 7
 230000001131 transforming Effects 0.000 claims description 5
 238000007477 logistic regression Methods 0.000 claims description 3
 230000004044 response Effects 0.000 claims description 3
 238000004364 calculation method Methods 0.000 claims description 2
 230000011218 segmentation Effects 0.000 claims 1
 230000000694 effects Effects 0.000 abstract description 4
 238000000034 method Methods 0.000 description 9
 230000015572 biosynthetic process Effects 0.000 description 3
 238000010586 diagram Methods 0.000 description 3
 238000005755 formation reaction Methods 0.000 description 3
 238000004422 calculation algorithm Methods 0.000 description 2
 230000002068 genetic Effects 0.000 description 2
 238000000691 measurement method Methods 0.000 description 2
 230000035515 penetration Effects 0.000 description 2
 244000020998 Acacia farnesiana Species 0.000 description 1
 235000010643 Leucaena leucocephala Nutrition 0.000 description 1
 210000004218 Nerve Net Anatomy 0.000 description 1
 125000004122 cyclic group Chemical group 0.000 description 1
 238000000354 decomposition reaction Methods 0.000 description 1
 230000000994 depressed Effects 0.000 description 1
 238000009795 derivation Methods 0.000 description 1
 238000001514 detection method Methods 0.000 description 1
 239000000428 dust Substances 0.000 description 1
 235000013399 edible fruits Nutrition 0.000 description 1
 238000005516 engineering process Methods 0.000 description 1
 238000009413 insulation Methods 0.000 description 1
 238000002955 isolation Methods 0.000 description 1
 230000004301 light adaptation Effects 0.000 description 1
 238000004519 manufacturing process Methods 0.000 description 1
 239000000203 mixture Substances 0.000 description 1
 230000004048 modification Effects 0.000 description 1
 238000006011 modification reaction Methods 0.000 description 1
 230000035772 mutation Effects 0.000 description 1
 230000001681 protective Effects 0.000 description 1
 230000000452 restraining Effects 0.000 description 1
 230000035945 sensitivity Effects 0.000 description 1
 238000004088 simulation Methods 0.000 description 1
 238000006467 substitution reaction Methods 0.000 description 1
Classifications

 G—PHYSICS
 G01—MEASURING; TESTING
 G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
 G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
 G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
Abstract
The invention discloses the lowvoltage distribution system series fault arc method of identifications based on all phase deep learning, it solves in existing lowvoltage network failure, recognition methods for series fault arc is interfered vulnerable to noise, spectrum leakage, influence recognition effect, recognition efficiency is not high, and the problem that stability is not high.The method of the present invention includes: to carry out current signal acquisition to the different loads in low tension loop under lowvoltage alternatingcurrent system；All phase discrete Fourier transform is carried out to collected current signal, all phase spectrum signature amount loaded is extracted, and constructs all phase spectrum signature vector；The deep learning neural network model returned based on Logistic is constructed, by carrying out deep learning training to all phase spectrum signature amount under different loads, different operating statuses, until model convergence；The examination to different loads type is completed using trained model and the identification of series fault arc whether occurs.
Description
Technical field
The present invention relates to lowvoltage distribution system series fault arc identification technology fields, and in particular to is based on all phase depth
The lowvoltage distribution system series fault arc method of identification of study.
Background technique
In lowvoltage distribution system, due to route or equipment insulation ag(e)ing, rupture or electrical missing contact when, be often associated with
Fault electric arc, and then cause fire hazard.Low voltage failure electric arc is divided into parallel arc fault and series fault arc, parallel failure
Electric arc is caused by short trouble occurs between distribution line, and short circuit current is larger, and protective device correctly can be identified and be acted.String
Connection fault electric arc is in contact bad or broken string by route and is caused, and fault curre is 5~30A, is unable to satisfy protection dress
The sensitivity requirements set.In practical lowvoltage network failure, most indiscernible is also series fault arc.
It to the detection method of series fault arc, can be divided into three classes at present: the emulation 1. based on Cassie Arc Modelling
Circuit；The physical features such as light, heat, the sound generated when 2. being occurred using electric arc are judged；3. analyzing electricity when serial arc occurs
Stream, voltage signal identify fault electric arc with mathematical method.The 3. for kind method due to wide adaptation range, high reliablity is main to collect
In Timefrequency Analysis and feature vector to electric current and voltage signal mathematical analysis etc..Wherein, analysis of timedomain characteristic is
Series fault electricity is identified using features such as electric current, the change rate of voltage waveform, the waveform slope value of mutation, " zero stops " times
Arc, the disadvantage is that the influence of ambient noise or the pollution of sampled data can all make time domain index generate biggish error, interference judgement
As a result.Frequency domain character analysis is spectrum signature, harmonic wave and the mAcetyl chlorophosphonazo content using signal on frequency domain to determine whether occurring
Series fault arc, the disadvantage is that the influence vulnerable to spectrum leakage, determining interference result.The Mathematical Method of feature vector includes
Autoregressive parameter model, fractal dimension and support vector machines, wavelet transformation and singular value decomposition etc., principle and testing process are more multiple
Miscellaneous, actually detected middle stability is not high, cannot preferably apply to engineering in practice.
