CN110069886A - The identification of cable initial failure and classification method based on VMD and CNN - Google Patents
The identification of cable initial failure and classification method based on VMD and CNN Download PDFInfo
- Publication number
- CN110069886A CN110069886A CN201910367000.8A CN201910367000A CN110069886A CN 110069886 A CN110069886 A CN 110069886A CN 201910367000 A CN201910367000 A CN 201910367000A CN 110069886 A CN110069886 A CN 110069886A
- Authority
- CN
- China
- Prior art keywords
- signal
- initial failure
- cnn
- vmd
- identification
- 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
- G06F18/21—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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Complex Calculations (AREA)
Abstract
The identification of cable initial failure and classification method based on VMD and CNN, comprising the following steps: Step 1: obtaining analog signal to be measured；Step 2: choosing bandwidth limiting factors α, noise margin τ and mode decomposition number K as parameter and is arranged parameter value；Step 3: carrying out variation mode decomposition to all kinds of analog signals, each mode and its centre frequency are obtained, realizes that frequency band divides；Modal characteristics and construction feature vector are decomposed Step 4: extracting；Step 5: various types of signal feature vector is inputted convolutional neural networks, tune, which is participated in training, to be practiced and obtains classification results；By using this method, accurately cable initial failure and overcurrent disturbance can be distinguished, cable maintenance is completed in time before initial failure becomes permanent fault, maintain the stable operation of power grid.
Description
Technical field
The present invention relates to cable initial failure identification technology fields, specifically the cable early stage event with a kind of based on VMD and CNN
Barrier identification and classification method.
Background technique
The development process of the capital equipment that cable is transmitted as power system information, failure is generally divided into three phases:
Shelf depreciation phase, early fault period and permanent fault phase.In the use process of cable, due to the defect of insulating layer, corrosion or
There is a series of partial discharge pulse at first in aging, forms electric branch or water tree, deteriorates with further, will develop
For the initial failure occurred with electric arc；Initial failure can occur repeatedly after occurring for the first time, until becoming irreversible permanent
Property failure.The generation of cable initial failure has uncertainty, and electric current is very small when failure occurs, so being not enough to cause
The safeguard protection of traditional overcurrent detecting device.Simultaneously as there are certain similar disturbances, for example, transformer excitation shove,
Overcurrent problem caused by constant impedance and capacitor switching failure, presently, there are most of cable fault EARLY RECOGNITION method tool
It has ready conditions limitation, and discrimination is not high, under the influence of overcurrent disturbance, cable initial failure and overcurrent disturb can not be accurate
Division identification.
Summary of the invention
Present invention aims at solve the problems, such as cable initial failure and overcurrent disturbance can not accurately distinguish identification, provide
The identification of cable initial failure and classification method based on VMD and CNN can be accurately to cable morning by using this method
Phase failure is distinguished with overcurrent disturbance, is completed cable maintenance in time before initial failure becomes permanent fault, is maintained power grid
Stable operation.
The present invention is achieved through the following technical solutions:
The identification of cable initial failure and classification method based on VMD and CNN, comprising the following steps:
Step 1: obtaining analog signal to be measured；
Step 2: choosing bandwidth limiting factors α, noise margin τ and mode decomposition number K as parameter and parameter is arranged taking
Value；
Step 3: carrying out variation mode decomposition to all kinds of analog signals, each mode and its centre frequency are obtained, realizes frequency
Band divides；
Modal characteristics and construction feature vector are decomposed Step 4: extracting；
Step 5: various types of signal feature vector is inputted convolutional neural networks, tune, which is participated in training, to be practiced and obtains classification results.
Cable early stage self-cleaning failure is generally singlephase earth fault, easily causes mutually initial failure indirectly, typical fault
Type mainly includes half cycles initial failure and more cycle initial failures.Cable fault features are concluded are as follows: the duration is short
Or current amplitude is low, failure occurs to continue for 1/4 period in voltage peaks and half cycles initial failure, works as current zero-crossing point
When, failure disappears automatically；More cycle initial failures generally continue 1-4 period, and after electric arc disappearance, failure is eliminated automatically,
Although cable initial failure has had obvious feature, if the effect is unsatisfactory by primary fault Direct Recognition, the present invention
The middle method using variation mode decomposition assumes that each mode decomposed is the finite bandwidth with corresponding centre frequency, will be optimal
Resolution problem is converted to variational methods problem, solves each mode and its centre frequency using alternating direction multipliers method, effectively
It realizes that frequency band divides, compares some other decomposition algorithm, variation mode decomposition has more solid theoretical foundation and robustness, becomes
Divide mode decomposition method successively to separate input signal from low to high, can effectively distinguish cable initial failure signal and phase
Like the characteristic information of disturbance different frequency range；Then according to signal characteristic complete parameter setting, but decompose after multilayer signal
With very big data volume, convolutional neural networks identification is directly inputted, not only network parameter selection difficulty increases, Er Qiehui
Cause the training time too long, in the present invention convolutional neural networks is inputted later to mode signals progress feature extraction again and known
Not；The multilayered structure feature that convolutional neural networks local sensing and weight are shared, makes it in variation mode decomposition feature extraction
On the basis of, secondary deep feature learning has also been carried out, more valuable information is excavated, has improved the accurate of cable initial failure classification
Rate, and then identification effectively is distinguished to cable initial failure and overcurrent disturbance, before initial failure becomes permanent fault
Cable maintenance is completed in time, maintains the stable operation of power grid；VMD is that the English of variation mode decomposition is write a Chinese character in simplified form in the present invention, and CNN is
The English of convolutional neural networks is write a Chinese character in simplified form.
