CN109740523A - A kind of method for diagnosing fault of power transformer based on acoustic feature and neural network - Google Patents
A kind of method for diagnosing fault of power transformer based on acoustic feature and neural network Download PDFInfo
- Publication number
- CN109740523A CN109740523A CN201811646299.2A CN201811646299A CN109740523A CN 109740523 A CN109740523 A CN 109740523A CN 201811646299 A CN201811646299 A CN 201811646299A CN 109740523 A CN109740523 A CN 109740523A
- Authority
- CN
- China
- Prior art keywords
- power transformer
- neural network
- gru
- state
- hidden layer
- 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.)
- Granted
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 103
- 238000000034 method Methods 0.000 title claims abstract description 98
- 238000003745 diagnosis Methods 0.000 claims abstract description 28
- 238000012549 training Methods 0.000 claims abstract description 28
- 238000003062 neural network model Methods 0.000 claims abstract description 19
- 239000011159 matrix material Substances 0.000 claims description 84
- 230000008569 process Effects 0.000 claims description 58
- 210000002569 neuron Anatomy 0.000 claims description 55
- 238000012545 processing Methods 0.000 claims description 18
- 238000009432 framing Methods 0.000 claims description 15
- 238000000605 extraction Methods 0.000 claims description 14
- 230000030279 gene silencing Effects 0.000 claims description 14
- 230000009467 reduction Effects 0.000 claims description 13
- 238000000354 decomposition reaction Methods 0.000 claims description 12
- 238000001914 filtration Methods 0.000 claims description 12
- 238000010276 construction Methods 0.000 claims description 8
- 230000001537 neural effect Effects 0.000 claims description 8
- 230000005611 electricity Effects 0.000 claims description 5
- 239000000284 extract Substances 0.000 claims description 5
- 238000011423 initialization method Methods 0.000 claims description 4
- 230000007257 malfunction Effects 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- XCWPUUGSGHNIDZ-UHFFFAOYSA-N Oxypertine Chemical compound C1=2C=C(OC)C(OC)=CC=2NC(C)=C1CCN(CC1)CCN1C1=CC=CC=C1 XCWPUUGSGHNIDZ-UHFFFAOYSA-N 0.000 claims description 3
- 230000001351 cycling effect Effects 0.000 abstract description 3
- 238000012544 monitoring process Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 8
- 238000000513 principal component analysis Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 7
- 239000007789 gas Substances 0.000 description 3
- 230000000306 recurrent effect Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000002405 diagnostic procedure Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 210000004218 nerve net Anatomy 0.000 description 2
- 230000005236 sound signal Effects 0.000 description 2
- 238000013024 troubleshooting Methods 0.000 description 2
- 102000002274 Matrix Metalloproteinases Human genes 0.000 description 1
- 108010000684 Matrix Metalloproteinases Proteins 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000005684 electric field Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 238000009413 insulation Methods 0.000 description 1
- 230000002045 lasting effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000004804 winding Methods 0.000 description 1
Landscapes
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention discloses a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network, the following steps are included: acquiring the voice signal obtained when power transformer is in each state using voice collection device, the voice signal of acquisition and the corresponding relationship of each state of power transformer are recorded；The voice signal of acquisition is pre-processed；It establishes and trains GRU neural network model；The voice signal to Diagnosis for Power Transformer is acquired, is inputted in the GRU neural network model that training finishes after pretreatment, the disconnected diagnosing fault of power transformer of follow-up is completed according to the output result of GRU neural network model.The frequency domain character of power transformer is extracted in voice signal when method of the invention is run from power transformer, thresholding cycling element neural network is trained using the frequency domain character of power transformer, operate relatively easy, cost is relatively low, is easier to realize on-line monitoring.
Description
Technical field
The invention belongs to diagnosing fault of power transformer technical fields, more particularly to one kind to be based on acoustic feature and nerve net
The method for diagnosing fault of power transformer of network.
Background technique
Power transformer as one of the important equipment in electric system, carry the transformation of electric system builtin voltage,
The key tasks such as electric energy distribution and transmission.Power transformer in the process of running it is possible that electric discharge, overheat, insulation degradation,
The failures such as winding and core slackness, the pollution of insulating oil solid-state.Method for diagnosing fault of power transformer is furtherd investigate, to electric system
Stable operation have a very important significance.
As the continuous development of machine Learning Theory is perfect, the non-linear mapping capability of neural network, self-learning capability and
Fault-tolerant ability constantly enhances, by Application of Neural Network in diagnosing fault of power transformer gradually at trend.Shi Xin et al. is in " depth
Practise application of the neural network in diagnosing fault of power transformer " in, it is based on Detection Ssytem of Dissolved Gases in Power Transformer Oil Base analytical technology,
Use H2、CH4、C2H6、C2H4、C2H2、CO、CO2The content value of this 7 kinds of gases is trained neural network.The program is in reality
In, need to carry out the gas content value dissolved in electric power transformer oil contact or contactless measurement, operation
It is cumbersome, higher cost, it is not easy to realize on-line monitoring.To sum up, a kind of novel power transformer neural network based is needed
Method for diagnosing faults.
Summary of the invention
In view of the deficiencies of the prior art, the present invention proposes a kind of based on acoustic feature and the event of the power transformer of neural network
Hinder diagnostic method.The frequency domain that power transformer is extracted in voice signal when method of the invention is run from power transformer is special
Sign, using power transformer frequency domain character to thresholding cycling element (Gated Recurrent Unit, GRU) neural network into
Row training, operation is relatively easy, and cost is relatively low, is easier to realize on-line monitoring.
In order to achieve the above objectives, the invention adopts the following technical scheme:
A kind of method for diagnosing fault of power transformer based on acoustic feature and neural network, comprising the following steps:
Step 1, the voice signal of power transformer is acquired；Become specifically, being acquired using voice collection device and obtaining electric power
Depressor is in voice signal when each state, records the voice signal of acquisition and the corresponding relationship of each state of power transformer；Institute
Each state for stating power transformer includes normal condition and various types of malfunctions；
Step 2, voice signal step 1 acquired pre-processes；The preprocessing process includes: low-pass filtering, letter
One of number de-noising, feature extraction and data dimension-reduction treatment are a variety of；
Step 3, GRU neural network model is established；It specifically includes: determining input layer and output layer neuron number, determination
GRU neuron node number, weights initialisation and the neural metwork training that hidden layer number and each hidden layer include；Institute
State the shape of frequency domain character and the corresponding power transformer of frequency domain character that neural metwork training is obtained with feature extraction in step 2
Training data of the state as neural network；
Step 4, voice signal of the acquisition to Diagnosis for Power Transformer；The sound to Diagnosis for Power Transformer of acquisition is believed
It number is pre-processed by the preprocess method of step 2；The pretreated voice signal to Diagnosis for Power Transformer is inputted
In the GRU neural network model that step 3 training finishes, follow-up is completed according to the output result of GRU neural network model and powers off power
Transformer fault diagnosis.