Summary of the invention
The technical problems to be solved by the present invention are: in existing lowvoltage network failure, for series fault arc
Recognition methods is interfered vulnerable to noise, spectrum leakage, influences recognition effect, and recognition efficiency is not high, and the problem that stability is not high, this
Invention provides the lowvoltage distribution system series fault arc method of identification based on all phase deep learning to solve the above problems.
The present invention is achieved through the following technical solutions:
Lowvoltage distribution system series fault arc method of identification based on all phase deep learning, this method comprises:
Under lowvoltage alternatingcurrent system, current signal acquisition is carried out to the different loads in low tension loop；
All phase discrete Fourier transform, all phase spectrum signature amount loaded are carried out to collected current signal
It extracts, and constructs all phase spectrum signature vector；
Using all phase spectrum signature vector built up, the deep learning neural network mould returned based on Logistic is constructed
Type, by carrying out deep learning training to all phase spectrum signature amount under different loads, different operating statuses, until the model
Until convergence；
The deep learning neural network model returned based on Logistic completed using learning training, to be identified point
Section load current waveform carries out all phase FFT spectrum analysis measurement, completes the examination to different loads type and whether series fault occurs
The identification of electric arc is exported according to the modelEach load is pressed type, normal operation or failure by it
Conditions at the arc are numbered,AndWherein u ≠ v, u, v=0,1 ..., 47, indicate the type number for identifying load
For round (u+1)/2, operating status is normal arc, and u is even number；Or operating status is fault electric arc, u is odd number；Wherein
Round () expression, which takes, to round up.
Working principle is: in existing lowvoltage network failure, for series fault arc recognition methods vulnerable to noise,
Spectrum leakage interference, influences recognition effect, recognition efficiency is not high, and the problem that stability is not high, and the present invention is using the above scheme
Using all phase spectrum signature formation mechenism of load current signal as point of penetration, the spectrum leakage and frequency spectrum interference composed using all phase
Rejection extracts the characteristic quantity of distorted signal and normal signal on all phase frequency spectrum；Building is based on Logistic
The deep learning neural network model of recurrence, using the deep learning classifying identification method returned based on Logistic, to input
All phase spectrum signature differentiated, realize examination to load type, the electric arc that whether breaks down；Utilize neural network
Nonlinear, adaptivity and faulttolerant ability carry out intelligence to series fault arc, efficiently identify.The present invention is for series connection event
The recognition methods for hindering electric arc is not interfered vulnerable to noise, spectrum leakage, and recognition efficiency is high, and stability is strong.
Further, the lowvoltage alternatingcurrent system includes power supply unit, failure generating unit, data sampling analytical unit,
The power supply unit uses lowvoltage alternatingcurrent experiment power supply AC, the lowvoltage alternatingcurrent experiment power supply AC to become by switch K1 connection isolation
Primary coil one end of depressor, the primary coil other end ground connection of isolating transformer, secondary coil one end of isolating transformer connects
Arc generator is connect, arc generator connects resistance R1, resistance R1 connection slide rheostat R_{p}, rheostat R_{p}Connected by switch K2
Connect the secondary coil other end of isolating transformer；The arc generator is used to generate stable series fault arc, and described
Switch K3 is parallel on arc generator；Data acquisition unit DA, the data acquisition unit DA are connected on the resistance R1
It connects computer PC and realizes analysis；Wherein, slide rheostat R_{p}, resistance R_{1}Respectively indicate load and sampling resistor.
Further, described that all phase discrete Fourier transform is carried out to collected current signal, wherein final full phase
The frequency offset formula of position discrete Fourier transform are as follows:
In formula: G (k_{i})、Y(k_{i}) it is respectively frequency peak spectral line serial number k_{i}Corresponding discrete Fourier transform and all phase are discrete
Fourier transformation spectral vector value,For corresponding phase spectrum, β_{i}k_{i}For peak spectral line k_{i}The frequency at place is inclined
Shifting amount, N are random natural number.
Specifically, the derivation process of the frequency offset formula of final all phase discrete Fourier transform is as follows:
For lowvoltage alternatingcurrent system, current signal acquisition, actual current letter are carried out to the different loads in low tension loop
Number it is expressed as the linear superposition of multiple signal components, load current sampled value y (n Δ t) formula is as follows:
(1) in formula: n ∈ [ N+1, N1], 2N1 are analysis data length；Δ t is sampling interval, ω_{i}、A_{i}、p_{i}Respectively
Angular frequency, amplitude and the initial phase of i signal component, M are signal component sum.
Discrete Fourier transform is carried out to the current signal in formula (1) respectively and all phase discrete Fourier transform (is ignored
The influence of negative frequency), respective spectrum expression formula are as follows:
(2) in (3) formula: G (k) is discrete Fourier transform frequency spectrum, and Y (k) is all phase discrete Fourier transform frequency spectrum, k
For spectral line serial number, β_{i}=ω_{i}/ Δ ω, Δ ω are angular frequency resolution ratio, ω_{i}For the angular frequency value of i frequency component, W (β_{i} k) be
Frequency offset β_{i} k locates corresponding rectangular window spectral function, A_{i}For the amplitude of i signal component, p_{i}For the initial phase of i signal component, M
For signal component sum；Wherein
If frequency component ω_{i}Spectral magnitude peak spectral line serial number k_{i}, in (4) formula: β_{i}k_{i}For peak spectral line k_{i}The frequency at place is inclined
Shifting amount, N are random natural number, β_{i}=ω_{i}/ Δ ω, Δ ω are angular frequency resolution ratio, ω_{i}For the angular frequency value of i frequency component.By
This, can be derived by the frequency offset of final all phase discrete Fourier transform are as follows:
(5) in formula: G (k_{i})、Y(k_{i}) it is respectively frequency peak spectral line serial number k_{i}Corresponding discrete Fourier transform and all phase
Discrete Fourier transform spectral vector value,For corresponding phase spectrum, β_{i}k_{i}For peak spectral line k_{i}The frequency at place
Offset, N are random natural number.