Further, parameter value is respectively bandwidth limiting factors α=2000 in the step 2, noise margin τ=0,
Mode decomposition number selects K=7.Bandwidth limiting factors α is the parameter of an influence decomposed signal bandwidth, with the increase of α, often
A decomposed signal is on the basis of centre frequency, and the frequency on both sides can decay faster, and decomposed signal bandwidth is smaller；Conversely, α is smaller,
Signal decaying in centre frequency both sides is slower, and decomposed signal bandwidth is bigger.Therefore, when input signal frequency range is very big, α value
It is smaller, it should be near hundreds of；When frequency input signal very concentrates range smaller, α value should become larger, near tens of thousands of,
Shove disturbance, constant impedance and capacitor of half cycles initial failure, more cycle initial failures, transformer excitation is thrown by inventor
It cuts the 5 class signal spectrum such as failure to be analyzed, finds frequency distribution between 0~380Hz, feature is that frequency range is larger, low
Frequency signal content is more, and when bandwidth limiting factors α=2000, signal characteristic abstraction effect is preferable；Noise margin τ, works as input signal
When containing very noisy, which is able to achieve the effect of denoising, and by the research of inventor, when τ=0 is able to satisfy decomposition and requires；Mould
State decompose number K, when decomposing to signal, K value is too small, and Decomposition order is very little, cannot be fitted completely input signal when
Frequency feature, K value is excessive, and Decomposition order is too many, interference signal excessive decomposition can be caused mode centre frequency aliasing.By hair
Bright people analyzes 5 class signals, and Decomposition order is optimal when K value is 7.
Further, to original signal progress variation mode decomposition, detailed process is as follows in the step 3:
Step 3.1, by input signal x (t) predecomposition be K mode function uk(t), Hilbert is carried out to each mode
Transformation, by uk(t) real signal becomes analytic signal:
Wherein δ (t) is Dirac function, and j indicates imaginary number, ukIt (t) is k-th of modal components, * indicates convolution algorithm；
Step 3.2 is estimated on each mode parsing signal center frequency and modulation spectrum to corresponding Base Band, realizes frequency
Rate mixing:
ω in formulakFor the centre frequency of k-th of modal components；
The bandwidth of step 3.3, each modal components of estimation, and the sum of estimation bandwidth for meeting each mode minimum, introduce
The L of constraint condition calculation formula (2) demodulated signal gradient2Norm, form are as follows:
In formula, { uk}={ u1,u2…ukIndicate K modal components, { ωk}={ ω1,ω2…ωkIndicate K component
Centre frequency；
Step 3.4 introduces Lagrange multiplier λ (t) and penalty factor α, Augmented Lagrangian Functions is constructed, by formula
(3) non-binding variational problem is converted to, form is as follows:
Step 3.5, model solution, process are as follows: alternately being updated by continuous iterationWithTo acquire
The saddle point of above-mentioned Lagrange formula (4), it is rightCarry out Fourier transformation, the solution of double optimizationIt may be expressed as:
Similarly obtain the more new formula of centre frequency and Lagrange multiplier:
Herein, τ indicates time step, rebuilds and constrains as noise；
Formula (5) meetsCondition, iterative process terminate, and obtainWithIt is rightFourier's inversion is carried out, real part is the modal components u of forms of time and spacek(t)。
Further, the decomposition modal characteristics that the step 4 is extracted are peak-to-peak value, root mean square, centre frequency, zero crossing
Number, mode relative energy ratio and instantaneous amplitude, in 1~K mode, the feature vector of any mode is configured to FVk=[peak peak
Value, root mean square, zero passage points, mode relative energy ratio, instantaneous amplitude, centre frequency], K modal vector is end to end, each
Signal obtains the one-dimensional vector F of 1 × 6 × K.Peak-to-peak value describes the size of signal value floating range in the present invention, it is selected to make
It is characterized and mainly identifies cable initial failure from the signal amplitude of different modalities；Root mean square, that is, virtual value, to measure one
The size of signal in a period, therefore this feature can separate multicycle fault-signal and other types signal.Centre frequency
As the important frequency-domain index of signal decomposition, it can be good at the frequency composition for reacting different disturbing signals；Zero passage points are used for
Distinguish the non-stationary property of the signal under different center frequency mode；Mode relative energy ratio describe each decomposition mode for
The contribution rate of entire signal；Instantaneous amplitude calculates the amplitude envelope of signal by mobile fixed window width, is short for distinguishing signal
When or sustained fault；Important frequency-domain index of the centre frequency as signal decomposition, can be good at reacting different disturbing signals
Frequency composition.Variation mode decomposition method successively separates input signal from low to high, has efficiently differentiated cable early stage
The characteristic information of fault-signal and similar disturbance different frequency range, however the multilayer signal after decomposing has bigger data volume, such as
Fruit is directly inputted convolutional neural networks identification, and not only network parameter selection difficulty increases, but also will lead to training time mistake
It is long, it is excessive that redundancy can effectively be reduced to the feature extraction of variation mode decomposition, solve the problems, such as that computational efficiency is low.
Further, the convolutional neural networks in the step 5 include input layer, convolutional layer, down-sampling layer, full connection
Layer and output layer；Input layer obtains one-dimensional vector F information, and convolutional layer obtains the depth characteristic mapping of input signal, down-sampling layer
The characteristic information generated to convolution is extracted and is filtered, and all kinds of probability are spliced and calculated to multiple feature vectors by full articulamentum,
Finally reach output layer.Convolutional neural networks are a kind of deep learning models with special networks structure, copy the view of biology
Perceptual pattern construction, using in hidden layer convolution kernel parameter sharing and interlayer partially connected reduce the calculating in learning process
Amount, furthermore down-sampling layer carries out secondary deep feature learning, excavates more valuable information, can more accurately realize cable morning
Phase failure modes.