Further, in step 1, the sample frequency f of voice collection devicesMore than or equal to predetermined threshold ft, expression formula is
fs≥ft, wherein ft=2000Hz.
Further, the preprocessing process of step 2 specifically includes:
Step 2.1, acquisition is obtained when power transformer is in each state using Butterworth lowpass digital filter
Voice signal carries out low-pass filtering；
Step 2.2, wavelet decomposition is carried out to the voice signal of the power transformer after step 2.1 low-pass filtering, to small echo
Wavelet coefficient after decomposition carries out threshold process, the electric power after signal noise silencing is obtained using the wavelet coefficient reconstruct after threshold process
The voice signal of transformer；
Step 2.3, to step 2.2 reconstruct obtain signal noise silencing after power transformer voice signal be normalized with
And framing windowing process, frequency domain character construction is all extracted for each frame obtains the one-dimensional matrix of frequency domain character；
Step 2.4, the one-dimensional matrix of each frequency domain character obtained for step 2.3 extracts data conduct therein
A line of sample matrix, obtains sample matrix, i.e., every data line in sample matrix has only one frequency domain character one-dimensional
Matrix is corresponding to it；Sample matrix is handled using one-dimensional PCA algorithm, contribution rate is chosen in processing result and is greater than threshold value
PCAthPreceding M main component as pretreated result.
Further, in step 2.1, the order N=8 of Butterworth lowpass digital filter used in low-pass filtering,
Cut-off frequecy of passband fp=1000Hz, stopband cut initial frequency fs=1200Hz leads to passband fluctuation minimal attenuation Rp=1dB, stopband
Interior minimal attenuation Rs=50dB；
In step 2.2, the detailed process for carrying out threshold process to the wavelet coefficient after wavelet decomposition in signal noise silencing is: setting
Carrying out the wavelet coefficient obtained after wavelet decomposition to sound S (n) is wi,jIf wavelet coefficient after processing isSet threshold
Value λ, is shown below, if | wi,j| >=λ, thenIf | wi,j| < λ, thenI.e.
Further, normalized detailed process is in step 2.3: setting in the processing of step 2.2 signal noise silencing and reconstructs electricity
The sound of power transformer is It is the numerical value x by successively occurring in the time domain0, x1, x2, x3, xnStructure
At, if xminFor x0, x1, x2, x3, xnIn minimum value, if xmaxFor x0, x1, x2, x3,
xnIn maximum value, to any xi∈{x0,x1,x2,x3,......,xnIts value after normalizing are as follows:
The detailed process of framing adding window is in step 2.3:
If the sound after normalized isFrame length is set as T, it is α that frame, which moves, and framing as existsWhen upper interception
Between length is one section of T and is used as a frame, the Chong Die part in the tail portion of former frame and the head of a later frame is frame shifting α, the knot of jth frame
Beam moment tendFor T+ (j-1) × (1- α) × T, the initial time t of jth framestartFor (j-1) × (1- α) × T, wherein j >=1
And j is integer；During framing, the part is abandoned if the remainder time span of S (n) is less than frame length T, not to this
Part carries out framing operation；
Windowing process is carried out after carrying out sub-frame processing to S (n), if jth frame is fj(n), using window function to fj(n) it carries out
Windowing process；Specifically, using hamming window hm (n) to fj(n) windowing process is carried out；The expression formula of the hamming window hm (n) used
Are as follows:
If to jth frame fj(n) result after progress windowing process isThen
Wherein * indicates convolution.
Further, in step 3, input layer number is M；If the fault type of power transformer is total N kind,
The neuron number of output layer is set as N+1；
It determines hidden layer number and the detailed process of GRU neuron node number that each hidden layer includes is:
GRU neural network includes two hidden layers, and the GRU neuron number that first hidden layer includes is h1, second
The GRU neuron number that hidden layer includes is h2, first hidden layer be connected to input layer and second hidden layer, and second hidden
First hidden layer and output layer are connected to containing layer；
In formula, M is input layer number, and N+1 is output layer neuron number.
Further, in step 3, weight is carried out using Glorot Initialization method in weights initialisation
Initialization, specific process is:
The update door weight of GRU neuron in first hidden layer is W1z, then W1zThe distribution mode initialized
Are as follows:
The resetting door weight of GRU neuron in first hidden layer is W1r, then W1rThe distribution mode initialized
Are as follows:
The input weight of the tanh of GRU neuron in first hidden layer is W1tanh, then W1tanhPoint initialized
Mode for cloth are as follows:
The update door weight of GRU neuron in second hidden layer is W2z, then W2zIt is carried out just according to following distribution mode
Beginningization:
The resetting door weight of GRU neuron in second hidden layer is W2r, then W2rIt is carried out just according to following distribution mode
Beginningization:
The input weight of the tanh of GRU neuron in second hidden layer is W2tanh, then W2tanhAccording to following distribution side
Formula is initialized:
The output layer of GRU neural network is set as full articulamentum, and output layer weight is wout, woutAccording to following distribution mode
It is initialized:
In formula, M is input layer number, and (N+1) is output layer neuron number, h1Include for first hidden layer
GRU neuron number, h2The GRU neuron number for including for second hidden layer.
Further, the frequency domain character extracted in step 2 feature extraction is DFT feature, STFT feature, MFCC feature and SC
One of feature is a variety of.
Further, in step 3, the detailed process of neural metwork training is: using mean square deviation as error function, uses
Training data of the state of frequency domain character and power transformer corresponding with frequency domain character as neural network, uses boarding steps
Spend the weight that descent algorithm updates GRU neural network；In the weight mistake for updating GRU neural network using stochastic gradient descent algorithm
Cheng Zhong, if the variation of mean square deviation is less than scheduled threshold value loss in continuous predetermined step number range stepthre, then stop updating
The process of the weight of GRU neural network；After the weight for stopping update GRU neural network, the calculated GRU mind of final step is saved
Weight through network.
Further, step 4 specifically includes:
Acquire the voice signal to Diagnosis for Power Transformer；The voice signal to Diagnosis for Power Transformer of acquisition is passed through
The preprocess method of step 2 is pre-processed to obtain according to dimensionality reduction matrix, is denoted asNoteForIn any a line；Use one
Tieing up matrix indicates that the state of power transformer, the number of the data of one-dimensional matrix are N+1, and the data in one-dimensional matrix include N number of 0
With 11, the different states of power transformer are distinguished by the difference of the position of number 1 in one-dimensional matrix, note indicates electricity
The matrix of power transformer state is statei；
To eachHave the state of a power transformer withIt is corresponding, byDetermine the state of power transformer
Process is: forIt is obtained after GRU neural network after input trainingIt calculatesWith stateiEuclidean distance,
stateiIt is middle selection withThe smallest state of Euclidean distanceiIt is denoted as statemin, with stateminThe shape of corresponding power transformer
State passes through the state to Diagnosis for Power Transformer that GRU neural network is judged.
Compared with prior art, the invention has the following advantages:
In the present invention, fault diagnosis is carried out to power transformer using power transformer frequency domain character, not by electric field and magnetic
The interference of field does not influence power transformer normal operation, easy to operate, at low cost.Using GRU neural network, nerve can be shortened
The training time of network improves the efficiency for carrying out fault diagnosis to power transformer using GRU neural network.