The then angular frequency value of i signal frequency component are as follows:
ω_{i}=β_{i}Δω (6)
The corresponding amplitude of i signal frequency component are as follows:
(7) in formula: G (k) is discrete Fourier transform frequency spectrum, and Y (k) is all phase discrete Fourier transform frequency spectrum, and k is spectrum
Line sequence number.
Formula (6) (7) is the corresponding angular frequency value of i signal frequency component and width based on all phase discrete Fourier transform
It is worth calculation formula.From formula (2) (3) as can be seen that when spectrum leakage occur, the amplitude of all phase discrete Fourier transform occurs
Decay at double, this is also the frequency spectrum of all phase discrete Fourier transform with good spectrum leakage and frequency spectrum interference rejection
Basic reason, therefore the frequency component measured has very high precision.
Further, all phase spectrum signature amount loaded is extracted, and is specifically included when load operates normally
Current spectrum feature extraction under current spectrum feature and series fault arc, by electric to operating normally and series fault occurring
Current waveform when arc carries out all phase spectrometry, extracts spectrum signature of the different loads under different conditions；By to each time
The amplitude distribution of harmonic component and the frequency of the mAcetyl chlorophosphonazo component to play a leading role, the identification of the feature of amplitude distribution situation, are realized
Whether break down the examination of electric arc to different loads；It the amplitude distribution by each harmonic component and plays a leading role
The frequency of mAcetyl chlorophosphonazo component, the identification of the feature of amplitude distribution situation, wherein the specific features of extraction specifically include it is leading between it is humorous
Wave frequency rate, leading mAcetyl chlorophosphonazo content, harmonic content.
Further, building all phase spectrum signature vector is by dominating mAcetyl chlorophosphonazo frequency f_{Ihm}, leading mAcetyl chlorophosphonazo content A_{Ihm}
With harmonic content A_{Hm}Form 150 dimensional feature vector X=[f_{Ihm}, A_{Ihm}, A_{Hm}], wherein m=1,2 ..., 50, Ihm indicate m
A mAcetyl chlorophosphonazo, Hm indicate mth of harmonic wave.
Further, all phase spectrum signature vector that the utilization is built up constructs the depth returned based on Logistic
Learning neural network model, specific as follows:
The present invention is based on the deep learning methods that Logistic is returned, this is because using traditional neural network to event
When hindering electric arc progress discriminant analysis, Accurate classification can not be carried out to complicated sample, while the convergence of result is by initial value shadow
The case where sound is larger, is as a result easily trapped into local optimum or overfitting or even does not restrain.And Logistic is returned not by initial
It is worth the influence chosen, result is still highly stable when initial value is unsatisfactory for assuming, has stronger robustness, therefore the present invention adopts
Sample training is carried out with the deep learning method returned based on Logistic.
To screen whether different loads generate fault electric arc, 50 subharmonic contents as defined in IEC6100047 standard are selected
A_{Hm}As characteristic parameter (m=1,2 ..., 50), while choose two subharmonic between leading mAcetyl chlorophosphonazo frequency f_{Ihm}, content
A_{Ihm}As characteristic parameter, 150 dimensional feature vector X=[f are collectively constituted_{Ihm},A_{Ihm},A_{Hm}] it is the defeated of neural network deep learning
Enter parameter, deep learning neural network includes 5 layers of hidden layer, the output response variable of deep learning neural network Each sample is loaded and is numbered by type, normal operation or fault electric arc state by it,And The type number for indicating to identify load is round (u+1)/2, and operating status is normal (u
For even number) or fault electric arc (u is odd number), wherein round () expression, which takes, rounds up.
Logistic regression model in the deep learning neural network model returned based on Logistic is nonthread
Property regression model, have following distribution form:
(8) in formula: E (X) indicates probability distribution, and p is the prediction probability value of corresponding input variable X, and ε is prediction error, β_{0}
And β_{1}For the nonlinear regression coefficient of model；
Logistic recurrence is the prediction probability value p for calculating Y=1 to the different level of input variable X, wherein probability value
Cutpoint, formula are used as using 0.5 are as follows:
(9) in formula: u=0,1 ..., 47；y_{u}=1 indicates that (u is surprise for normal condition (u is even number) or fault electric arc state
Number).
The present invention has the advantage that and the utility model has the advantages that
1, the present invention is using all phase spectrum signature formation mechenism of load current signal as point of penetration, to distorted signal and normally
Characteristic quantity of the signal on all phase frequency spectrum extracts, not vulnerable to noise, spectrum leakage, spectral interference；It is composed based on all phase
The measurement method of analysis can effectively extract current characteristic parameter under the various operating statuses of load, mention for the identification of fault electric arc
For theoretical foundation；
2, the deep learning neural network model that present invention building is returned based on Logistic, neural network have nonthread
Property, adaptivity and faulttolerant ability, prototype network can stablize, fast convergence；Utilize the depth returned based on Logistic
Classifying identification method is practised, all phase spectrum signature of input is differentiated, is realized to load type, whether break down electric arc
Examination；
3, the present invention for series fault arc recognition methods process it is reasonable, easy for operation, this method not vulnerable to
Noise, spectrum leakage interference, recognition efficiency is high, and stability is strong.