Further, two convolutional layers and two down-sampling layers are alternately present in the convolutional neural networks structure.
Further, the convolutional layer formula expression are as follows:
F indicates the feature vector of input, and p indicates that the convolution kernel having a size of 1 × G, b are bias, and C is convolutional layer output knot
Fruit；Wherein, 1≤g≤G, n=6 × K-g+1；；
It helps to express more complicated Feature Mapping using activation primitive after convolution, the whole process of convolution sum activation is such as
Under:
Wherein, f is activation primitive, is indicated are as follows:
Wherein, Cl jIt is l layers of jth width Feature Mapping as a result, FlAnd blRespectively l layers of input signal and biasing
Value,For the weight coefficient of j-th of convolution kernel of (l+1) layer.
Convolutional layer can be realized the feature extraction to input signal, and internal includes many convolution kernels, and forms convolution kernel
Each element have corresponding weight coefficient and deviation again, be equivalent to the neuron of feedforward neural network, convolution kernel is according to one
Input signal that fixed step size is inswept completes convolution algorithm in the corresponding every part of convolution kernel and is superimposed deviation, each convolution kernel
The input as next layer of neural network is exported with convolution results currently entered, activation primitive is normally used for after convolution,
It can be used to assist to express more complicated Feature Mapping.
Further, the down-sampling layer formula expression are as follows:
Assuming that the size of each Feature Mapping C of above formula is 1 × M, the size in down-sampling region is 1 × d, wherein
D is that down-sampling layer exports result.
Down-sampling layer has used mean value pond, and the texture information of signal characteristic is remained with sacrificial features figure size.
Further, the output head and the tail of down-sampling layer are spliced into one-dimensional row vector by the full articulamentum.
Further, the output layer obtains the label of classification using normalization exponential function softmax.
Compared with prior art, the present invention having the following advantages and benefits: using variation mode point in the present invention
Solving feature extracting method can be from the different characteristic information successfully extracted under similar signal in original signal, while to variation
Mode decomposition structure, which carries out feature extraction, can effectively reduce that redundancy is excessive, the low problem of computational efficiency；Finally roll up
The multilayered structure feature that product neural network local sensing and weight are shared, makes it on the basis of variation mode decomposition feature extraction
On, secondary deep feature learning has also been carried out, more valuable information is excavated, has improved the accuracy rate of cable initial failure classification,
And then identification effectively is distinguished to cable initial failure and overcurrent disturbance, before initial failure becomes permanent fault in time
Cable maintenance is completed, the stable operation of power grid is maintained.
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 flow chart based on VMD Yu CNN algorithm；
Fig. 2 is half cycles cable initial failure, more cycle cable initial failures, normal signal, transformer excitation shove and disturb
Dynamic, the disturbance of constant impedance failure, capacitor switching current waveform figure；
Fig. 3 is a) multicycle cable initial failure and b) the VMD result and spectrogram of inrush current of transformer interference；
Fig. 4 is the display diagram of 7 decomposition modal characteristics vectors；
Fig. 5 is the CNN structure chart that the present invention uses；
Fig. 6 is influence diagram of the CNN repetitive exercise to nicety of grading.
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 1:
As shown in Figure 1, the identification of cable initial failure and classification method based on VMD and CNN, comprising the following steps:
Step 1: obtaining analog signal to be measured；
Step 2: choosing bandwidth limiting factors α, noise margin τ and mode decomposition number K as parameter and parameter is arranged taking
Value；
Step 3: carrying out variation mode decomposition to all kinds of analog signals, each mode and its centre frequency are obtained, realizes frequency
Band divides；
Modal characteristics and construction feature vector are decomposed Step 4: extracting；
Step 5: various types of signal feature vector is inputted convolutional neural networks, tune, which is participated in training, to be practiced and obtains classification results.
Half cycles cable initial failure, more cycle cable initial failures, normal signal, constant impedance are chosen in the present embodiment
Failure, capacitor switching and transformer excitation shove disturbance as analog signal.
6 kinds of analog signals to be measured pass through PSCARD/EMTDC software emulation and obtain in the present embodiment.As shown in Fig. 2, Fig. 2
In (a) be half cycles cable initial failure, (b) be more cycle cable initial failures, (c) be normal signal, (d) for transformer swash
Magnetic shoves disturbance, (e) is constant impedance failure, (f) disturbs for capacitor switching.As seen from Figure 2, all sample abort situation
With randomness and uniformity signal waveform slightly has difference.
In order to make variation mode decomposition process preferably realize that parameter value is respectively in step 2 described in the present embodiment
Bandwidth limiting factors α=2000, noise margin τ=0, mode decomposition number select K=7.Bandwidth limiting factors α is an influence
The parameter of decomposed signal bandwidth, when input signal frequency range is very big, α value is smaller, should be near hundreds of；Work as input signal
When frequency very concentrates range smaller, α value should become larger, near tens of thousands of, by inventor to half cycles initial failure, more
Cycle initial failure, the transformer excitation 5 class signal spectrum such as disturbance, constant impedance and capacitor switching failure that shoves are analyzed,
It was found that frequency distribution, between 0~380Hz, feature is that frequency range is larger, and low frequency signal content is more, thus when bandwidth limitation because
When sub- α=2000, signal characteristic abstraction effect is preferable；Noise margin τ influences about beam intensity of the lagrangian multiplier when rebuilding
Degree.In general, 0 can be set to if not needing Exact Reconstruction under strong noise.By the research of inventor, τ=0
When be able to satisfy herein decompose require；Mode decomposition number K, when decomposing to signal, K value is too small, and Decomposition order is very little, no
It can be fitted the time-frequency characteristics of input signal completely；K value is excessive, and Decomposition order is too many, interference signal excessive decomposition can be caused mould
State centre frequency aliasing.By inventor to cable initial failure and overcurrent disturbing signal using centre frequency observation point
Analysis, decomposition result is more excellent when K value is 7.