Further, GRU neural network is initialized using Glorot Initialization method, can be solved
The initial value tender subject of GRU neural network.
Further, during GRU neural metwork training, stopped in time by the judgement to mean square deviation variation range
The right value update of GRU neural network can avoid GRU neural network and over-fitting occur, can be improved using GRU neural network to electricity
The accuracy of power transformer progress fault diagnosis.
Detailed description of the invention
Fig. 1 is a kind of process of method for diagnosing fault of power transformer based on acoustic feature and neural network of the invention
Schematic block diagram；
Fig. 2 is a kind of electric power of method for diagnosing fault of power transformer based on acoustic feature and neural network of the invention
The pretreated schematic process flow diagram of transformer voice signal；
Fig. 3 is a kind of electric power of method for diagnosing fault of power transformer based on acoustic feature and neural network of the invention
The schematic process flow diagram of the pretreated signal noise silencing of transformer voice signal；
Fig. 4 is a kind of electric power of method for diagnosing fault of power transformer based on acoustic feature and neural network of the invention
The schematic process flow diagram of the pretreated feature extraction of transformer voice signal；
Fig. 5 is a kind of electric power of method for diagnosing fault of power transformer based on acoustic feature and neural network of the invention
The schematic process flow diagram of the pretreated Data Dimensionality Reduction of transformer voice signal；
Fig. 6 is a kind of GRU of method for diagnosing fault of power transformer based on acoustic feature and neural network of the invention
The structural schematic block diagram of the single GRU neuron of neural network；
Fig. 7 is a kind of foundation of method for diagnosing fault of power transformer based on acoustic feature and neural network of the invention
The schematic process flow diagram of neural network model.
Specific embodiment
Invention is further described in detail in the following with reference to the drawings and specific embodiments, so that those skilled in the art's energy
The protection scope that the solution of the present invention is enough more clearly understood, but is not intended to limit the present invention.
Referring to Fig. 1, Fig. 1 is that a kind of Power Transformer Faults based on acoustic feature and neural network of the present invention are examined
The flow chart of disconnected method.A kind of method for diagnosing fault of power transformer based on acoustic feature and neural network of the invention, packet
Include following steps:
Step 1: the sound signal collecting to Diagnosis for Power Transformer.
Step 2: power transformer voice signal pretreatment.
Step 3: neural network model is established.
Step 4: diagnosing fault of power transformer.
Step 1 is worked normally and is gone out to Diagnosis for Power Transformer specifically, being collected and recorded using voice collection device
Sound when existing various types failure, the corresponding relationship of record sound and the disconnected Power Transformer Condition of follow-up；Follow-up powers off power and becomes
Depressor state includes normal condition and various types of malfunctions.
In the present invention, mode training GRU (Gated Recurrent Unit, thresholding cycling element) of supervised learning is used
Neural network, using the sound of the power transformer collected and recorded in step 1 through subsequent processing as the instruction of GRU neural network
Practice data, the present invention converts the troubleshooting issue of power transformer to the classification problem of GRU neural network, GRU nerve net
The output of network corresponds to the state of power transformer, and the state of power transformer includes normal condition and various types of events
Barrier state.To ensure that GRU neural network being capable of various states (including normal condition and various types of to power transformer
Malfunction) correct judgement is made, the sound of the power transformer collected and recorded in the step 1 needs to include that electric power becomes
There is sound when various failures in sound and power transformer when depressor works normally.
Specifically, the present invention carries out fault diagnosis to power transformer using the frequency domain character of the sound of power transformer,
Since the frequency domain energy of the sound of power transformer is concentrated mainly on low frequency part, therefore the sample frequency of the voice collection device
fsCannot be too small, the sample frequency f of the voice collection devicesThreshold value f need to be greater than or equal tot, i.e. fs≥ft, it is preferred that ft=
2000Hz。
Referring to Fig. 2, Fig. 2 is a kind of Power Transformer Faults based on acoustic feature and neural network of the present invention
The pretreated flow chart of power transformer voice signal of diagnostic method, power transformer voice signal pretreatment packet in step 2
It includes:
Step (2-1): low-pass filtering.
As previously mentioned, the frequency domain energy of the sound of power transformer is concentrated mainly on low frequency part, therefore use Butterworth
Lowpass digital filter carries out low-pass filtering to the sound of the power transformer collected and recorded in step 1, reduces ambient noise
The interference of medium-high frequency part.
It preferably, is the approximation for realizing preferable passband and stopband, Butterworth lowpass digital filter in this step
Order N=8, cut-off frequecy of passband fp=1000Hz, stopband cut initial frequency fs=1200Hz leads to passband fluctuation minimal attenuation
Rp=1dB, minimal attenuation R in stopbands=50dB.
Step (2-2): signal noise silencing.Fig. 3 is of the present invention a kind of based on acoustic feature and the change of the electric power of neural network
The flow chart of the pretreated signal noise silencing of power transformer voice signal of depressor method for diagnosing faults, as shown in figure 3, signal disappears
It makes an uproar and includes:
Step (2-2-1): wavelet decomposition.If the sound through step (2-1) treated power transformer is S (n), use
Db4 small echo carries out 5 layers of wavelet decomposition to S (n).
Step (2-2-2): wavelet coefficient threshold processing.If being carried out through step (2-2-1) wavelet decomposition to sound S (n) small
The wavelet coefficient obtained after Wave Decomposition is wi,jIf through step (2-2-2) wavelet coefficient threshold, treated that wavelet coefficient isGiven threshold λ, is shown below, if | wi,j| >=λ, thenIf | wi,j| < λ, thenExpression formula are as follows:
Preferably, threshold value λ is set as empirical value 0.025.
Step (2-2-3): wavelet reconstruction.With step (2-2-2) wavelet coefficient threshold treated wavelet coefficientReconstruct
Sound out
Step (2-3): feature extraction.Fig. 4 is of the present invention a kind of based on acoustic feature and the change of the electric power of neural network
The flow chart of the pretreated feature extraction of power transformer voice signal of depressor method for diagnosing faults, as shown in figure 4, feature mentions
It takes and includes:
Step (2-3-1): normalization.
As described in preceding step 1, during collecting and recording power transformer sound using voice collection device,
Voice collection device is at a distance from power transformer, the factors such as the relative angle of voice collection device and power transformer can be to institute
The quality for stating the sound of the power transformer collected and recorded has an impact, therefore to the sound that step (2-2-3) reconstructs
It is normalized, so that sound amplitude value is in same section.
The sound reconstructedIt is the numerical value x by successively occurring in the time domain0, x1, x2, x3, xnStructure
At, if xminFor x0, x1, x2, x3, xnIn minimum value, if xmaxFor x0, x1, x2, x3,
xnIn maximum value, to any xi∈{x0,x1,x2,x3,......,xnIts value after normalizing are as follows:
Step (2-3-2): framing adding window.