Detailed description of the invention
Attached drawing described herein is used to provide to further understand the embodiment of the present invention, constitutes one of the application
Point, do not constitute the restriction to the embodiment of the present invention.In the accompanying drawings:
Fig. 1 is the lowvoltage distribution system series fault arc method of identification process of the invention based on all phase deep learning
Figure.
Fig. 2 is series fault arc test platform schematic diagram of the invention.
Fig. 3 (a) is the normal current waveform figure of linear load of the invention.
Fig. 3 (b) is the normal current spectrum figure of linear load of the invention.
Fig. 4 (a) is the normal current waveform figure of nonlinear load of the invention.
Fig. 4 (b) is the normal current spectrum figure of nonlinear load of the invention.
Current waveform figure when Fig. 5 (a) is linear load fault electric arc of the invention.
Current spectrum figure when Fig. 5 (b) is linear load fault electric arc of the invention.
Current waveform figure when Fig. 6 (a) is nonlinear load fault electric arc of the invention.
Current spectrum figure when Fig. 6 (b) is nonlinear load fault electric arc of the invention.
Fig. 7 is the deep learning neural network model of the invention returned based on Logistic.
Fig. 8 is the deep learning training error figure of neural network of the invention.
Fig. 9 is the network training output error figure of the invention with other each methods.
Figure 10 (a) is the content tendency chart of each harmonic of the invention.
Figure 10 (b) is the content tendency chart of each secondary mAcetyl chlorophosphonazo of the invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below with reference to embodiment and attached drawing, to this
Invention is described in further detail, and exemplary embodiment of the invention and its explanation for explaining only the invention, are not made
For limitation of the invention.
Embodiment
As shown in Fig. 1 to Figure 10 (b), the lowvoltage distribution system series fault arc identification based on all phase deep learning
Method, this method comprises:
Under lowvoltage alternatingcurrent system, current signal acquisition is carried out to the different loads in low tension loop；
All phase discrete Fourier transform, all phase spectrum signature amount loaded are carried out to collected current signal
It extracts, and constructs all phase spectrum signature vector；
Using all phase spectrum signature vector built up, the deep learning neural network mould returned based on Logistic is constructed
Type, by carrying out deep learning training to all phase spectrum signature amount under different loads, different operating statuses, until the model
Until convergence；
The deep learning neural network model returned based on Logistic completed using learning training, to be identified point
Section load current waveform carries out all phase FFT spectrum analysis measurement, completes the examination to different loads type and whether series fault occurs
The identification of electric arc.
Specifically, described under lowvoltage alternatingcurrent system, current signal acquisition, electricity are carried out to the different loads in low tension loop
Stream signal is expressed as the linear superposition of multiple signal components, wherein load current samples formula are as follows:
In formula: n ∈ [ N+1, N1], 2N1 are analysis data length；Δ t is sampling interval, ω_{i}、A_{i}、p_{i}Respectively i letter
Angular frequency, amplitude and the initial phase of number component, M are signal component sum.
Specifically, described that all phase discrete Fourier transform is carried out to collected current signal, wherein final all phase
The frequency offset formula of discrete Fourier transform are as follows:
In formula: G (k_{i})、Y(k_{i}) it is respectively frequency peak spectral line serial number k_{i}Corresponding discrete Fourier transform and all phase are discrete
Fourier transformation spectral vector value,For corresponding phase spectrum, β_{i}k_{i}For peak spectral line k_{i}The frequency at place is inclined
Shifting amount, N are random natural number.
Specifically, all phase spectrum signature amount loaded is extracted, and specifically includes electricity when load operates normally
The current spectrum feature extraction under spectrum signature and series fault arc is flowed, by normal operation and generation series fault arc
When current waveform carry out all phase spectrometry, extract spectrum signature of the different loads under different conditions；By humorous to each time
The amplitude distribution of wave component and the frequency of the mAcetyl chlorophosphonazo component to play a leading role, the identification of the feature of amplitude distribution situation, realization pair
Whether different loads break down the examination of electric arc.It the amplitude distribution by each harmonic component and plays a leading role
The frequency of mAcetyl chlorophosphonazo component, the identification of the feature of amplitude distribution situation, wherein the specific features of extraction specifically include leading mAcetyl chlorophosphonazo
Frequency, leading mAcetyl chlorophosphonazo content, harmonic content.
The all phase spectrum signature amount that is loaded is extracted, specifically include current spectrum feature when load operates normally and
Current spectrum feature extraction under series fault arc, wherein all phase spectrum signature amount analytic process is as follows:
(1) current spectrum feature when load operates normally
According to all phase frequency spectrum formation mechenism abovementioned, to the corresponding linear load of formula (1), nonlinear load current signal
It carries out all phase discrete Fourier transform to go forward side by side the frequency and amplitude measurement of rowfrequency component, obtains its time domain waveform and spectrogram
Respectively as shown in Fig. 3 (a), Fig. 3 (b) and Fig. 4 (a), Fig. 4 (b).
