CN109325526A - A kind of distribution network failure classification method using convolution depth confidence network - Google Patents
A kind of distribution network failure classification method using convolution depth confidence network Download PDFInfo
- Publication number
- CN109325526A CN109325526A CN201811029500.2A CN201811029500A CN109325526A CN 109325526 A CN109325526 A CN 109325526A CN 201811029500 A CN201811029500 A CN 201811029500A CN 109325526 A CN109325526 A CN 109325526A
- Authority
- CN
- China
- Prior art keywords
- layer
- feature
- frequency
- pond
- time
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Locating Faults (AREA)
Abstract
The present invention relates to a kind of distribution network failure classification methods using convolution depth confidence network, three-phase voltage, residual voltage and the main transformer low-pressure side three-phase current for acquiring main transformer low-voltage bus bar first intercept the signal waveform data of each cycle in failure front and back as training sample to all kinds of fault conditions respectively；Then Time-frequency Decomposition is carried out using training sample data of the discrete analog method to step S1, seeks time-frequency matrix, and then construct the picture element matrix of time-frequency spectrum, and construct time-frequency spectrum, the input as subsequent CDBN model；Then CDBN model is constructed, two convolution of training are limited Boltzmann machine in a manner of unsupervised learning, and softmax classifier is added after the 2nd CRBM, and training network model realizes the effective extraction and automatic classification of fault signature；Finally classified using trained model realization distribution network failure.The present invention can be realized accurate fault location.
Description
Technical field
The present invention relates to battery testing field, especially a kind of distribution network failure using convolution depth confidence network is classified
Method.
Background technique
Distribution net work structure is increasingly complicated, breaks down unavoidable, wherein short circuit, ground fault it is most commonly seen.Event occurs
After barrier, no matter network reconfiguration, fault location or crash analysis, investigation maintenance are all very dependent on to the accurate of fault type
Classification.But the presence of many disturbing factors accurately identifies fault type to power distribution network and manufactured problem: power distribution network is vulnerable to user
The interference such as the noise and harmonic wave at scene, so that fault signature is more fuzzy；Further, since constituent is complicated, branch is more, match
Electrical measure feature when electric network fault, is affected by factors such as neutral grounding mode and fault resstances；Also, with distribution
Net more automated, the upload of a large amount of fault datas is difficult to operator to distinguish the mode of fault type according to experience
It realizes, traditional fault type recognition method can not adapt to the complicated and changeable of power distribution network.To sum up, a kind of event of high efficient and reliable is found
Hinder classification method to accurate fault location, quickly repair faulty line, maintains power distribution network safe operation and raising power supply can
Have great importance by property.
Currently, the basic step of Fault Classification is: obtaining fault transient electrical quantity and carry out signal decomposition, in conjunction with number
Method carries out feature extraction, selection, and suitable mode identification method is selected to carry out failure modes.But the side based on this step
Method, related signal decomposition extract characteristic quantity and mode identification method, and fault characteristic value requires manually to extract, vulnerable to people
It for factor interference and need to take considerable time, increase the uncertainty of result.
Summary of the invention
In view of this, the purpose of the present invention is to propose to a kind of distribution network failure classification sides using convolution depth confidence network
Method can be realized accurate fault location.
The present invention is realized using following scheme: a kind of distribution network failure classification method using convolution depth confidence network,
Specifically includes the following steps:
Step S1: three-phase voltage, residual voltage and the main transformer low-pressure side three-phase current of acquisition main transformer low-voltage bus bar, to all kinds of
Fault condition intercepts the signal waveform data of each cycle in failure front and back as training sample respectively；
Step S2: Time-frequency Decomposition is carried out using training sample data of the discrete analog method to step S1, seeks time-frequency
Matrix, and then the picture element matrix of time-frequency spectrum is constructed, and construct time-frequency spectrum, the input as subsequent CDBN model；
Step S3: building CDBN model, two convolution of training are limited Boltzmann machine in a manner of unsupervised learning, the 2nd
Softmax classifier is added after a CRBM, training network model realizes the effective extraction and automatic classification of fault signature；
Step S4: time-frequency spectrum is constructed with the step S1 test sample obtained and step S2, step S3 is input to and has instructed
Practice in perfect CDBN model, realizes distribution network failure classification.
Further, step S1 specifically: establish a radiant type medium-voltage distribution pessimistic concurrency control, 10 kinds of events are obtained by the model
The emulation of the electrical quantity including three-phase voltage, residual voltage and three-phase current of each 1 frequency cycle in front and back occurs for barrier
Waveform.
Further, 10 kinds of failures include A, B, C singlephase earth fault, AB, AC, BC double earthfault, AB,
AC, BC two-phase phase fault and ABC three-phase phase fault.