If the sound after step (2-3-1) normalized isFrame length is set as T, it is α that frame, which moves,.Framing is
?Upper interception time length is one section of T and is used as a frame, and the Chong Die part in the tail portion of former frame and the head of a later frame is
Frame moves α, then the finish time t of jth (j >=1 and j is integer) frameendFor T+ (j-1) × (1- α) × T, the initial time of jth frame
tstartFor (j-1) × (1- α) × T.During framing, ifRemainder time span be less than frame length T then by the portion
Divide and abandon, framing operation is not carried out to the part.Preferably, frame length T value is 0.5s, and it is 1/3 that frame, which moves α,.
RightWindowing process is carried out after carrying out sub-frame processing.Specifically, setting jth frame as fj(n), using window function pair
fj(n) windowing process is carried out, it is preferred that using hamming window hm (n) to fj(n) windowing process is carried out.The hamming window hm (n) used
It is as follows:
If to jth frame fj(n) result after progress windowing process isThen
Wherein * indicates convolution.
Step (2-3-3): the construction one-dimensional matrix of frequency domain character.
Feature extraction processing is all carried out to each frame after step (2-3-2) framing adding window, as previously mentioned,For to
The processing result of the framing adding window of j frame.The present invention diagnoses Power Transformer Faults using frequency domain character, therefore extractsFrequency domain character, specifically, DFT (discrete Fourier transform, discrete Fourier transform) can be used
It is rightIt is handled to obtain the DFT feature of its frequency domain, STFT (short-time Fourier also can be used
Transform, Short Time Fourier Transform) it is rightIt is handled to obtain the STFT feature of its frequency domain, MFCC also can be used
(Mel-scale Frequency Cepstral Coefficients, mel cepstrum coefficients) are rightIt is handled to obtain
It is right that SC (spectral centroid composes mass center) also can be used in the MFCC feature of its frequency domainIt is handled to obtain it
Any group of above-mentioned four kinds of frequency domain character extracting modes (including DFT, STFT, MFCC, SC) also can be used in the SC feature of frequency domain
Conjunction pairIt is handled to extract frequency domain character, such as uses DFT pairsIt is handled to obtain frequency domain DFT feature, simultaneously
Use MFCC pairsIt is handled to obtain frequency domain MFCC feature.
Further, rightWhat progress DFT was handled is one-dimensional matrix, rightCarry out what STFT was handled
It is two-dimensional matrix, it is rightWhat progress MFCC was handled is two-dimensional matrix, rightWhat progress SC was handled is one-dimensional square
Battle array needs to handle STFT to guarantee that step (2-3-3) constructs the consistency of the feature extraction result of the one-dimensional matrix of frequency domain character
Obtained two-dimensional matrix is converted to one-dimensional matrix, and the two-dimensional matrix for handling MFCC is needed to be converted to one-dimensional matrix, will
The two-dimensional matrix that STFT is handled is converted to the mode of one-dimensional matrix and the two-dimensional matrix for handling MFCC is converted to one
The mode for tieing up matrix is identical, is all that two-dimensional matrix is converted to one-dimensional matrix according in a manner of behavior master.
The case where for a variety of frequency domain characters are extracted, it is first determined whether need to be converted to two-dimensional matrix into one-dimensional matrix,
If desired it converts, is converted in a manner of behavior master according to aforementioned, and multiple one-dimensional matrixes are directly subjected to row and are spliced
To one-dimensional matrix；Multiple one-dimensional matrixes are directly subjected to row splicing if not needing conversion and obtain one-dimensional matrix.
By the treatment process of above-mentioned steps (2-3-3) construction one-dimensional matrix of frequency domain character, one is all obtained for each frame
A one-dimensional matrix of corresponding frequency domain character, the present embodiment are the frequency domain character for using the one-dimensional matrix to indicate the frame.
Step (2-4): Data Dimensionality Reduction.
As described in preceding step (2-3-3) construction one-dimensional matrix of frequency domain character, one is all obtained for each frame and is corresponding to it
The one-dimensional matrix of frequency domain character, find in the actual operation process, the data volume that the one-dimensional matrix of frequency domain character includes is very big,
Directly performance is exactly that the one-dimensional matrix of frequency domain character is very long, this will make the calculation amount in subsequent process very big, in addition, being not frequency
The one-dimensional matrix of characteristic of field each of data it is all advantageous to GRU neural network classification, therefore, use one-dimensional PCA
(Principal Component Analysis, principal component analysis) algorithm carries out dimension-reduction treatment to the one-dimensional matrix of frequency domain character.
Fig. 5 is a kind of power transformer of the method for diagnosing fault of power transformer based on acoustic feature and neural network of the present invention
The flow chart of the pretreated Data Dimensionality Reduction of voice signal, as shown in figure 5, Data Dimensionality Reduction includes:
Step (2-4-1): construction sample matrix.
As described in preceding step (2-3-3) construction one-dimensional matrix of frequency domain character, one is all obtained for each frame and is corresponding to it
The one-dimensional matrix of frequency domain character, matrix one-dimensional for each frequency domain character extracts data therein as sample matrix
A line, i.e., every data line in sample matrix have the one-dimensional matrix of only one frequency domain character to be corresponding to it.
Step (2-4-2): one-dimensional PCA processing.Sample matrix is handled using one-dimensional PCA algorithm, in processing result
Middle selection contribution rate is greater than threshold value PCAthPreceding M main component.Preferably, PCAthValue is 0.7.Specifically, using one-dimensional
PCA algorithm obtains a two-dimensional matrix after handling sample matrix, M column are dropped as data before choosing in the two-dimensional matrix
Matrix is tieed up, every a line in Data Dimensionality Reduction matrix is the result after the corresponding one-dimensional matrix dimension-reduction treatment of frequency domain character.
Step 3: neural network model is established.
Power transformer sound is a kind of time series, often uses LSTM (LongShort-Term in field of neural networks
Memory, the memory of long short) neural network handles time series, but since LSTM neural network model is more multiple
It is miscellaneous, there is a problem of training time length in neural network model training process.(Gated Recurrent Unit, thresholding follow GRU
Ring element) neural network is simplification to LSTM neural network, the training time of neural network model can be shortened.Present invention wound
New property power transformer sound is handled using GRU neural network, to shorten the training time of neural network model.
It include GRU neuron in GRU neural network, Fig. 6 is of the present invention a kind of based on acoustic feature and neural network
Method for diagnosing fault of power transformer GRU neural network single GRU neuron structure chart, as shown in fig. 6, xtIt indicates
In the input of the t moment GRU neuron, ht-1Indicate the output in (t-1) the moment GRU neuron, htIndicating should in t moment
The output of GRU neuron, the GRU neuron include x in the input of t momenttAnd ht-1.σ indicates activation primitive, invention activation
Function σ selects sigmoid function, and tanh indicates hyperbolic tangent function.As shown in fig. 6,
zt=σ (Wz·[ht-1,xt])
rt=σ (Wr·[ht-1,xt])
Wherein, WzIndicate the weight of update door (Update Gate), WrIndicate the weight of resetting door (Reset Gate), W
Indicate the input weight of tanh.