Carrying out variation mode decomposition to original signal in the present embodiment, in the step 3, detailed process is as follows:
Step 3.1, by input signal x (t) predecomposition be K=7 mode function uk(t) (k=1 ..., 7), to each mould
State carries out Hilbert transformation, by uk(t) real signal becomes analytic signal:
Wherein δ (t) is Dirac function, and j indicates imaginary number, ukIt (t) is k-th of modal components, * indicates convolution algorithm；
Step 3.2 is estimated on each mode parsing signal center frequency and modulation spectrum to corresponding Base Band, realizes frequency
Rate mixing:
ω in formulakFor the centre frequency of k-th of modal components；
The bandwidth of step 3.3, each modal components of estimation, and the sum of estimation bandwidth for meeting each mode minimum, introduce
The L of constraint condition calculation formula (2) demodulated signal gradient2Norm, form are as follows:
In formula, { uk}={ u1,u2…u7Indicate 7 modal components, { ωk}={ ω1,ω2…ω7Indicate 7 mode point
The centre frequency of amount；
Step 3.4 introduces Lagrange multiplier λ (t) and penalty factor α, Augmented Lagrangian Functions is constructed, by formula
(3) non-binding variational problem is converted to, form is as follows:
Step 3.5, model solution, process are as follows: alternately being updated by continuous iterationWithTo acquire
The saddle point of above-mentioned Lagrange formula (4), it is rightCarry out Fourier transformation, the solution of double optimizationIt may be expressed as:
Similarly obtain the more new formula of centre frequency and Lagrange multiplier:
Herein, τ indicates time step, rebuilds and constrains as noise；
Formula (5) meetsCondition, iterative process terminate, and obtainWithIt is rightFourier's inversion is carried out, real part is the modal components u of forms of time and spacek(t)。
As shown in figure 3, showing a) multicycle cable initial failure in the present embodiment and b) transformer excitation shoves interference
The result and frequency spectrum of VMD, the analytic process of other signals are similar.Such as Fig. 3 as can be seen that corresponding to each pattern of unlike signal
Shape differences it is very big.
The feature extraction of variation mode decomposition can effectively reduce that redundancy is excessive, solve that computational efficiency is low to ask
Topic, the decomposition modal characteristics that step 4 described in the present embodiment is extracted are peak-to-peak value, root mean square, centre frequency, zero passage points, mould
State relative energy ratio and instantaneous amplitude, in 1~K mode, the feature vector of any mode be configured to FVk=[peak-to-peak value, just
Root, zero passage points, mode relative energy ratio, instantaneous amplitude, centre frequency], K modal vector is end to end, and each signal obtains
To the one-dimensional vector F of 1 × 6 × K, K=7 in the present embodiment, each signal obtains 1 × 42 one-dimensional vector F.
Peak-to-peak value (PTP): PTP=max (uk[n])-min(uk[n])
uk[n] indicates k-th of decomposition mode signals for containing N number of sampled point, wherein 1≤n≤N；
Size of the root mean square (RMS) to measure periodic signal；
Centre frequency (ωk): VMD process is directly realized by a frequency domain, therefore selects ωkAs feature, reduce to each
The demand of a frequency domain information extra computation for decomposing mode；
Zero passage points (ZC) are used to distinguish the non-stationary property of the signal under different center frequency mode；
The each mode of decomposing of mode relative energy ratio (MRER) description is for the contribution rate of entire signal；
Instantaneous amplitude (IA) is for distinguishing signal in short-term or sustained fault；
Herein, w is the window that length is 20；M=1,2 ... (N-w+1).
The feature of the 7 decomposition mode u1~u2 done based on features described above shown as shown in figure 4, as seen from Figure 4,
Under 7 mode, the same feature differentiation of unlike signal is more obvious in most of feature vectors.
Convolutional neural networks in step 5 described in the present embodiment include input layer, convolutional layer, down-sampling layer, full connection
Layer and output layer；Input layer obtains one-dimensional vector F information, and convolutional layer obtains the depth characteristic mapping of input signal, down-sampling layer
The characteristic information generated to convolution is extracted and is filtered, and all kinds of probability are spliced and calculated to multiple feature vectors by full articulamentum,
Finally reach output layer.Convolutional neural networks are a kind of deep learning models with special networks structure, copy the view of biology
Perceptual pattern construction, using in hidden layer convolution kernel parameter sharing and interlayer partially connected reduce the calculating in learning process
Amount, furthermore down-sampling layer carries out secondary deep feature learning, excavates more valuable information, can more accurately realize cable morning
Phase failure modes.
In the present embodiment, two convolutional layers and two down-sampling layers are alternately present in the convolutional neural networks structure.
In the present embodiment, the convolutional layer formula expression are as follows:
F indicates the feature vector of input, and p indicates that the convolution kernel having a size of 1 × G, b are bias, and C is convolutional layer output knot
Fruit；Wherein, 1≤g≤G, n=6 × 7-g+1；
It helps to express more complicated Feature Mapping using activation primitive after convolution, the whole process of convolution sum activation is such as
Under:
Wherein, f is activation primitive, is indicated are as follows:
Wherein, Cl jIt is l layers of jth width Feature Mapping as a result, FlAnd blRespectively l layers of input signal and biasing
Value,For the weight coefficient of j-th of convolution kernel of (l+1) layer.