Further, step S2 specifically includes the following steps:
Step S21: the good db4 of regularity is chosen as wavelet basis function, Decomposition order is 4 layers, to selected voltage
Time-frequency Decomposition is carried out with current signal；
Step S22: being reconstructed the 4th layer of wavelet coefficient, respectively obtains the low frequency part S of signal4,0And radio-frequency head
Divide S4,1-S4,15；
Step S23: S is chosen4,0-S4,4As the frequency range of signal analysis, time-frequency matrix is constructed；
Step S24: the time-frequency matrix obtained after discrete analog method is changed into 35 × 380 time-frequency spectrum.
Wherein, in step S23, by taking A phase voltage as an example, sample frequency 10kHz, sampling number n=400,5 son frequencies
Data point with waveform is aij(i=1,2 ..., 5；J=1,2 ... n), then time-frequency matrix can be obtained are as follows:
Wherein, the row of time-frequency matrix A indicates the reconfiguration waveform of fault waveform each sub-band after discrete analog method
Data, column indicate the sampling instant of fault waveform.Time-frequency matrix A be completely demonstrated by fault waveform on each sub-band when
Frequency feature enumerates the temporary steady state information in fault waveform, and is separated by frequency band to them, is conducive to mentioning for characteristic quantity
It takes.Each time-frequency matrix is changed into the picture element matrix of time-frequency spectrum, and by three-phase voltage, three-phase current and residual voltage sequentially from upper
35 rows are built up under and, for the influence for inhibiting boundary effect, crop each 10 sampled points of head and the tail, obtaining image pattern size is 35
× 380 time-frequency spectrum.
Further, in step S3, the CDBN model uses 7 model of a layered structure, comprising: visible layer 1, hidden layer 1, pond
Change layer 1, visible layer 2, hidden layer 2, pond layer 2, full articulamentum and output layer, wherein pond layer 1 is same with visible layer 2
Layer includes 2 visible layers, 2 hidden layers, 2 pond layers, 1 full articulamentum and 1 output layer；The CRBM model is abided by
Greedy order training method principle is followed, using the operation of convolution sum pondization to obtain, level is deeper, the stronger fault signature of ability to express；
Convolutional calculation be in CDBN model visible layer to the calculating of hidden layer.By convolution operation, K are generated in hidden layer
Characteristic pattern as next layer of input to extract more advanced feature, the hiding layer unit of k-th of characteristic pattern the i-th row jth column
Condition activates probability are as follows:
In formula, v represents visible layer, bkIndicate the biasing of k-th of characteristic pattern,It indicates in visible layer and hidden layer k-th
The interlayer weight of characteristic pattern, σ=1/ (1+e-x) it is Sigmoid activation primitive；
Pond is the calculating of hidden layer to pond layer, down-sampled to hidden layer progress using maximum probability pond method, with
Hidden layer is not divided into several B by overlaid windowsαRegion；Maximum value in selection area is taken to obtain as output valve by severalInput of the pond layer feature of region composition as next CRBM；As the operation of a dimensionality reduction and regularization, Chi Huacao
Work can reduce the training complexity of model, reduce computation burden.The condition of k-th of pond layer unit activates probability are as follows:
The characteristic pattern that CRBM is extracted successively is pressed into column expansion, obtains a feature vector, maps that sample labeling sky
Between, a new feature representation form is obtained, calculates current each sample in the corresponding probability of each type are as follows:
T (i)=Y (li)；
In formula, Y indicates soft-max classifier functions, liNumber indicates the calculated result of the i-th class input data；In T (i) most
The number of big value position is the label of fault type.Table specific as follows.
Further, step S4 specifically includes the following steps:
Step S41: in the time-frequency spectrum of 1 input fault waveform of visible layer, convolution is carried out to input picture using convolution kernel
Operation exports fault signature figure in hidden layer 1；
Step S42: input picture of the output image of hidden layer 1 as pond layer 1 carries out maximum pondization operation, passes through
Pond layer 1 is extracted the integration that obtained feature carries out the adjacent similar features of regional area, reaches effective by the mode in maximum pond
Reduce the purpose of intrinsic dimensionality；
Step S43: sending the output image of pond layer 1 into hidden layer 2 and carry out the second wheel convolution operation, so as to will before
The feature that face is extracted further is abstracted, and feature more advanced, that discrimination is bigger is obtained；
Step S44: the characteristic pattern that hidden layer 2 exports needs to carry out primary maximum pondization operation using pond layer 2, compares
The characteristic pattern that pond layer 1 exports is more regular, this is the feature extraction more profound to original image, characterizes original image
Inherent feature obtains advanced features figure；
Step S45: advanced features figure is successively unfolded by column in full articulamentum and is stacked into feature vector；
Step S46: feature vector carries out connecting operation entirely with output layer in full articulamentum, and final output differentiates result.
Compared with prior art, the invention has the following beneficial effects: the method that the present invention uses deep learning, can be automatic
Fault characteristic value is extracted, and has accurate failure modes rate.
Detailed description of the invention
Fig. 1 is the method flow schematic diagram of the embodiment of the present invention.
The CDBN structure chart that Fig. 2 designs for the embodiment of the present invention.
Fig. 3 is the software phantom figure of the 10kV power distribution network of the embodiment of the present invention.