Fig. 7 is a kind of building for method for diagnosing fault of power transformer based on acoustic feature and neural network of the present invention
The flow chart of vertical neural network model, as shown in fig. 7, establishing neural network model and including:
Step (3-1): input layer and output layer neuron number are determined.
As described in step (2-4-2), Data Dimensionality Reduction matrix includes M column, i.e. every a line of Data Dimensionality Reduction matrix has M number
According to every a line in Data Dimensionality Reduction matrix is all used as the input of the GRU neural network by this programme, therefore input layer is a
Number is M.
As described in step 1, the present invention converts the troubleshooting issue of power transformer to the classification of GRU neural network
Problem, the sound of the power transformer collected and recorded in step 1 needs include sound when power transformer works normally with
And sound of power transformer when there are various failures, if the fault type of power transformer is total N kind, then the present invention is by institute
The output layer for stating GRU neural network is set as full articulamentum, and sets (N+1) for the neuron number of output layer.
The GRU neuron node number that step (3-2): determining hidden layer number and each hidden layer includes.
It is generally acknowledged that in the case where containing only a hidden layer, as long as the neuron node number of hidden layer is enough,
Neural network can be fitted arbitrary function, in addition, neural network error can be reduced by increasing hidden layer number.But increase simply
Add the number of hidden layer and the number of neuron that each hidden layer includes, it will be so that neural network becomes complicated, in addition, meeting
So that the training time of neural network is elongated, it is also possible to lead to over-fitting.Based on this, GRU neural network packet that the present invention uses
Two hidden layers are included, if the GRU neuron number that first hidden layer includes is h1If the GRU mind that second hidden layer includes
It is h through first number2.First hidden layer is connected to input layer and second hidden layer, second hidden layer connection
In first hidden layer and output layer.Preferably,
Wherein, M is input layer number, and (N+1) is output layer neuron number.
Step (3-3): weights initialisation.
Neural network there are initial value tender subject, weights initialisation is improper will lead to neural network occur training speed it is slow,
Gradient explosion falls into the problems such as locally optimal solution.Based on this, this programme is using Glorot Initialization method to power
It is initialized again.
Specifically, setting the update door weight of the GRU neuron in first hidden layer as W1z, then W1zAccording to following distribution
Mode is initialized:
If the resetting door weight of the GRU neuron in first hidden layer is W1r, then W1rIt is carried out according to following distribution mode
Initialization:
If the input weight of the tanh of the GRU neuron in first hidden layer is W1tanh, then W1tanhAccording to following distribution
Mode is initialized:
If the update door weight of the GRU neuron in second hidden layer is W2z, then W2zIt is carried out according to following distribution mode
Initialization:
If the resetting door weight of the GRU neuron in second hidden layer is W2r, then W2rIt is carried out according to following distribution mode
Initialization:
If the input weight of the tanh of the GRU neuron in second hidden layer is W2tanh, then W2tanhAccording to following distribution
Mode is initialized:
As described in preceding step (3-1), the output layer of the GRU neural network is set as full articulamentum, if output layer weight
For wout, then woutIt is initialized according to following distribution mode:
In above-mentioned steps (3-3), M is input layer number, and (N+1) is output layer neuron number, h1It is
The GRU neuron number that one hidden layer includes, h2The GRU neuron number for including for second hidden layer.
Step (3-4): neural metwork training.
As described in preceding step (2-4-2), if Data Dimensionality Reduction matrixAre as follows:
Data Dimensionality Reduction matrixIn every a line be all used as the input data of the GRU neural metwork training, be denoted asExample
Such asThe shape of power transformer is indicated using one-dimensional matrix
State is denoted as In data number be (N+1),In data include N number of 0 and 11, byMiddle digital 1
The difference of position distinguishes the different states of power transformer, such asIndicate power transformer
Device is in normal operating conditions,Indicate that power transformer is in the state of fault type one.
Use mean square deviation as error function, usesAnd withIt is correspondingAs the training data of neural network,
The weight of the GRU neural network is updated using stochastic gradient descent algorithm.Described in being updated using stochastic gradient descent algorithm
During the weight of GRU neural network, if the variation of mean square deviation is less than certain threshold in continuous certain step number range step
Value lossthre, then stop the process for updating the weight of the GRU neural network, prevent over-fitting.Stop described in update
After the weight of GRU neural network, the weight of the calculated GRU neural network of final step is saved.Preferably, step value is 5,
lossthreValue is 0.013458.
In this way, through the above steps one, Step 2: step 3, the sound of the power transformer based on acquisition establish mind
Through network model.
Step 4: diagnosing fault of power transformer.
Diagnosing fault of power transformer be process lasting in real time, using voice collection device continuous collecting and record to
The sound of Diagnosis for Power Transformer, voice collection device used in voice collection device and step 1 used in the step 4
It is identical.
By the sound for collecting and recording to Diagnosis for Power Transformer in the step 4 carry out step (2-1) low-pass filtering,
Step (2-2) signal noise silencing, step (2-3) feature extraction and step (2-4-1) construct the processing of sample matrix, in step 4
The sound progress to Diagnosis for Power Transformer collected and recorded rapid (2-1) low-pass filtering, step (2-2) signal noise silencing, step
The process of (2-3) feature extraction and the processing of step (2-4-1) construction sample matrix is low with the step (2-1) in abovementioned steps two
Pass filter, step (2-2) signal noise silencing, step (2-3) feature extraction and step (2-4-1) construct the treatment process of sample matrix
It is identical.Later, sample matrix is handled using one-dimensional PCA algorithm identical with step (2-4-2), is then directly being handled
As a result M main component is denoted as Data Dimensionality Reduction matrix before choosing in
NoteForIn any a line,As the input data of the GRU neural network, such as Length can be obtained after inputting the GRU neural network
For the one-dimensional matrix of (N+1), it is denoted as
As described in preceding step (3-4), the state of power transformer, of the data of one-dimensional matrix are indicated using one-dimensional matrix
Number is (N+1), and the data in one-dimensional matrix include N number of 0 and 11, by the difference of the position of number 1 in one-dimensional matrix come
The different states of power transformer are distinguished, remember that the matrix for indicating Power Transformer Condition is statei.Such as state1=
[1 00 ... 0 0] indicate that power transformer is in normal operating conditions, state2=[0 10 ... 0 0] indicate electric power
Transformer is in the state of fault type one.
To eachHave the state of a power transformer withIt is corresponding, byDetermine the state of power transformer
Process is: forIt inputs after the GRU neural network and to obtainIt calculatesWith stateiEuclidean distance, in stateiIn
Selection withThe smallest state of Euclidean distanceiIt is denoted as statemin, with stateminThe state of corresponding power transformer is this
The state for the power transformer judged described in scheme by GRU neural network, and then realize the electric power based on GRU neural network
The fault diagnosis of transformer.Such asIt is obtained after inputting the GRU neural networkPass through
Calculate discoveryWith state2=[0 10 ... 0 0] Euclidean distance is nearest, with state2The state of corresponding power transformer
For the state of fault type one, then determine that the failure of fault type one occurs in power transformer.