Convolutional layer can be realized the feature extraction to input signal, and internal includes many convolution kernels, and forms convolution kernel
Each element have corresponding weight coefficient and deviation again, be equivalent to the neuron of feedforward neural network, convolution kernel is according to one
Input signal that fixed step size is inswept completes convolution algorithm in the corresponding every part of convolution kernel and is superimposed deviation, each convolution kernel
The input as next layer of neural network is exported with convolution results currently entered, activation primitive is normally used for after convolution,
It can be used to assist to express more complicated Feature Mapping.
In the present embodiment, the down-sampling layer formula expression are as follows:
Assuming that the size of each Feature Mapping C of above formula is 1 × M, the size in down-sampling region is 1 × d, wherein
D is that down-sampling layer exports result.
Down-sampling layer has used mean value pond, and the texture information of signal characteristic is remained with sacrificial features figure size.
In the present embodiment, the output head and the tail of down-sampling layer are spliced into one-dimensional row vector by the full articulamentum.
In the present embodiment, the output layer obtains the label of classification using normalization exponential function softmax.Softmax
Loss function may be expressed as:
pvIndicate scoring function, the i.e. probability that calculating sample x belongs to v class.Softmax function, which makes correctly to classify, to be obtained
Bigger probability makes the classification of mistake obtain smaller probability.Secondly sample is belonged in one kind of maximum probability, thus real
Now classify.
In the present embodiment, specific CNN structural model is as shown in figure 5, table 1 show CNN network architecture parameters.For
Training parameter in CNN mainly includes learning rate lr, each number of training Bs and each sample training times N e.Learning rate lr
As parameter important in deep learning, can which determine objective functions to converge to local minimum at the right time.
When lr setting it is too small when, convergence process will become very slowly.And when lr setting it is excessive when, gradient may be in minimum value
It nearby shakes back and forth, in some instances it may even be possible to can not restrain.In the present embodiment, lr of the inventor by the different organizational levels of setting, observation
Mean square error curve after comparing training, is finally arranged lr=1；The size of each number of training Bs determines gradient value, with
And the frequency that weight updates.When Bs is set as entire training sample set, although the gradient of backpropagation is very accurate, calculate
Time-consuming is easily trapped into local optimum.And when Bs is set as the sample number of very little, gradiometer is not calculated accurately really, will lead to network instruction
White silk is not restrained.The selection of each sample training times N e and the size of Bs are closely related, and when Ne is equal to 1, represent every batch of training
Sample is sent into the process that CNN completes a forward calculation and backpropagation.However in training, multiple sample set repetitive exercises one
It is secondary be it is inadequate, need repeated multiple times to be fitted convergence.By the experimental analysis of inventor, network query function efficiency is comprehensively considered
And precision, Bs=60, Ne=150 are finally set.After the completion of adjusting ginseng, entire training set iteration reaches 30000 times.Training precision is such as
Shown in Fig. 6.
Table 1.CNN structural model
Following part is the comparison and analysis of entire result of implementation.6 class sample of signal quantity are as shown in table 2:
The distribution of 2 sample size of table
In order to verify the validity of VMD characteristic extraction procedure, utilize in the present embodiment with structure C NN classifier, from the time
In the angle of accuracy rate, compare after feature of present invention extracts with original signal Direct Classification as a result, as shown in table 3, table 3 is
Classifying quality contrast table after VMD feature extraction.
Classifying quality contrast table after 3 VMD feature extraction of table
Although signal is classified after VMD feature extraction and is slightly below directly divided on classification accuracy as can be seen from Table 3
Class 2.5%, but significant shortening 2.2 hours in time, in comparison, VMD feature extracting method is in training speed and identification
There is certain application value in accuracy rate.
In order to verify the superiority for taking CNN classifier, by CNN classifier and decision tree, BP nerve net in the present embodiment
Four kinds of network, Bayes, support vector machines traditional classifier results are assessed (as shown in table 4):
4 different classifications device result of table compares
It can be obtained by table 4, decision tree, K- neighbour, BP neural network, support vector cassification accuracy rate be respectively 85.31%,
93.45%, 79.1%, 80.74%.Compared to other three kinds of traditional algorithms, K- nearest neighbour classification effect is more excellent, and reachable 93.45%,
But compared with context of methods, still there is certain gap.The same of signal extraction time can shortened using VMD and CNN in the present invention
When cable initial failure is accurately classified and is identified
Above-described 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 (10)
1. the identification of cable initial failure and classification method based on VMD and CNN, which comprises the following steps:
Step 1: obtaining analog signal to be measured；
Step 2: choosing bandwidth limiting factors α, noise margin τ and mode decomposition number K as parameter and is arranged parameter value；
Step 3: carrying out variation mode decomposition to all kinds of analog signals, each mode and its centre frequency are obtained, realizes that frequency band is drawn
Point；
Modal characteristics and construction feature vector are decomposed Step 4: extracting；
Step 5: various types of signal feature vector is inputted convolutional neural networks, tune, which is participated in training, to be practiced and obtains classification results.