Specific embodiment
The present invention will be further described with reference to the accompanying drawings and embodiments.
It is noted that described further below be all exemplary, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has usual with the application person of an ordinary skill in the technical field
The identical meanings of understanding.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular
Also it is intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or their combination.
As shown in Figure 1, a kind of distribution network failure classification method using convolution depth confidence network is present embodiments provided,
Specifically includes the following steps:
Step S1: three-phase voltage, residual voltage and the main transformer low-pressure side three-phase current of acquisition main transformer low-voltage bus bar, to all kinds of
Fault condition intercepts the signal waveform data of each cycle in failure front and back as training sample respectively；
Step S2: Time-frequency Decomposition is carried out using training sample data of the discrete analog method to step S1, seeks time-frequency
Matrix, and then the picture element matrix of time-frequency spectrum is constructed, and construct time-frequency spectrum, the input as subsequent CDBN model；
Step S3: building CDBN model, two convolution of training are limited Boltzmann machine in a manner of unsupervised learning, the 2nd
Softmax classifier is added after a CRBM, training network model realizes the effective extraction and automatic classification of fault signature；
Step S4: time-frequency spectrum is constructed with the step S1 test sample obtained and step S2, step S3 is input to and has instructed
Practice in perfect CDBN model, realizes distribution network failure classification.
In the present embodiment, step S1 specifically: establish a radiant type medium-voltage distribution pessimistic concurrency control and (utilize PSCAD/EMTDC
10kV electricity distribution network model is built by simulation software), include by what the model obtained 10 kinds of each 1 frequency cycles in failures generation front and back
The simulation waveform of electrical quantity including three-phase voltage, residual voltage and three-phase current.
In the present embodiment, 10 kinds of failures include A, B, C singlephase earth fault, AB, AC, BC double earthfault,
AB, AC, BC two-phase phase fault and ABC three-phase phase fault.
In the present embodiment, step S2 specifically includes the following steps:
Step S21: the good db4 of regularity is chosen as wavelet basis function, Decomposition order is 4 layers, to selected voltage
Time-frequency Decomposition is carried out with current signal；
Step S22: being reconstructed the 4th layer of wavelet coefficient, respectively obtains the low frequency part S of signal4,0And radio-frequency head
Divide S4,1-S4,15；
Step S23: S is chosen4,0-S4,4As the frequency range of signal analysis, time-frequency matrix is constructed；
Step S24: the time-frequency matrix obtained after discrete analog method is changed into 35 × 380 time-frequency spectrum.
Particularly, during wavelet package transforms, each sub-band sequence can be made to interlock the decomposition of high frequency band, i.e., frequency band is handed over
Wrong phenomenon.When carrying out 4 layers of WAVELET PACKET DECOMPOSITION to the signal that sample frequency is f, in the 2nd layer of decomposition, S2,2With S2,3There is frequency band
Staggeredly phenomenon, practical S2,2For high frequency band, S2,3For low-frequency band.Staggered frequency band will further generate friendship in the 3rd layer of decomposition
It is wrong.After from low to high sub-band being rearranged according to frequency, respectively obtain the low frequency part S of signal4,0With high frequency section S4,1-
S4,15, choose S4,0-S4,4Frequency range as signal analysis.
Wherein, in step S23, by taking A phase voltage as an example, sample frequency 10kHz, sampling number n=400,5 son frequencies
Data point with waveform is aij(i=1,2 ..., 5；J=1,2 ... n), then time-frequency matrix can be obtained are as follows:
Wherein, the row of time-frequency matrix A indicates the reconfiguration waveform of fault waveform each sub-band after discrete analog method
Data, column indicate the sampling instant of fault waveform.Time-frequency matrix A be completely demonstrated by fault waveform on each sub-band when
Frequency feature enumerates the temporary steady state information in fault waveform, and is separated by frequency band to them, is conducive to mentioning for characteristic quantity
It takes.Each time-frequency matrix is changed into the picture element matrix of time-frequency spectrum, and by three-phase voltage, three-phase current and residual voltage sequentially from upper
35 rows are built up under and, for the influence for inhibiting boundary effect, crop each 10 sampled points of head and the tail, obtaining image pattern size is 35
× 380 time-frequency spectrum.