To sum up, the invention proposes a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network,
It include: step 1: power transformer sound signal collecting；Step 2: power transformer voice signal pretreatment；Step 3: it builds
Vertical neural network model；Step 4: diagnosing fault of power transformer.In voice signal when the present invention is run from power transformer
The frequency domain character for extracting power transformer is trained GRU neural network using the frequency domain character of power transformer.To electricity
Power transformer carries out the whole process of fault diagnosis not by the interference of electric and magnetic fields, can be not influence power transformer normal
Fault diagnosis is carried out to power transformer under conditions of operation.
The above is only embodiments of the present invention are described, not by technical solution of the present invention limited to this,
Those skilled in the art's made any known deformation on the basis of major technique design of the invention belongs to institute of the present invention
Technology scope to be protected, the specific protection scope of the present invention are subject to the record of claims.
Claims (10)
1. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network, which is characterized in that including following
Step:
Step 1, the voice signal of power transformer is acquired；Power transformer is obtained specifically, acquiring using voice collection device
Voice signal when in each state records the voice signal of acquisition and the corresponding relationship of each state of power transformer；The electricity
Each state of power transformer includes normal condition and various types of malfunctions；
Step 2, voice signal step 1 acquired pre-processes；The preprocessing process includes: that low-pass filtering, signal disappear
It makes an uproar, one of feature extraction and data dimension-reduction treatment or a variety of；
Step 3, GRU neural network model is established；It specifically includes: determining input layer and output layer neuron number, determines and imply
GRU neuron node number, weights initialisation and the neural metwork training that layer number and each hidden layer include；The mind
The state of the frequency domain character and the corresponding power transformer of frequency domain character that are obtained through network training with feature extraction in step 2 is made
For the training data of neural network；
Step 4, voice signal of the acquisition to Diagnosis for Power Transformer；The voice signal to Diagnosis for Power Transformer of acquisition is led to
The preprocess method for crossing step 2 is pre-processed；By the pretreated voice signal input step 3 to Diagnosis for Power Transformer
In the GRU neural network model that training finishes, completed according to the output result of GRU neural network model to Diagnosis for Power Transformer
Fault diagnosis.
2. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 1,
It is characterized in that, in step 1, the sample frequency f of voice collection devicesMore than or equal to predetermined threshold ft, expression formula fs≥
ft, wherein ft=2000Hz.
3. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 1,
It is characterized in that, the preprocessing process of step 2 specifically includes:
Step 2.1, sound when power transformer is in each state is obtained to acquisition using Butterworth lowpass digital filter
Signal carries out low-pass filtering；
Step 2.2, wavelet decomposition is carried out to the voice signal of the power transformer after step 2.1 low-pass filtering, to wavelet decomposition
Wavelet coefficient afterwards carries out threshold process, the power transformer after signal noise silencing is obtained using the wavelet coefficient reconstruct after threshold process
The voice signal of device；
Step 2.3, the voice signal for reconstructing the power transformer after obtaining signal noise silencing to step 2.2 is normalized and divides
Frame windowing process all extracts each frame frequency domain character construction and obtains the one-dimensional matrix of frequency domain character；
Step 2.4, the one-dimensional matrix of each frequency domain character obtained for step 2.3 extracts data therein as sample
A line of matrix, obtains sample matrix, i.e., every data line in sample matrix has the one-dimensional matrix of only one frequency domain character
It is corresponding to it；Sample matrix is handled using one-dimensional PCA algorithm, contribution rate is chosen in processing result and is greater than threshold value PCAth
Preceding M main component as pretreated result.
4. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 3,
It is characterized in that,
In step 2.1, the order N=8 of Butterworth lowpass digital filter used in low-pass filtering, cut-off frequecy of passband fp
=1000Hz, stopband cut initial frequency fs=1200Hz leads to passband fluctuation minimal attenuation Rp=1dB, minimal attenuation R in stopbands=
50dB；
In step 2.2, the detailed process for carrying out threshold process to the wavelet coefficient after wavelet decomposition in signal noise silencing is: setting to sound
It is w that sound S (n), which carries out the wavelet coefficient obtained after wavelet decomposition,i,jIf wavelet coefficient after processing isGiven threshold λ,
It is shown below, if | wi,j| >=λ, thenIf | wi,j| < λ, thenI.e.
5. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 3,
It is characterized in that,
Normalized detailed process is in step 2.3: setting the sound that power transformer is reconstructed in the processing of step 2.2 signal noise silencing
For It is the numerical value x by successively occurring in the time domain0, x1, x2, x3... ... xnIt constitutes, if xminFor x0, x1, x2,
x3... ... xnIn minimum value, if xmaxFor x0, x1, x2, x3... ... xnIn maximum value, to any xi∈{x0,x1,x2,
x3,......,xnIts value after normalizing are as follows:
The detailed process of framing adding window is in step 2.3:
If the sound after normalized isFrame length is set as T, it is α that frame, which moves, and framing as existsUpper interception time is long
Degree is one section of T and is used as a frame, and the Chong Die part in the tail portion of former frame and the head of a later frame is frame shifting α, at the end of jth frame
Carve tendFor T+ (j-1) × (1- α) × T, the initial time t of jth framestartFor (j-1) × (1- α) × T, wherein j >=1 and j is
Integer；During framing, ifRemainder time span be less than frame length T then the part is abandoned, not to the part
Carry out framing operation；
It is rightWindowing process is carried out after carrying out sub-frame processing, if jth frame is fj(n), using window function to fj(n) adding window is carried out
Processing；Specifically, using hamming window hm (n) to fj(n) windowing process is carried out；The expression formula of the hamming window hm (n) used are as follows:
If to jth frame fj(n) result after progress windowing process isThen
Wherein * indicates convolution.
6. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 3,
It is characterized in that, input layer number is M in step 3；If the fault type of power transformer is total N kind, output layer
Neuron number be set as N+1；
It determines hidden layer number and the detailed process of GRU neuron node number that each hidden layer includes is:
GRU neural network includes two hidden layers, and the GRU neuron number that first hidden layer includes is h1, second hidden layer
The GRU neuron number for including is h2, first hidden layer be connected to input layer and second hidden layer, and second hidden layer connects
It is connected to first hidden layer and output layer；
In formula, M is input layer number, and N+1 is output layer neuron number.
7. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 6,
It is characterized in that, weight is initialized using Glorot Initialization method in weights initialisation in step 3,
Specific process is:
The update door weight of GRU neuron in first hidden layer is W1z, then W1zThe distribution mode initialized are as follows:
The resetting door weight of GRU neuron in first hidden layer is W1r, then W1rThe distribution mode initialized are as follows:
The input weight of the tanh of GRU neuron in first hidden layer is W1tanh, then W1tanhThe distribution side initialized
Formula are as follows:
The update door weight of GRU neuron in second hidden layer is W2z, then W2zIt is initialized according to following distribution mode:
The resetting door weight of GRU neuron in second hidden layer is W2r, then W2rIt is initialized according to following distribution mode:
The input weight of the tanh of GRU neuron in second hidden layer is W2tanh, then W2tanhAccording to following distribution mode into
Row initialization:
The output layer of GRU neural network is set as full articulamentum, and output layer weight is wout, woutIt is carried out according to following distribution mode
Initialization:
In formula, M is input layer number, and (N+1) is output layer neuron number, h1The GRU for including for first hidden layer
Neuron number, h2The GRU neuron number for including for second hidden layer.
8. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 1,
It is characterized in that, the frequency domain character extracted in step 2 feature extraction is in DFT feature, STFT feature, MFCC feature and SC feature
It is one or more.
9. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 1,
It is characterized in that, the detailed process of neural metwork training is in step 3: using mean square deviation as error function, use frequency domain spy
The training data of sign and the state of power transformer corresponding with frequency domain character as neural network, uses stochastic gradient descent
The weight of algorithm update GRU neural network；
During updating the weight of GRU neural network using stochastic gradient descent algorithm, if in continuous predetermined step number range
The variation of mean square deviation is less than scheduled threshold value loss in stepthre, then stop the process for updating the weight of GRU neural network；Stop
After the weight for only updating GRU neural network, the weight of the calculated GRU neural network of final step is saved.
10. a kind of method for diagnosing fault of power transformer based on acoustic feature and neural network according to claim 1,
It is characterized in that, step 4 specifically includes:
Acquire the voice signal to Diagnosis for Power Transformer；The voice signal to Diagnosis for Power Transformer of acquisition is passed through into step
2 preprocess method is pre-processed to obtain according to dimensionality reduction matrix, is denoted asNoteForIn any a line；
Indicate that the state of power transformer, the number of the data of one-dimensional matrix are N+1 using one-dimensional matrix, the number in one-dimensional matrix
According to including N number of 0 and 11, the different shapes of power transformer are distinguished by the difference of the position of number 1 in one-dimensional matrix
State, note indicate that the matrix of Power Transformer Condition is statei；
To eachHave the state of a power transformer withIt is corresponding, byDetermine the process of the state of power transformer
It is: forIt is obtained after GRU neural network after input trainingIt calculatesWith stateiEuclidean distance, in stateiIn
Selection withThe smallest state of Euclidean distanceiIt is denoted as statemin, with stateminThe state of corresponding power transformer passes through
The state to Diagnosis for Power Transformer that GRU neural network is judged.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811646299.2A CN109740523B (en) | 2018-12-29 | 2018-12-29 | Power transformer fault diagnosis method based on acoustic features and neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811646299.2A CN109740523B (en) | 2018-12-29 | 2018-12-29 | Power transformer fault diagnosis method based on acoustic features and neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109740523A true CN109740523A (en) | 2019-05-10 |
| CN109740523B CN109740523B (en) | 2020-12-29 |
Family
ID=66362802
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811646299.2A Active CN109740523B (en) | 2018-12-29 | 2018-12-29 | Power transformer fault diagnosis method based on acoustic features and neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740523B (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110322896A (en) * | 2019-06-26 | 2019-10-11 | 上海交通大学 | A kind of transformer fault sound identification method based on convolutional neural networks |
| CN110415709A (en) * | 2019-06-26 | 2019-11-05 | 深圳供电局有限公司 | Transformer working condition recognition methods based on Application on Voiceprint Recognition model |
| CN110503133A (en) * | 2019-07-26 | 2019-11-26 | 东北大学 | A kind of centrifugal compressor failure prediction method based on deep learning |
| CN110534118A (en) * | 2019-07-29 | 2019-12-03 | 安徽继远软件有限公司 | Transformer/reactor method for diagnosing faults based on Application on Voiceprint Recognition and neural network |
| CN110634493A (en) * | 2019-09-09 | 2019-12-31 | 国网湖南省电力有限公司 | Transformer state identification method, system and medium based on voiceprint image characteristics |
| CN111579056A (en) * | 2020-05-19 | 2020-08-25 | 北京快鱼电子股份公司 | Transformer direct-current magnetic bias prediction method and system |
| CN111626370A (en) * | 2020-05-29 | 2020-09-04 | 展讯通信（上海）有限公司 | Neural network-based hollow-out identification method and device, storage medium and terminal |
| CN111912519A (en) * | 2020-07-21 | 2020-11-10 | 国网安徽省电力有限公司 | Transformer fault diagnosis method and device based on voiceprint frequency spectrum separation |
| CN112114215A (en) * | 2020-09-17 | 2020-12-22 | 陕西师范大学 | Transformer aging evaluation method and system based on error back propagation algorithm |
| CN112147444A (en) * | 2020-09-25 | 2020-12-29 | 广东电网有限责任公司佛山供电局 | Power transformer working state monitoring method and system |
| CN112329914A (en) * | 2020-10-26 | 2021-02-05 | 华翔翔能科技股份有限公司 | Fault diagnosis method and device for buried transformer substation and electronic equipment |
| CN112506690A (en) * | 2020-12-09 | 2021-03-16 | 北京地平线信息技术有限公司 | Method and device for controlling processor |
| CN112599134A (en) * | 2020-12-02 | 2021-04-02 | 国网安徽省电力有限公司 | Transformer sound event detection method based on voiceprint recognition |
| CN114019310A (en) * | 2021-11-08 | 2022-02-08 | 国网湖南省电力有限公司 | Method, equipment and medium for prejudging open circuit fault of energy storage converter |
| CN114157023A (en) * | 2021-11-19 | 2022-03-08 | 广东电网有限责任公司电力调度控制中心 | Distribution transformer early warning information acquisition method |
| CN114783152A (en) * | 2022-03-30 | 2022-07-22 | 郑州熙禾智能科技有限公司 | Energy storage power station fire alarm method and system based on gas-sound information fusion |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5598509A (en) * | 1992-08-28 | 1997-01-28 | Hitachi, Ltd. | Method of configuring a neural network and a diagnosis/recognition system using the same |
| CN104964738A (en) * | 2015-06-10 | 2015-10-07 | 国家电网公司 | Noise separation-based power transformer noise assessment method |
| CN106324502A (en) * | 2016-08-22 | 2017-01-11 | 哈尔滨理工大学 | Asynchronous motor fault analysis method based on wavelet |
| CN106546892A (en) * | 2016-11-10 | 2017-03-29 | 华乘电气科技（上海）股份有限公司 | The recognition methodss of shelf depreciation ultrasonic audio and system based on deep learning |
| CN107293301A (en) * | 2017-05-27 | 2017-10-24 | 深圳大学 | Recognition methods and system based on dental articulation sound |
| CN108960077A (en) * | 2018-06-12 | 2018-12-07 | 南京航空航天大学 | A kind of intelligent failure diagnosis method based on Recognition with Recurrent Neural Network |
-
2018