From Fig. 3 (a), Fig. 3 (b) and Fig. 4 (a), Fig. 4 (b) as can be seen that waveform is close ideal when linear load operates normally
Sine wave, contain only fundametal compoment in spectrogram, all phase spectrum measurement result contains only fundamental voltage amplitude；Nonlinear load is normally transported
More regular cyclic distortion occurs for waveform when row, and each spectrum line in spectrogram forms spectral band, all phase spectrometry knot
Fruit includes each harmonic component and fewer parts mAcetyl chlorophosphonazo component, and wherein oddorder harmonic components are more prominent.Load operates normally
When extracted by the spectrum signature that linear load and nonlinear load may be implemented in all phase spectrometry method, and it is different types of
Load has different spectrum signatures, therefore can distinguish under normal operating condition different types of linear load and nonlinear negative
It carries.
(2) the current spectrum feature under series fault arc
By building series fault arc analogue test platform, the conventional linear under series fault arc is loaded, is nonthread
Property load current signal waveform carry out time and frequency domain analysis, obtain waveform diagram and spectrogram respectively such as Fig. 5 (a), Fig. 5 (b) and
Shown in Fig. 6 (a), Fig. 6 (b).
From 5 (a), Fig. 5 (b) and Fig. 6 (a), Fig. 6 (b) as can be seen that linear load, nonlinear load are in series fault electricity
Irregularity distortion occurs for waveform when arc, and waveform carries out all phase spectrum measurement with burr, to it, obtains humorous comprising each time
The mAcetyl chlorophosphonazo component of wave component and different frequency.Comparison diagram 3 (a), Fig. 3 (b) and Fig. 4 (a), Fig. 4 (b) simultaneously, when connecting
Current waveform when fault electric arc can be regarded as the polyteny superposition of the current distortion waveform when operating normally, aggravate letter
Number distortion degree, form spectral band on spectrogram, and with significantly different frequency, the mAcetyl chlorophosphonazo component of amplitude and outstanding
Harmonic component (such as 3 times, 5 subharmonic).
By carrying out all phase spectrometry to current waveform when operating normally and series fault arc occurs, can extract
To spectrum signature of the different loads under different conditions.By the amplitude distribution to each harmonic component and between playing a leading role
The frequency of harmonic component, the identification of the feature of amplitude distribution situation, the Zhen for the electric arc that may be implemented whether to break down to different loads
Not.
The building all phase spectrum signature vector is by dominating mAcetyl chlorophosphonazo frequency f_{Ihm}, leading mAcetyl chlorophosphonazo content A_{Ihm}It is harmonious
Wave content A_{Hm}Form 150 dimensional feature vector X=[f_{Ihm}, A_{Ihm}, A_{Hm}], wherein m=1,2 ..., 50, Ihm indicate mth between
Harmonic wave, Hm indicate mth of harmonic wave.
The all phase spectrum signature vector that the utilization is built up constructs the deep learning nerve net returned based on Logistic
Network model, specific as follows:
150 dimensional feature vector X=[f_{Ihm}, A_{Ihm}, A_{Hm}] input parameter as neural network deep learning, deep learning
Neural network includes 5 layers of hidden layer, the output response variable of deep learning neural networkIt will
Each sample load is numbered by type, normal operation or fault electric arc state,And The type number for indicating to identify load is round (u+1)/2, operating status be normal (u is even number) or
Fault electric arc (u is odd number), wherein round () expression, which takes, rounds up.
Logistic regression model in the deep learning neural network model returned based on Logistic is nonthread
Property regression model, have following distribution form:
In formula: E (X) indicates probability distribution, and p is the prediction probability value of corresponding input variable X, and ε is prediction error, β_{0}And β_{1}
For the nonlinear regression coefficient of model；
Logistic recurrence is the prediction probability value p for calculating Y=1 to the different level of input variable X, wherein probability value
Cutpoint, formula are used as using 0.5 are as follows:
U=0 in formula, 1 ..., 47；y_{u}=1 indicates normal condition (u is even number) or fault electric arc state (u is odd number).
As shown in fig. 7, for neural network training process in the deep learning network model returned based on Logistic, examination
Sample is tested using resistance box, insulating pot, incandescent lamp, power drill, the microwave oven, vacuum met under the different capacity of UL1699 standard
Total 24 kinds of loads such as dust catcher, LED light, fluorescent lamp, desk lamp with dimmer switch, and type, normal operation or fault electric arc are pressed into each load
State is numbered (0~47), takes 100 normal operation waveform samples, 100 fault electric arc waveform samples under every kind of load,
Total 4800 samples carry out neural network deep learning shown in Fig. 7.Sample is finally restrained, and does not occur traditional neural network
The case where not restraining, elearning accuracy is up to 100%.Neural network is returning the training in deep learning based on Logistic
Error mean square value as shown in figure 8, with deep learning frequency of training increase, network error gradient decline index it is more satisfactory, net
Network can stablize, fast convergence.
Depth is returned using traditional BP neural network (method 1), genetic algorithm neural network (method 2) and Logistisc
The output error of learning neural network (method 3) is as shown in Figure 9.
As seen from Figure 9, returning deep learning neural network (method 3) based on Logistic has preferable convergence
Property, the rapidity and the more other two methods of stability in convergence process are more excellent.Traditional BP neural network convergence is avoided simultaneously
Slowly, the neural network based on genetic algorithm restrains unstable disadvantage.