In the present embodiment, in step S3, the CDBN model uses 7 model of a layered structure, comprising: visible layer 1, hidden layer
1, pond layer 1, visible layer 2, hidden layer 2, pond layer 2, full articulamentum and output layer, wherein pond layer 1 be with visible layer 2
Same layer includes 2 visible layers, 2 hidden layers, 2 pond layers, 1 full articulamentum and 1 output layer；The CRBM mould
Type follows greedy order training method principle, and using the operation of convolution sum pondization to obtain, level is deeper, the stronger failure of ability to express
Feature；
Convolutional calculation be in CDBN model visible layer to the calculating of hidden layer.By convolution operation, K are generated in hidden layer
Characteristic pattern as next layer of input to extract more advanced feature, the hiding layer unit of k-th of characteristic pattern the i-th row jth column
Condition activates probability are as follows:
In formula, v represents visible layer, bkIndicate the biasing of k-th of characteristic pattern,It indicates in visible layer and hidden layer k-th
The interlayer weight of characteristic pattern, σ=1/ (1+e-x) it is Sigmoid activation primitive；
Pond is the calculating of hidden layer to pond layer, down-sampled to hidden layer progress using maximum probability pond method, with
Hidden layer is not divided into several B by overlaid windowsαRegion；Maximum value in selection area is taken to obtain as output valve by severalInput of the pond layer feature of region composition as next CRBM；As the operation of a dimensionality reduction and regularization, Chi Huacao
Work can reduce the training complexity of model, reduce computation burden.The condition of k-th of pond layer unit activates probability are as follows:
The characteristic pattern that CRBM is extracted successively is pressed into column expansion, obtains a feature vector, maps that sample labeling sky
Between, a new feature representation form is obtained, calculates current each sample in the corresponding probability of each type are as follows:
T (i)=Y (li)；
In formula, Y indicates soft-max classifier functions, liNumber indicates the calculated result of the i-th class input data；In T (i) most
The number of big value position is the label of fault type.Table specific as follows.
In the present embodiment, step S4 specifically includes following steps, and wherein model is as shown in Figure 2:
Step S41: in the time-frequency spectrum of 1 input fault waveform of visible layer, convolution is carried out to input picture using convolution kernel
Operation exports fault signature figure in hidden layer 1；
Step S42: input picture of the output image of hidden layer 1 as pond layer 1 carries out maximum pondization operation, passes through
Pond layer 1 is extracted the integration that obtained feature carries out the adjacent similar features of regional area, reaches effective by the mode in maximum pond
Reduce the purpose of intrinsic dimensionality；
Step S43: sending the output image of pond layer 1 into hidden layer 2 and carry out the second wheel convolution operation, so as to will before
The feature that face is extracted further is abstracted, and feature more advanced, that discrimination is bigger is obtained；
Step S44: the characteristic pattern that hidden layer 2 exports needs to carry out primary maximum pondization operation using pond layer 2, compares
The characteristic pattern that pond layer 1 exports is more regular, this is the feature extraction more profound to original image, characterizes original image
Inherent feature obtains advanced features figure；
Step S45: advanced features figure is successively unfolded by column in full articulamentum and is stacked into feature vector；
Step S46: feature vector carries out connecting operation entirely with output layer in full articulamentum, and final output differentiates result.
In order to allow those skilled in the art to further appreciate that technical solution proposed by the present invention, combined with specific embodiments below
It is illustrated.As shown in figure 3, in the present embodiment, being obtained using the 10kV power distribution network software phantom that simulation software is built
The electric quantity signals such as main transformer low-pressure side inlet wire current and busbar voltage with discrete analog method and construct time-frequency matrix, by time-frequency
Matrix conversion is used as CDBN to input at after the picture element matrix of time-frequency spectrum, independently extracts fault characteristic value through CDBN, is finally completed
Distribution network failure classification.Wherein, training sample is 1080, and test sample is 6480.
Distribution network failure classifying step are as follows:
(1) acquisition of time-frequency spectrum
According to above-mentioned technical proposal provided in this embodiment, intercepts failure and each 1 frequency cycle in front and back occurs (to inhibit side
Boundary's effect gives up front and back 10 sampled points, totally 380 sampled points) bus three-phase voltage, residual voltage and main transformer low-pressure side
The simulation waveform of three-phase current.Discrete analog method processing is done respectively to each waveform, selection db4 is wavelet basis function, point
Solving the number of plies is 4 layers.5 components before reconstructing the 4th layer are stacked by the sequence of three-phase voltage, three-phase current, residual voltage, obtain figure
As the time-frequency spectrum that sample size is 35 × 380.
(2) failure modes
According to above-mentioned technical proposal provided in this embodiment, time-frequency spectrum that the input picture of input layer is 35 × 380；
Visible layer 1 uses 6 sizes to carry out convolution operation, convolution moving step length to input picture for 4 × 13 convolution kernel
It is 1, exports 6 32 × 368 characteristic patterns in hidden layer 1；
Pond layer 1 is using maximum pond mode, and sampling window size 2 × 2, transverse and longitudinal step-length is 2, exports 6 16 × 184
Characteristic pattern；
Visible layer 2 (pond layer 1) carries out convolution operation to 6 characteristic patterns using 6 × 12=72 5 × 13 convolution kernel,
Convolution moving step length is 1, obtains 12 12 × 172 output characteristic patterns in hidden layer 2；
Layer 2 parameter setting in pond is identical as pond layer 1, the characteristic pattern that output is 12 6 × 86；
12 characteristic patterns that pond layer exports are unfolded by column and are stacked into 6192 × 1 feature vector by full articulamentum,
Full articulamentum carries out connecting operation entirely with output layer；
Output layer exports one 10 × 1 type identification vector, obtains classification results.