2. the identification of cable initial failure and classification method based on VMD and CNN as described in claim 1, which is characterized in that institute
Stating parameter value in step 2 is respectively bandwidth limiting factors α=2000, noise margin τ=0, mode decomposition number selection K=
7。
3. the identification of cable initial failure and classification method based on VMD and CNN as described in claim 1, which is characterized in that institute
State in step 3 to original signal carry out variation mode decomposition detailed process is as follows:
Step 3.1, by input signal x (t) predecomposition be K mode function uk(t), Hilbert transformation is carried out to each mode,
By uk(t) real signal becomes analytic signal:
Wherein δ (t) is Dirac function, and j indicates imaginary number, ukIt (t) is k-th of modal components, * indicates convolution algorithm；
Step 3.2 is estimated on each mode parsing signal center frequency and modulation spectrum to corresponding Base Band, realizes that frequency is mixed
It closes:
ω in formulakFor the centre frequency of k-th of modal components；
The bandwidth of step 3.3, each modal components of estimation, and the sum of estimation bandwidth for meeting each mode minimum, introduce constraint
The L of condition calculation formula (2) demodulated signal gradient2Norm, form are as follows:
In formula, { uk}={ u1,u2…ukIndicate K modal components, { ωk}={ ω1,ω2…ωkIndicate K component center
Frequency；
Step 3.4 introduces Lagrange multiplier λ (t) and penalty factor α, constructs Augmented Lagrangian Functions, and formula (3) are turned
It is changed to non-binding variational problem, form is as follows:
Step 3.5, model solution, process are as follows: alternately being updated by continuous iterationWithIt is above-mentioned to acquire
The saddle point of Lagrangian formula (4), it is rightCarry out Fourier transformation, the solution of double optimizationIt may be expressed as:
Similarly obtain the more new formula of centre frequency and Lagrange multiplier:
Herein, τ indicates time step, rebuilds and constrains as noise；
Formula (5) meetsCondition, iterative process terminate, and obtainWithIt is rightFourier's inversion is carried out, real part is the modal components u of forms of time and spacek(t)。
4. the identification of cable initial failure and classification method based on VMD and CNN as described in claim 1, which is characterized in that institute
State step 4 extraction decomposition modal characteristics be peak-to-peak value, root mean square, centre frequency, zero passage points, mode relative energy ratio and
Instantaneous amplitude, in 1~K mode, the feature vector of any mode is configured to FVk=[peak-to-peak value, root mean square, zero passage points, mould
State relative energy ratio, instantaneous amplitude, centre frequency], K modal vector is end to end, and each signal obtains the one-dimensional of 1 × 6 × K
Vector F.
5. the identification of cable initial failure and classification method based on VMD and CNN as described in claim 1, which is characterized in that institute
Stating the convolutional neural networks in step 5 includes input layer, convolutional layer, down-sampling layer, full articulamentum and output layer；Input layer obtains
One-dimensional vector F information is taken, convolutional layer obtains the depth characteristic mapping of input signal, the characteristic information that down-sampling layer generates convolution
It extracts and filters, all kinds of probability are spliced and calculated to multiple feature vectors by full articulamentum, finally reaches output layer.
6. the identification of cable initial failure and classification method based on VMD and CNN as claimed in claim 5, which is characterized in that institute
Two convolutional layers and two down-sampling layers in convolutional neural networks structure are stated to be alternately present.
7. the identification of cable initial failure and classification method based on VMD and CNN as claimed in claim 6, which is characterized in that institute
State the expression of convolutional layer formula are as follows:
F indicates the feature vector of input, and p indicates that the convolution kernel having a size of 1 × G, b are bias, and C is that convolutional layer exports result；
Wherein, 1≤g≤G, n=6 × K-g+1；
It helps to express more complicated Feature Mapping using activation primitive after convolution, the whole process of convolution sum activation is as follows:
Wherein, f is activation primitive, is indicated are as follows:
Wherein,It is l layers of jth width Feature Mapping as a result, FlAnd blRespectively l layers of input signal and bias,
For the weight coefficient of j-th of convolution kernel of (l+1) layer.
8. the identification of cable initial failure and classification method based on VMD and CNN as claimed in claim 7, which is characterized in that institute
State the expression of down-sampling layer formula are as follows:
Assuming that the size of each Feature Mapping C of above formula is 1 × M, the size in down-sampling region is 1 × d, whereinD is
Down-sampling layer exports result.
9. the identification of cable initial failure and classification method based on VMD and CNN as claimed in claim 8, which is characterized in that institute
It states full articulamentum and the output head and the tail of down-sampling layer is spliced into one-dimensional row vector.
10. the identification of cable initial failure and classification method based on VMD and CNN as claimed in claim 9, which is characterized in that
The output layer obtains the label of classification using normalization exponential function softmax.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910367000.8A CN110069886A (en) | 2019-05-05 | 2019-05-05 | The identification of cable initial failure and classification method based on VMD and CNN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910367000.8A CN110069886A (en) | 2019-05-05 | 2019-05-05 | The identification of cable initial failure and classification method based on VMD and CNN |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110069886A true CN110069886A (en) | 2019-07-30 |