Therefore, the fault electric arc identification process figure based on all phase spectrum and deep learning is as shown in Figure 1.
For the lowvoltage distribution system series fault arc method of identification of the invention based on all phase deep learning, for simulation
Electrical quantity situation of change when true generation series fault arc builds series fault electricity according to UL16992008 AFCI standard
Arc analogue test platform is as shown in Fig. 2, the lowvoltage alternatingcurrent system includes power supply unit, failure generating unit, data sampling point
Unit is analysed, the power supply unit uses lowvoltage alternatingcurrent experiment power supply AC, the lowvoltage alternatingcurrent experiment power supply AC to connect by switch K1
Connect primary coil one end of isolating transformer, the primary coil other end ground connection of isolating transformer, the secondary sideline of isolating transformer
It encloses one end and connects arc generator, arc generator connects resistance R1, resistance R1 connection slide rheostat R_{p}, rheostat R_{p}Pass through
The secondary coil other end of switch K2 connection isolating transformer；R_{p}、R_{1}=0.1 Ω respectively indicates test load and sampling resistor；Institute
Arc generator is stated for generating stable series fault arc, and is parallel with switch K3 on the arc generator；The electricity
Data acquisition unit DA is connected on resistance R1, the data acquisition unit DA connection computer PC realizes analysis.
To 4 sections of load current waves of State Grid Sichuan Electric Power Company's production and operation central data where the test platform of building
Shape carries out all phase FFT spectrum analysis measurement, and the variation for obtaining each (with highest 10 times for example) harmonic wave, leading mAcetyl chlorophosphonazo content becomes
Gesture such as Figure 10 (a), shown in 10 (b).
The visible characteristic quantity based on all phase FFT spectrum analysis is more apparent in Figure 10 (a), the harmonic content concentration of signal 1, signal 3
It is distributed at odd harmonics, and the harmonic content of signal 1 is integrally minimum, highest content about 2.5%, the harmonic content of signal 3
Whole higher, highest content about 26.6%；Signal 2, signal 4 are distributed at each harmonic, from the point of view of harmonic content, signal
4 integral level is higher than signal 2.See that signal 2, the mAcetyl chlorophosphonazo content of signal 4 are whole higher from Figure 10 (b), signal 1 and letter
Numbers 3 mAcetyl chlorophosphonazo content is integrally lower.Further stimulus is quantified, obtains the feature value part composed based on all phase
Measurement result is as shown in table 1.
Part measurement result of the table 1 based on all phase FFT spectrum analysis
Abovementioned test result shows that the measurement method based on all phase FFT spectrum analysis can effectively extract the various operations of load
Current characteristic parameter under state provides theoretical foundation for the identification of fault electric arc.
By the leading mAcetyl chlorophosphonazo frequency f based on all phase spectrometry_{Ihm}, content A_{Ihm}With harmonic content A_{Hm}150 dimension of composition
Feature vector (m=1,2 ..., 50), as be completed in Fig. 7 Logistic return deep learning nerve network input parameter,
Output and the fault electric arc recognition result for obtaining 4 sections of stimulus are as shown in table 2.The conclusion (of pressure testing) is consistent with test preset condition
It closes, demonstrates the correct of the lowvoltage distribution system series fault arc method of identification based on all phase deep learning that the present invention is mentioned
Property and validity, realize the identification to load type, the electric arc that whether breaks down, and recognition efficiency is high, method stability is strong.
The test result of each segment signal of table 2
Abovedescribed specific embodiment has carried out further the purpose of the present invention, technical scheme and beneficial effects
It is described in detail, it should be understood that being not intended to limit the present invention the foregoing is merely a specific embodiment of the invention
Protection scope, all within the spirits and principles of the present invention, any modification, equivalent substitution, improvement and etc. done should all include
Within protection scope of the present invention.
Claims (9)
1. the lowvoltage distribution system series fault arc method of identification based on all phase deep learning, which is characterized in that this method packet
It includes:
Under lowvoltage alternatingcurrent system, current signal acquisition is carried out to the different loads in low tension loop；
All phase discrete Fourier transform is carried out to collected current signal, all phase spectrum signature amount loaded mentions
It takes, and constructs all phase spectrum signature vector；
Using all phase spectrum signature vector built up, the deep learning neural network model returned based on Logistic is constructed,
By carrying out deep learning training to all phase spectrum signature amount under different loads, different operating statuses, until the model is received
Until holding back；
The deep learning neural network model returned based on Logistic completed using learning training, it is negative to segmentation to be identified
It carries current waveform and carries out all phase FFT spectrum analysis measurement, carry out the examination to different loads type and whether series fault arc occurs
Identification, according to the model exportEach load is pressed type, normal operation or fault electric arc by it
State is numbered,AndWherein u ≠ v, u, v=0,1 ..., 47, indicate to identify that the type number of load is
Round (u+1)/2, operating status are normal arc, and u is even number；Or operating status is fault electric arc, u is odd number；Wherein
Round () expression, which takes, to round up.
2. the lowvoltage distribution system series fault arc method of identification according to claim 1 based on all phase deep learning,
It is characterized in that, it is described under lowvoltage alternatingcurrent system, current signal acquisition, electric current letter are carried out to the different loads in low tension loop
Number it is expressed as the linear superposition of multiple signal components, wherein load current samples formula are as follows:
In formula: n ∈ [ N+1, N1], 2N1 are analysis data length；Δ t is sampling interval, ω_{i}、A_{i}、p_{i}Respectively i signal point
Angular frequency, amplitude and the initial phase of amount, M are signal component sum.