Each element value in result is exported in [0,1], represents the probability for being determined as such.It is maximized where element
Position Number as finally differentiate fault type number.
Classification results: classification accuracy rate is up to 99% or more.
The foregoing is merely presently preferred embodiments of the present invention, all equivalent changes done according to scope of the present invention patent with
Modification, is all covered by the present invention.
Claims (6)
1. a kind of distribution network failure classification method using convolution depth confidence network, it is characterised in that: the following steps are included:
Step S1: three-phase voltage, residual voltage and the main transformer low-pressure side three-phase current of acquisition main transformer low-voltage bus bar, to all kinds of failures
Operating condition intercepts the signal waveform data of each cycle in failure front and back as training sample respectively；
Step S2: carrying out Time-frequency Decomposition using training sample data of the discrete analog method to step S1, seek time-frequency matrix,
And then the picture element matrix of time-frequency spectrum is constructed, and construct time-frequency spectrum, the input as subsequent CDBN model；
Step S3: building CDBN model, two convolution of training are limited Boltzmann machine in a manner of unsupervised learning, at the 2nd
Softmax classifier is added after CRBM, training network model realizes the effective extraction and automatic classification of fault signature；
Step S4: time-frequency spectrum is constructed with the step S1 test sample obtained and step S2, step S3 is input to and has trained
In kind CDBN model, distribution network failure classification is realized.
2. a kind of distribution network failure classification method using convolution depth confidence network according to claim 1, feature
It is: step S1 specifically: establish a radiant type medium-voltage distribution pessimistic concurrency control, 10 kinds of failures are obtained by the model, front and back occurs respectively
The simulation waveform of the electrical quantity including three-phase voltage, residual voltage and three-phase current of 1 frequency cycle.
3. a kind of distribution network failure classification method using convolution depth confidence network according to claim 2, feature
Be: 10 kinds of failures include A, B, C singlephase earth fault, AB, AC, BC double earthfault, and AB, AC, BC two-phase are alternate
Short trouble and ABC three-phase phase fault.
4. a kind of distribution network failure classification method using convolution depth confidence network according to claim 1, feature
Be: step S2 specifically includes the following steps:
Step S21: choosing the good db4 of regularity as wavelet basis function, and Decomposition order is 4 layers, to selected voltage and electricity
It flows signal and carries out Time-frequency Decomposition；
Step S22: being reconstructed the 4th layer of wavelet coefficient, respectively obtains the low frequency part S of signal4,0And high frequency section
S4,1-S4,15；
Step S23: S is chosen4,0-S4,4As the frequency range of signal analysis, time-frequency matrix is constructed；
Step S24: the time-frequency matrix obtained after discrete analog method is changed into 35 × 380 time-frequency spectrum.
5. a kind of distribution network failure classification method using convolution depth confidence network according to claim 1, feature
Be: in step S3, the CDBN model uses 7 model of a layered structure, comprising: visible layer 1, hidden layer 1, pond layer 1, visible layer
2, hidden layer 2, pond layer 2, full articulamentum and output layer, wherein pond layer 1 and visible layer 2 are same layer；The CRBM mould
Type follows greedy order training method principle, and using the operation of convolution sum pondization to obtain, level is deeper, the stronger failure of ability to express
Feature；
By convolution operation, K characteristic pattern is generated as next layer of input to extract more advanced feature, kth in hidden layer
The hiding layer unit of a characteristic pattern the i-th row jth columnCondition activate probability are as follows:
In formula, v represents visible layer, bkIndicate the biasing of k-th of characteristic pattern,Indicate k-th of feature in visible layer and hidden layer
The interlayer weight of figure, σ=1/ (1+e-x) it is Sigmoid activation primitive；
Pond is the calculating of hidden layer to pond layer, down-sampled to hidden layer progress using maximum probability pond method, not weigh
Hidden layer is divided into several B by folded windowαRegion；Maximum value in selection area is taken to obtain as output valve by severalArea
Input of the pond layer feature of domain composition as next CRBM；The condition of k-th of pond layer unit activates probability are as follows:
The characteristic pattern that CRBM is extracted successively is pressed into column expansion, a feature vector is obtained, maps that sample labeling space, obtain
The feature representation form new to one calculates current each sample in the corresponding probability of each type are as follows:
T (i)=Y (li)；
In formula, Y indicates soft-max classifier functions, liNumber indicates the calculated result of the i-th class input data；Maximum value in T (i)
The number of position is the label of fault type.