Family
ID=67369862
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910367000.8A Pending CN110069886A (en) | 2019-05-05 | 2019-05-05 | The identification of cable initial failure and classification method based on VMD and CNN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110069886A (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110686899A (en) * | 2019-09-21 | 2020-01-14 | 天津大学 | Internal combustion engine noise source identification method |
| CN110766264A (en) * | 2019-07-31 | 2020-02-07 | 贵州电网有限责任公司 | Transmission cable health state monitoring and evaluating method |
| CN111130053A (en) * | 2020-01-08 | 2020-05-08 | 华南理工大学 | Power distribution network overcurrent protection method based on deep reinforcement learning |
| CN111141519A (en) * | 2019-12-27 | 2020-05-12 | 西安交通大学 | Rolling bearing fault quantitative evaluation method |
| CN111259949A (en) * | 2020-01-13 | 2020-06-09 | 山东超越数控电子股份有限公司 | Fault identification model construction method, model and identification method for aircraft environmental control system |
| CN111413588A (en) * | 2020-03-31 | 2020-07-14 | 陕西省地方电力（集团）有限公司咸阳供电分公司 | Power distribution network single-phase earth fault line selection method |
| CN111504646A (en) * | 2020-06-05 | 2020-08-07 | 合肥工业大学 | Weak signal fault feature classification method and system for early failure of bearing |
| CN111664365A (en) * | 2020-06-07 | 2020-09-15 | 东北石油大学 | Oil and gas pipeline leakage detection method based on improved VMD and 1DCNN |
| CN112051480A (en) * | 2020-08-06 | 2020-12-08 | 国网安徽省电力有限公司电力科学研究院 | Neural network power distribution network fault diagnosis method and system based on variational modal decomposition |
| CN112101144A (en) * | 2020-08-27 | 2020-12-18 | 云南电网有限责任公司昭通供电局 | Self-adaptive method for improving transformer vibration signal processing precision |
| CN112580699A (en) * | 2020-12-04 | 2021-03-30 | 中国船舶重工集团公司第七一一研究所 | Sensor data anomaly detection method and application thereof |
| CN113093050A (en) * | 2021-03-31 | 2021-07-09 | 中国矿业大学 | Cable early fault identification method and system based on cable grounding wire current time-frequency characteristics |
| CN113203914A (en) * | 2021-04-08 | 2021-08-03 | 华南理工大学 | Underground cable early fault detection and identification method based on DAE-CNN |
| CN113740671A (en) * | 2020-10-22 | 2021-12-03 | 国网河北省电力有限公司雄安新区供电公司 | Fault arc identification method based on VMD and ELM |
| WO2022067562A1 (en) * | 2020-09-29 | 2022-04-07 | 西门子股份公司 | Method and device for diagnosing fault arc, and computer-readable storage medium |
| CN115166585A (en) * | 2022-09-08 | 2022-10-11 | 石家庄科林电气股份有限公司 | Ground fault detection anti-misjudgment method and device and electronic equipment |
| WO2022227491A1 (en) * | 2021-04-27 | 2022-11-03 | 京东方科技集团股份有限公司 | Display defect detection apparatus and detection method therefor, and display defect detection system |
| CN117929952A (en) * | 2024-03-21 | 2024-04-26 | 国网(山东)电动汽车服务有限公司 | Novel arc fault detection method for electric automobile charging pile |
| CN117929952B (en) * | 2024-03-21 | 2024-05-28 | 国网(山东)电动汽车服务有限公司 | Novel arc fault detection method for electric automobile charging pile |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060112028A1 (en) * | 2004-11-24 | 2006-05-25 | Weimin Xiao | Neural Network and Method of Training |
| CN108197648A (en) * | 2017-12-28 | 2018-06-22 | 华中科技大学 | A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models |
| CN108414226A (en) * | 2017-12-25 | 2018-08-17 | 哈尔滨理工大学 | Fault Diagnosis of Roller Bearings under the variable working condition of feature based transfer learning |
| CN108896296A (en) * | 2018-04-18 | 2018-11-27 | 北京信息科技大学 | A kind of wind turbine gearbox method for diagnosing faults based on convolutional neural networks |
| CN109633271A (en) * | 2019-01-17 | 2019-04-16 | 长沙理工大学 | The traveling wave Time-Frequency Analysis Method being distributed based on variation mode decomposition and Eugene Wigner Weir |
-
2019
- 2019-05-05 CN CN201910367000.8A patent/CN110069886A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060112028A1 (en) * | 2004-11-24 | 2006-05-25 | Weimin Xiao | Neural Network and Method of Training |
| CN108414226A (en) * | 2017-12-25 | 2018-08-17 | 哈尔滨理工大学 | Fault Diagnosis of Roller Bearings under the variable working condition of feature based transfer learning |
| CN108197648A (en) * | 2017-12-28 | 2018-06-22 | 华中科技大学 | A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models |
| CN108896296A (en) * | 2018-04-18 | 2018-11-27 | 北京信息科技大学 | A kind of wind turbine gearbox method for diagnosing faults based on convolutional neural networks |
| CN109633271A (en) * | 2019-01-17 | 2019-04-16 | 长沙理工大学 | The traveling wave Time-Frequency Analysis Method being distributed based on variation mode decomposition and Eugene Wigner Weir |
Non-Patent Citations (2)
| Title |
|---|
| 刘长良 等: "基于变分模态分解和模糊C均值聚类的滚动轴承故障诊断", 《中国电机工程学报》 * |
| 曾学琼: "基于卷积神经网络的变速器故障分类识别研究", 《中国优秀硕士学位论文全文数据库 (工程科技Ⅱ辑)》 * |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110766264A (en) * | 2019-07-31 | 2020-02-07 | 贵州电网有限责任公司 | Transmission cable health state monitoring and evaluating method |
| CN110686899B (en) * | 2019-09-21 | 2021-01-29 | 天津大学 | Internal combustion engine noise source identification method |
| CN110686899A (en) * | 2019-09-21 | 2020-01-14 | 天津大学 | Internal combustion engine noise source identification method |
| CN111141519A (en) * | 2019-12-27 | 2020-05-12 | 西安交通大学 | Rolling bearing fault quantitative evaluation method |
| CN111130053A (en) * | 2020-01-08 | 2020-05-08 | 华南理工大学 | Power distribution network overcurrent protection method based on deep reinforcement learning |
| CN111130053B (en) * | 2020-01-08 | 2021-02-19 | 华南理工大学 | Power distribution network overcurrent protection method based on deep reinforcement learning |
| CN111259949A (en) * | 2020-01-13 | 2020-06-09 | 山东超越数控电子股份有限公司 | Fault identification model construction method, model and identification method for aircraft environmental control system |