3. the lowvoltage distribution system series fault arc method of identification according to claim 1 based on all phase deep learning,
It is characterized in that, described carry out all phase discrete Fourier transform to collected current signal, wherein final all phase is discrete
The frequency offset formula of Fourier transformation are as follows:
In formula: G (k_{i})、Y(k_{i}) it is respectively frequency peak spectral line serial number k_{i}Corresponding discrete Fourier transform and all phase direct computation of DFT
Leaf transformation spectral vector value,For corresponding phase spectrum, β_{i}k_{i}For peak spectral line k_{i}The frequency offset at place,
N is random natural number.
4. the lowvoltage distribution system series fault arc method of identification according to claim 1 based on all phase deep learning,
It is characterized in that, the lowvoltage alternatingcurrent system includes power supply unit, failure generating unit, data sampling analytical unit, the electricity
Source unit uses lowvoltage alternatingcurrent experiment power supply AC, the lowvoltage alternatingcurrent experiment power supply AC to pass through switch K1 connection isolating transformer
Primary coil one end, the primary coil other end ground connection of isolating transformer, secondary coil one end of isolating transformer connects electric arc
Generator, arc generator connect resistance R1, resistance R1 connection slide rheostat R_{p}, rheostat R_{p}It is isolated by switch K2 connection
The secondary coil other end of transformer；The arc generator is for generating stable series fault arc, and the electric arc is sent out
Switch K3 is parallel on raw device；Data acquisition unit DA, the data acquisition unit DA connection meter are connected on the resistance R1
Calculation machine PC realizes analysis；Wherein, slide rheostat R_{p}, resistance R_{1}Respectively indicate load and sampling resistor.
5. the lowvoltage distribution system series fault arc method of identification according to claim 1 based on all phase deep learning,
It is characterized in that, all phase spectrum signature amount loaded is extracted, electric current frequency when load operates normally is specifically included
Current spectrum feature extraction under spectrum signature and series fault arc, when by normal operation and generation series fault arc
Current waveform carries out all phase spectrometry, extracts spectrum signature of the different loads under different conditions；By to each harmonic point
The amplitude distribution of amount and the frequency of the mAcetyl chlorophosphonazo component to play a leading role, the identification of the feature of amplitude distribution situation, are realized to difference
Load the examination for the electric arc that whether breaks down.
6. the lowvoltage distribution system series fault arc method of identification according to claim 5 based on all phase deep learning,
It is characterized in that, the amplitude distribution by each harmonic component and the frequency of the mAcetyl chlorophosphonazo component to play a leading role, width
The feature of Distribution value situation identifies, wherein the specific features of extraction specifically include leading mAcetyl chlorophosphonazo frequency, leading mAcetyl chlorophosphonazo contains
Amount, harmonic content.
7. the lowvoltage distribution system series fault arc method of identification according to claim 6 based on all phase deep learning,
It is characterized in that, the building all phase spectrum signature vector is by dominating mAcetyl chlorophosphonazo frequency f_{Ihm}, leading mAcetyl chlorophosphonazo content A_{Ihm}With
Harmonic content A_{Hm}Form 150 dimensional feature vector X=[f_{Ihm}, A_{Ihm}, A_{Hm}], wherein m=1,2 ..., 50, Ihm indicate mth
MAcetyl chlorophosphonazo, Hm indicate mth of harmonic wave.
8. the lowvoltage distribution system series fault arc method of identification according to claim 7 based on all phase deep learning,
It is characterized in that, all phase spectrum signature vector that the utilization is built up, constructs the deep learning mind returned based on Logistic
It is specific as follows through network model:
150 dimensional feature vector X=[f_{Ihm}, A_{Ihm}, A_{Hm}] input parameter as neural network deep learning, deep learning nerve
Network includes 5 layers of hidden layer, the output response variable of deep learning neural networkIt is by various kinds
This load is numbered by type, normal operation or fault electric arc state,AndWherein u ≠ v, u, v=0,1 ...,
47, indicate to identify that the type number of load is round (u+1)/2, operating status is normal arc, and u is even number；Or operation
State is fault electric arc, and u is odd number；Wherein round () expression, which takes, rounds up.
9. the lowvoltage distribution system series fault arc method of identification according to claim 8 based on all phase deep learning,
It is characterized in that, the Logistic regression model in the deep learning neural network model returned based on Logistic is non
Linear regression model (LRM) has following distribution form:
In formula: E (X) indicates probability distribution, and p is the prediction probability value of corresponding input variable X, and ε is prediction error, β_{0}And β_{1}For mould
The nonlinear regression coefficient of type；
Logistic recurrence is the prediction probability value p for calculating Y=1 to the different level of input variable X, and wherein probability value uses
0.5 is used as cutpoint, formula are as follows:
U=0 in formula, 1 ..., 47；y_{u}=1 indicates normal condition, and u is even number；Or y_{u}=1 indicates fault electric arc state, and u is surprise
Number.