6. a kind of distribution network failure classification method using convolution depth confidence network according to claim 5, feature
Be: step S4 specifically includes the following steps:
Step S41: in the time-frequency spectrum of 1 input fault waveform of visible layer, carrying out convolution operation to input picture using convolution kernel,
Fault signature figure is exported in hidden layer 1；
Step S42: input picture of the output image of hidden layer 1 as pond layer 1 carries out maximum pondization operation, passes through maximum
Pond layer 1 is extracted the integration that obtained feature carries out the adjacent similar features of regional area, reaches and be effectively reduced by the mode in pond
The purpose of intrinsic dimensionality；
Step S43: the output image of pond layer 1 is sent into hidden layer 2 and carries out the second wheel convolution operation, so as to mention front
The feature got further is abstracted, and feature more advanced, that discrimination is bigger is obtained；
Step S44: the characteristic pattern that hidden layer 2 exports needs to carry out primary maximum pondization operation using pond layer 2, compares pond
The characteristic pattern of 1 output of layer is more regular, this is the feature extraction more profound to original image, characterizes the intrinsic of original image
Feature obtains advanced features figure；
Step S45: advanced features figure is successively unfolded by column in full articulamentum and is stacked into feature vector；
Step S46: feature vector carries out connecting operation entirely with output layer in full articulamentum, and final output differentiates result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811029500.2A CN109325526A (en) | 2018-09-05 | 2018-09-05 | A kind of distribution network failure classification method using convolution depth confidence network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811029500.2A CN109325526A (en) | 2018-09-05 | 2018-09-05 | A kind of distribution network failure classification method using convolution depth confidence network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109325526A true CN109325526A (en) | 2019-02-12 |
Family
ID=65263720
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811029500.2A Pending CN109325526A (en) | 2018-09-05 | 2018-09-05 | A kind of distribution network failure classification method using convolution depth confidence network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109325526A (en) |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059555A (en) * | 2019-03-14 | 2019-07-26 | 中国人民解放军战略支援部队信息工程大学 | Based on the signal modulation classifying identification methods and device for differentiating spectrogram watershed image segmentation and decision tree more |
| CN110119455A (en) * | 2019-04-23 | 2019-08-13 | 西安理工大学 | A kind of image classification search method based on convolution depth confidence network |
| CN110222610A (en) * | 2019-05-27 | 2019-09-10 | 青岛海信网络科技股份有限公司 | A kind of method and device of Modulation recognition |
| CN110320436A (en) * | 2019-07-06 | 2019-10-11 | 河南理工大学 | Flexible direct current power distribution network high resistance earthing fault detection method based on color relationship classifier |
| CN110417005A (en) * | 2019-07-23 | 2019-11-05 | 清华大学 | In conjunction with the transient stability catastrophe failure screening technique of deep learning and simulation calculation |
| CN110490218A (en) * | 2019-06-10 | 2019-11-22 | 内蒙古工业大学 | A kind of rolling bearing fault self-learning method based on two-stage DBN |
| CN110543921A (en) * | 2019-10-14 | 2019-12-06 | 四川大学 | cable early fault identification method based on waveform learning |
| CN110705373A (en) * | 2019-09-10 | 2020-01-17 | 中南大学 | Tank voltage signal classification method based on wavelet packet analysis and residual error network |
| CN110852295A (en) * | 2019-10-15 | 2020-02-28 | 深圳龙岗智能视听研究院 | Video behavior identification method based on multitask supervised learning |
| CN110991240A (en) * | 2019-10-31 | 2020-04-10 | 湖南世优电力科技股份有限公司 | Single-phase fault identification and classification method for low-voltage neutral point ungrounded system |
| CN111160241A (en) * | 2019-12-27 | 2020-05-15 | 华中科技大学 | Power distribution network fault classification method, system and medium based on deep learning |
| CN111709447A (en) * | 2020-05-14 | 2020-09-25 | 中国电力科学研究院有限公司 | Power grid abnormality detection method and device, computer equipment and storage medium |
| CN112177865A (en) * | 2020-12-02 | 2021-01-05 | 南京智谷人工智能研究院有限公司 | Method for solving marking noise and insufficient marks in fan fault detection |
| CN112180217A (en) * | 2020-09-30 | 2021-01-05 | 福州大学 | Power distribution network single-phase earth fault section positioning method and system |
| CN112240964A (en) * | 2019-07-16 | 2021-01-19 | 北京映翰通网络技术股份有限公司 | Method for identifying fault type of power distribution network |
| CN112327165A (en) * | 2020-09-21 | 2021-02-05 | 电子科技大学 | Battery SOH prediction method based on unsupervised transfer learning |
| CN113673630A (en) * | 2021-09-15 | 2021-11-19 | 国网能源研究院有限公司 | Power distribution network fault classification method, system and device |