| CN111413588A (en) * | 2020-03-31 | 2020-07-14 | 陕西省地方电力（集团）有限公司咸阳供电分公司 | Power distribution network single-phase earth fault line selection method |
| CN111504646A (en) * | 2020-06-05 | 2020-08-07 | 合肥工业大学 | Weak signal fault feature classification method and system for early failure of bearing |
| CN111504646B (en) * | 2020-06-05 | 2022-10-14 | 合肥工业大学 | Weak signal fault feature classification method and system for early failure of bearing |
| CN111664365A (en) * | 2020-06-07 | 2020-09-15 | 东北石油大学 | Oil and gas pipeline leakage detection method based on improved VMD and 1DCNN |
| CN111664365B (en) * | 2020-06-07 | 2022-01-28 | 东北石油大学 | Oil and gas pipeline leakage detection method based on improved VMD and 1DCNN |
| CN112051480A (en) * | 2020-08-06 | 2020-12-08 | 国网安徽省电力有限公司电力科学研究院 | Neural network power distribution network fault diagnosis method and system based on variational modal decomposition |
| CN112101144A (en) * | 2020-08-27 | 2020-12-18 | 云南电网有限责任公司昭通供电局 | Self-adaptive method for improving transformer vibration signal processing precision |
| WO2022067562A1 (en) * | 2020-09-29 | 2022-04-07 | 西门子股份公司 | Method and device for diagnosing fault arc, and computer-readable storage medium |
| CN113740671A (en) * | 2020-10-22 | 2021-12-03 | 国网河北省电力有限公司雄安新区供电公司 | Fault arc identification method based on VMD and ELM |
| CN112580699A (en) * | 2020-12-04 | 2021-03-30 | 中国船舶重工集团公司第七一一研究所 | Sensor data anomaly detection method and application thereof |
| CN113093050A (en) * | 2021-03-31 | 2021-07-09 | 中国矿业大学 | Cable early fault identification method and system based on cable grounding wire current time-frequency characteristics |
| CN113093050B (en) * | 2021-03-31 | 2023-07-07 | 中国矿业大学 | Cable early fault identification method and system based on time-frequency characteristics of cable grounding wire current |
| CN113203914A (en) * | 2021-04-08 | 2021-08-03 | 华南理工大学 | Underground cable early fault detection and identification method based on DAE-CNN |
| WO2022227491A1 (en) * | 2021-04-27 | 2022-11-03 | 京东方科技集团股份有限公司 | Display defect detection apparatus and detection method therefor, and display defect detection system |
| CN115166585A (en) * | 2022-09-08 | 2022-10-11 | 石家庄科林电气股份有限公司 | Ground fault detection anti-misjudgment method and device and electronic equipment |
| CN115166585B (en) * | 2022-09-08 | 2022-11-29 | 石家庄科林电气股份有限公司 | Ground fault detection anti-misjudgment method and device and electronic equipment |
| CN117929952A (en) * | 2024-03-21 | 2024-04-26 | 国网(山东)电动汽车服务有限公司 | Novel arc fault detection method for electric automobile charging pile |
| CN117929952B (en) * | 2024-03-21 | 2024-05-28 | 国网(山东)电动汽车服务有限公司 | Novel arc fault detection method for electric automobile charging pile |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110069886A (en) | The identification of cable initial failure and classification method based on VMD and CNN | |
| CN110334580A (en) | The equipment fault classification method of changeable weight combination based on integrated increment | |
| Schäfer et al. | Detection of gravitational-wave signals from binary neutron star mergers using machine learning | |
| CN109271975A (en) | A kind of electrical energy power quality disturbance recognition methods based on big data multi-feature extraction synergetic classification | |
| CN105488466B (en) | A kind of deep-neural-network and Acoustic Object vocal print feature extracting method | |
| CN109948647A (en) | A kind of electrocardiogram classification method and system based on depth residual error network | |
| CN109616138A (en) | Voice signal blind separating method and ears hearing assistance system based on segmentation frequency point selection | |
| CN110070102B (en) | Method for establishing sequence-to-sequence model for identifying power quality disturbance type | |
| CN108305619A (en) | Voice data collection training method and apparatus | |
| CN106601230A (en) | Logistics sorting place name speech recognition method, system and logistics sorting system based on continuous Gaussian mixture HMM | |
| CN102279358A (en) | MCSKPCA based neural network fault diagnosis method for analog circuits | |
| CN110349597A (en) | A kind of speech detection method and device | |
| CN102982351A (en) | Porcelain insulator vibrational acoustics test data sorting technique based on back propagation (BP) neural network | |
| CN114237046B (en) | Partial discharge pattern recognition method based on SIFT data feature extraction algorithm and BP neural network model | |
| Caicedo et al. | A systematic review of real-time detection and classification of power quality disturbances | |
| Venkatesh et al. | Orthogonal least square center selection technique–a robust scheme for multiple source partial discharge pattern recognition using radial basis probabilistic neural network | |
| CN109948726A (en) | A kind of Power Quality Disturbance Classification Method based on depth forest | |
| CN114966312A (en) | Power distribution network fault detection and positioning method and system based on migration convolutional neural network | |
| McIsaac et al. | Using machine learning to autotune chi-squared tests for gravitational wave searches | |
| CN113378472B (en) | Mixed boundary electromagnetic backscattering imaging method based on generation countermeasure network | |
| Kan et al. | Itunet: Integration of transformers and unet for organs-at-risk segmentation | |
| WO2024099422A1 (en) | Unet-based surface wave frequency dispersion quality rapid evaluation method | |
| CN114218988A (en) | Method for identifying unidirectional ground fault feeder line under unbalanced samples | |
| CN117630569A (en) | Low-current single-phase earth fault multi-criterion fusion line selection method based on GRU neural network | |
| CN113076892A (en) | Sag type identification method, device and system based on deep learning SDAE-BP |
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: 20190730 |
|
| RJ01 | Rejection of invention patent application after publication |