Priority Applications (1)
Application Number  Priority Date  Filing Date  Title 

CN201910739837.0A CN110376497A (en)  20190812  20190812  Lowvoltage distribution system series fault arc method of identification based on all phase deep learning 
Applications Claiming Priority (1)
Application Number  Priority Date  Filing Date  Title 

CN201910739837.0A CN110376497A (en)  20190812  20190812  Lowvoltage distribution system series fault arc method of identification based on all phase deep learning 
Publications (1)
Publication Number  Publication Date 

CN110376497A true CN110376497A (en)  20191025 
Family
ID=68258858
Family Applications (1)
Application Number  Title  Priority Date  Filing Date 

CN201910739837.0A Pending CN110376497A (en)  20190812  20190812  Lowvoltage distribution system series fault arc method of identification based on all phase deep learning 
Country Status (1)
Country  Link 

CN (1)  CN110376497A (en) 
Citations (6)
Publication number  Priority date  Publication date  Assignee  Title 

CN103543375A (en) *  20130826  20140129  上海交通大学  Method for detecting alternatingcurrent fault arcs on basis of wavelet transformation and timedomain hybrid features 
CN106093569A (en) *  20160530  20161109  中国民用航空总局第二研究所  Navigation signal measuring method, system and navigator inline diagnosis method, system 
CN106483374A (en) *  20160927  20170308  福州大学  A kind of harmonic wave harmonic detection method based on Nuttall double window whole phase FFT 
CN106597022A (en) *  20161223  20170426  北京化工大学  Accelerometer dynamic model parameter identification method based on allphase spectrum estimation 
CN108646149A (en) *  20180428  20181012  国网江苏省电力有限公司苏州供电分公司  Fault electric arc recognition methods based on current characteristic extraction 
CN109738771A (en) *  20190312  20190510  长春工程学院  Tandem type fault electric arc test platform and its test method based on deep learning 

2019
 20190812 CN CN201910739837.0A patent/CN110376497A/en active Pending
Patent Citations (6)
Publication number  Priority date  Publication date  Assignee  Title 

CN103543375A (en) *  20130826  20140129  上海交通大学  Method for detecting alternatingcurrent fault arcs on basis of wavelet transformation and timedomain hybrid features 
CN106093569A (en) *  20160530  20161109  中国民用航空总局第二研究所  Navigation signal measuring method, system and navigator inline diagnosis method, system 
CN106483374A (en) *  20160927  20170308  福州大学  A kind of harmonic wave harmonic detection method based on Nuttall double window whole phase FFT 
CN106597022A (en) *  20161223  20170426  北京化工大学  Accelerometer dynamic model parameter identification method based on allphase spectrum estimation 
CN108646149A (en) *  20180428  20181012  国网江苏省电力有限公司苏州供电分公司  Fault electric arc recognition methods based on current characteristic extraction 
CN109738771A (en) *  20190312  20190510  长春工程学院  Tandem type fault electric arc test platform and its test method based on deep learning 
NonPatent Citations (2)
Title 

张庆丰: ""利用Logistic二次回归法提高判别分析效率"", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * 
李海峰 等: ""消除外部环境因素干扰的MOA 在线监测参数修正方法"", 《高电压技术》 * 
Similar Documents
Publication  Publication Date  Title 

Khokhar et al.  A new optimal feature selection algorithm for classification of power quality disturbances using discrete wavelet transform and probabilistic neural network  
CN103136587B (en)  Based on the power distribution network running status classifying identification method of support vector machine  
CN105429135B (en)  The identification decisionmaking technique and system that a kind of nonintrusive electrical load decomposes  
Elkalashy et al.  Modeling and experimental verification of high impedance arcing fault in medium voltage networks  
El Safty et al.  Applying wavelet entropy principle in fault classification  
Wang  A novel extension method for transformer fault diagnosis  
Koziy et al.  A lowcost powerquality meter with series arcfault detection capability for smart grid  
Cui et al.  Hilberttransformbased transient/intermittent earth fault detection in noneffectively grounded distribution systems  
Lai et al.  Highimpedance fault detection using discrete wavelet transform and frequency range and RMS conversion  
Sedighi et al.  High impedance fault detection based on wavelet transform and statistical pattern recognition  
CN103454497B (en)  Based on the method for measuring phase difference improving windowed DFT  
CN106646096B (en)  Transformer fault classification and recognition methods based on vibration analysis method  
Manikandan et al.  Detection and classification of power quality disturbances using sparse signal decomposition on hybrid dictionaries  
Zhang et al.  An artificial neural network approach to transformer fault diagnosis  
CN104217110B (en)  A kind of the GIS state evaluating methods and device of the analysis of utilization benchmark state  
CN106556781A (en)  Shelf depreciation defect image diagnostic method and system based on deep learning  
Majidi et al.  Improving pattern recognition accuracy of partial discharges by new data preprocessing methods  
Hettiwatte et al.  Simulation of a transformer winding for partial discharge propagation studies  
CN103091612B (en)  Separation and recognition algorithm for transformer oiled paper insulation multiple partial discharging source signals  
CN104914394A (en)  Current transformer operation evaluation method  
CN102103183B (en)  Partial discharge measurement device under impulse voltage on site and signal processing method thereof  
Baqui et al.  High impedance fault detection methodology using wavelet transform and artificial neural networks  
CN104316487B (en)  Forbidden objects detection method and device  
Jiang et al.  Evaluating the safety condition of porcelain insulators by the time and frequency characteristics of LC based on artificial pollution tests  
Barbaro et al.  A novel approach based on nonactive power for the identification of disturbing loads in power systems 
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 