| CN114910742A (en) * | 2022-05-05 | 2022-08-16 | 湖南腾河智慧能源科技有限公司 | Single-phase fault grounding monitoring method and system, electronic equipment and storage medium |
| CN116184124A (en) * | 2023-04-26 | 2023-05-30 | 华东交通大学 | Power distribution network fault type identification method |
| CN117314883A (en) * | 2023-10-27 | 2023-12-29 | 国网四川省电力公司电力科学研究院 | Power distribution network fault line selection method and system based on EWT and VGGNet |
| CN118171049A (en) * | 2024-05-13 | 2024-06-11 | 西南交通大学 | Big data-based battery management method and system for edge calculation |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106226635A (en) * | 2016-07-14 | 2016-12-14 | 国网福建晋江市供电有限公司 | A kind of distribution feeder fault type recognition method and device |
| CN106680576A (en) * | 2017-01-11 | 2017-05-17 | 福州大学 | Power distribution network internal overvoltage identification method based on partitioned time-frequency spectrum and deep learning algorithm |
| CN107167702A (en) * | 2017-05-04 | 2017-09-15 | 国网福建省电力有限公司 | A kind of distribution feeder fault type recognition method and device |
-
2018
- 2018-09-05 CN CN201811029500.2A patent/CN109325526A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106226635A (en) * | 2016-07-14 | 2016-12-14 | 国网福建晋江市供电有限公司 | A kind of distribution feeder fault type recognition method and device |
| CN106680576A (en) * | 2017-01-11 | 2017-05-17 | 福州大学 | Power distribution network internal overvoltage identification method based on partitioned time-frequency spectrum and deep learning algorithm |
| CN107167702A (en) * | 2017-05-04 | 2017-09-15 | 国网福建省电力有限公司 | A kind of distribution feeder fault type recognition method and device |
Non-Patent Citations (4)
| Title |
|---|
| HONGLAK LEE 等,: "Unsupervised learning of hierarchical representations with convolutional deep belief networks", 《COMMUNICATIONS OF THE ACM》 * |
| 何灼彬，: "基于卷积深度置信网络的歌手识别", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
| 王林 等，: "卷积深度置信网络的场景文本检测", 《计算机系统应用》 * |
| 童利标 等，: "《无线传感器网络与信息融合》", 30 June 2008, 安徽人民出版社 * |
Cited By (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059555A (en) * | 2019-03-14 | 2019-07-26 | 中国人民解放军战略支援部队信息工程大学 | Based on the signal modulation classifying identification methods and device for differentiating spectrogram watershed image segmentation and decision tree more |
| CN110059555B (en) * | 2019-03-14 | 2021-04-20 | 中国人民解放军战略支援部队信息工程大学 | Signal modulation classification identification method and device based on multi-resolution spectrogram watershed image segmentation and decision tree |
| CN110119455A (en) * | 2019-04-23 | 2019-08-13 | 西安理工大学 | A kind of image classification search method based on convolution depth confidence network |
| CN110222610A (en) * | 2019-05-27 | 2019-09-10 | 青岛海信网络科技股份有限公司 | A kind of method and device of Modulation recognition |
| CN110490218A (en) * | 2019-06-10 | 2019-11-22 | 内蒙古工业大学 | A kind of rolling bearing fault self-learning method based on two-stage DBN |
| CN110490218B (en) * | 2019-06-10 | 2022-11-29 | 内蒙古工业大学 | Rolling bearing fault self-learning method based on two-stage DBN |
| CN110320436A (en) * | 2019-07-06 | 2019-10-11 | 河南理工大学 | Flexible direct current power distribution network high resistance earthing fault detection method based on color relationship classifier |
| CN112240964B (en) * | 2019-07-16 | 2023-06-20 | 北京映翰通网络技术股份有限公司 | Method for identifying fault type of power distribution network |
| CN112240964A (en) * | 2019-07-16 | 2021-01-19 | 北京映翰通网络技术股份有限公司 | Method for identifying fault type of power distribution network |
| CN110417005A (en) * | 2019-07-23 | 2019-11-05 | 清华大学 | In conjunction with the transient stability catastrophe failure screening technique of deep learning and simulation calculation |
| CN110705373A (en) * | 2019-09-10 | 2020-01-17 | 中南大学 | Tank voltage signal classification method based on wavelet packet analysis and residual error network |
| CN110705373B (en) * | 2019-09-10 | 2022-12-02 | 中南大学 | Tank voltage signal classification method based on wavelet packet analysis and residual error network |
| CN110543921A (en) * | 2019-10-14 | 2019-12-06 | 四川大学 | cable early fault identification method based on waveform learning |
| CN110852295A (en) * | 2019-10-15 | 2020-02-28 | 深圳龙岗智能视听研究院 | Video behavior identification method based on multitask supervised learning |
| CN110852295B (en) * | 2019-10-15 | 2023-08-25 | 深圳龙岗智能视听研究院 | Video behavior recognition method based on multitasking supervised learning |
| CN110991240B (en) * | 2019-10-31 | 2023-06-09 | 湖南世优电力科技股份有限公司 | Single-phase fault identification and classification method for low-voltage neutral point ungrounded system |
| CN110991240A (en) * | 2019-10-31 | 2020-04-10 | 湖南世优电力科技股份有限公司 | Single-phase fault identification and classification method for low-voltage neutral point ungrounded system |
| CN111160241B (en) * | 2019-12-27 | 2022-08-12 | 华中科技大学 | Power distribution network fault classification method, system and medium based on deep learning |
| CN111160241A (en) * | 2019-12-27 | 2020-05-15 | 华中科技大学 | Power distribution network fault classification method, system and medium based on deep learning |
| CN111709447A (en) * | 2020-05-14 | 2020-09-25 | 中国电力科学研究院有限公司 | Power grid abnormality detection method and device, computer equipment and storage medium |
| CN112327165A (en) * | 2020-09-21 | 2021-02-05 | 电子科技大学 | Battery SOH prediction method based on unsupervised transfer learning |
| CN112327165B (en) * | 2020-09-21 | 2021-07-13 | 电子科技大学 | Battery SOH prediction method based on unsupervised transfer learning |
| CN112180217A (en) * | 2020-09-30 | 2021-01-05 | 福州大学 | Power distribution network single-phase earth fault section positioning method and system |
| CN112177865A (en) * | 2020-12-02 | 2021-01-05 | 南京智谷人工智能研究院有限公司 | Method for solving marking noise and insufficient marks in fan fault detection |
| CN113673630A (en) * | 2021-09-15 | 2021-11-19 | 国网能源研究院有限公司 | Power distribution network fault classification method, system and device |
| CN114910742A (en) * | 2022-05-05 | 2022-08-16 | 湖南腾河智慧能源科技有限公司 | Single-phase fault grounding monitoring method and system, electronic equipment and storage medium |
| CN114910742B (en) * | 2022-05-05 | 2024-05-28 | 湖南腾河智慧能源科技有限公司 | Single-phase fault grounding monitoring method and monitoring system, electronic equipment and storage medium |
| CN116184124A (en) * | 2023-04-26 | 2023-05-30 | 华东交通大学 | Power distribution network fault type identification method |
| CN116184124B (en) * | 2023-04-26 | 2023-07-07 | 华东交通大学 | Power distribution network fault type identification method |
| CN117314883A (en) * | 2023-10-27 | 2023-12-29 | 国网四川省电力公司电力科学研究院 | Power distribution network fault line selection method and system based on EWT and VGGNet |
| CN117314883B (en) * | 2023-10-27 | 2024-04-16 | 国网四川省电力公司电力科学研究院 | Power distribution network fault line selection method and system based on EWT and VGGNet |
| CN118171049A (en) * | 2024-05-13 | 2024-06-11 | 西南交通大学 | Big data-based battery management method and system for edge calculation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109325526A (en) | A kind of distribution network failure classification method using convolution depth confidence network | |
| CN108279364B (en) | Power distribution network single-phase earth fault line selection method based on convolutional neural network | |
| CN108107324A (en) | A kind of electrical power distribution network fault location method based on depth convolutional neural networks | |
| CN107909118A (en) | A kind of power distribution network operating mode recording sorting technique based on deep neural network | |
| CN103728535B (en) | A kind of extra-high-voltage direct-current transmission line fault location based on wavelet transformation transient state energy spectrum | |
| CN106680576B (en) | Power distribution network internal overvoltage recognition methods based on piecemeal time-frequency spectrum and deep learning algorithm | |
| CN109271975A (en) | A kind of electrical energy power quality disturbance recognition methods based on big data multi-feature extraction synergetic classification | |
| CN108089099A (en) | The diagnostic method of distribution network failure based on depth confidence network | |
| CN106203382A (en) | A kind of excitation surge current based on kernel function extreme learning machine and fault current recognition methods | |
| CN109142851B (en) | Novel power distribution network internal overvoltage identification method | |
| CN108154223A (en) | Power distribution network operating mode recording sorting technique based on network topology and long timing information | |
| CN102510044A (en) | Excitation inrush current identification method based on wavelet transformation and probabilistic neural network (PNN) | |
| CN112240965A (en) | Grounding line selection device and method based on deep learning algorithm | |
| CN109597396B (en) | A kind of distribution transforming on-line fault diagnosis method based on high amount of traffic and transfer learning | |
| CN106547990A (en) | The modeling method of the integrated wideband equivalent-circuit model of extra-high voltage direct current converter valve tower valve layer | |
| CN107589342A (en) | A kind of one-phase earthing failure in electric distribution network localization method and system | |
| CN109858503A (en) | The traction converter failure diagnostic method of decision tree is promoted based on gradient | |
| CN107765139A (en) | A kind of resonant earthed system fault line selection method for single-phase-to-ground fault of high-accuracy | |
| Cheng et al. | Fault line selection of distribution network based on modified CEEMDAN and GoogLeNet neural network | |
| CN109596945B (en) | Novel power distribution network fault line selection method based on correlation coefficient vector similarity degree | |
| CN110058110A (en) | A kind of active inverter intermittent fault diagnostic method | |
| CN110532596A (en) | A kind of method and system generating grid simulation sample based on LSTM algorithm | |
| CN107462810A (en) | A kind of fault section location method suitable for active power distribution network | |
| CN109541395A (en) | Distribution Section Location based on Hausdorff under a kind of feature band | |
| Guo et al. | High-impedance fault detection methodology using time–frequency spectrum and transfer convolutional neural network in distribution networks |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190212 |
|
| RJ01 | Rejection of invention patent application after publication |