- 2018-12-29 CN CN201811646299.2A patent/CN109740523B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5598509A (en) * | 1992-08-28 | 1997-01-28 | Hitachi, Ltd. | Method of configuring a neural network and a diagnosis/recognition system using the same |
| CN104964738A (en) * | 2015-06-10 | 2015-10-07 | 国家电网公司 | Noise separation-based power transformer noise assessment method |
| CN106324502A (en) * | 2016-08-22 | 2017-01-11 | 哈尔滨理工大学 | Asynchronous motor fault analysis method based on wavelet |
| CN106546892A (en) * | 2016-11-10 | 2017-03-29 | 华乘电气科技（上海）股份有限公司 | The recognition methodss of shelf depreciation ultrasonic audio and system based on deep learning |
| CN107293301A (en) * | 2017-05-27 | 2017-10-24 | 深圳大学 | Recognition methods and system based on dental articulation sound |
| CN108960077A (en) * | 2018-06-12 | 2018-12-07 | 南京航空航天大学 | A kind of intelligent failure diagnosis method based on Recognition with Recurrent Neural Network |
Non-Patent Citations (4)
| Title |
|---|
| XAVIER GLOROT 等: "Understanding the difficulty of training deep feedforward neural networks", 《APPEARING IN PROCEEDINGS OF THE 13TH INTERNATIONAL CONFERENCE》 * |
| 张生军著: "《基于视觉的无标记手势识别》", 30 June 2016, 长春：吉林大学出版社 * |
| 珠杰著: "《藏文文本自动处理方法研究》", 31 May 2018, 成都：西南交通大学出版社 * |
| 陈绍荣等编著: "《数字信号处理》", 30 September 2016, 北京：国防工业出版社 * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110322896A (en) * | 2019-06-26 | 2019-10-11 | 上海交通大学 | A kind of transformer fault sound identification method based on convolutional neural networks |
| CN110415709A (en) * | 2019-06-26 | 2019-11-05 | 深圳供电局有限公司 | Transformer working condition recognition methods based on Application on Voiceprint Recognition model |
| CN110415709B (en) * | 2019-06-26 | 2022-01-25 | 深圳供电局有限公司 | Transformer working state identification method based on voiceprint identification model |
| CN110503133A (en) * | 2019-07-26 | 2019-11-26 | 东北大学 | A kind of centrifugal compressor failure prediction method based on deep learning |
| CN110534118A (en) * | 2019-07-29 | 2019-12-03 | 安徽继远软件有限公司 | Transformer/reactor method for diagnosing faults based on Application on Voiceprint Recognition and neural network |
| CN110534118B (en) * | 2019-07-29 | 2021-10-08 | 安徽继远软件有限公司 | Transformer/reactor fault diagnosis method based on voiceprint recognition and neural network |
| CN110634493A (en) * | 2019-09-09 | 2019-12-31 | 国网湖南省电力有限公司 | Transformer state identification method, system and medium based on voiceprint image characteristics |
| CN111579056A (en) * | 2020-05-19 | 2020-08-25 | 北京快鱼电子股份公司 | Transformer direct-current magnetic bias prediction method and system |
| CN111626370A (en) * | 2020-05-29 | 2020-09-04 | 展讯通信（上海）有限公司 | Neural network-based hollow-out identification method and device, storage medium and terminal |
| CN111626370B (en) * | 2020-05-29 | 2022-07-12 | 展讯通信（上海）有限公司 | Neural network-based hollow-out identification method and device, storage medium and terminal |
| CN111912519A (en) * | 2020-07-21 | 2020-11-10 | 国网安徽省电力有限公司 | Transformer fault diagnosis method and device based on voiceprint frequency spectrum separation |
| CN112114215A (en) * | 2020-09-17 | 2020-12-22 | 陕西师范大学 | Transformer aging evaluation method and system based on error back propagation algorithm |
| CN112147444A (en) * | 2020-09-25 | 2020-12-29 | 广东电网有限责任公司佛山供电局 | Power transformer working state monitoring method and system |
| CN112329914A (en) * | 2020-10-26 | 2021-02-05 | 华翔翔能科技股份有限公司 | Fault diagnosis method and device for buried transformer substation and electronic equipment |
| CN112329914B (en) * | 2020-10-26 | 2024-02-02 | 华翔翔能科技股份有限公司 | Fault diagnosis method and device for buried transformer substation and electronic equipment |
| CN112599134A (en) * | 2020-12-02 | 2021-04-02 | 国网安徽省电力有限公司 | Transformer sound event detection method based on voiceprint recognition |
| CN112506690A (en) * | 2020-12-09 | 2021-03-16 | 北京地平线信息技术有限公司 | Method and device for controlling processor |
| CN114019310A (en) * | 2021-11-08 | 2022-02-08 | 国网湖南省电力有限公司 | Method, equipment and medium for prejudging open circuit fault of energy storage converter |
| CN114019310B (en) * | 2021-11-08 | 2024-02-20 | 国网湖南省电力有限公司 | Open-circuit fault pre-judging method, equipment and medium for energy storage converter |
| CN114157023A (en) * | 2021-11-19 | 2022-03-08 | 广东电网有限责任公司电力调度控制中心 | Distribution transformer early warning information acquisition method |
| CN114157023B (en) * | 2021-11-19 | 2023-12-05 | 广东电网有限责任公司电力调度控制中心 | Distribution transformer early warning information acquisition method |
| CN114783152A (en) * | 2022-03-30 | 2022-07-22 | 郑州熙禾智能科技有限公司 | Energy storage power station fire alarm method and system based on gas-sound information fusion |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109740523B (en) | 2020-12-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740523A (en) | A kind of method for diagnosing fault of power transformer based on acoustic feature and neural network | |
| Li et al. | Understanding and improving deep learning-based rolling bearing fault diagnosis with attention mechanism | |
| CN112885372B (en) | Intelligent diagnosis method, system, terminal and medium for power equipment fault sound | |
| CN107013449B (en) | The method and system of voice signal identification compressor fault based on deep learning | |
| CN111325095B (en) | Intelligent detection method and system for equipment health state based on acoustic wave signals | |
| CN109044396B (en) | Intelligent heart sound identification method based on bidirectional long-time and short-time memory neural network | |
| CN108062572A (en) | A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on DdAE deep learning models | |
| CN109473120A (en) | A kind of abnormal sound signal recognition method based on convolutional neural networks | |
| CN110534118A (en) | Transformer/reactor method for diagnosing faults based on Application on Voiceprint Recognition and neural network | |
| CN113707176B (en) | Transformer fault detection method based on acoustic signal and deep learning technology | |
| CN109213121B (en) | Method for diagnosing clamping cylinder fault of fan braking system | |
| CN108426713A (en) | Rolling bearing Weak fault diagnostic method based on wavelet transformation and deep learning | |
| CN109555566A (en) | A kind of turbine rotor method for diagnosing faults based on LSTM | |
| CN111759345B (en) | Heart valve abnormality analysis method, system and device based on convolutional neural network | |
| CN109002847A (en) | A kind of axial plunger pump Multiple faults diagnosis approach of the deepness belief network based on index | |
| CN109214356A (en) | A kind of fan transmission system intelligent fault diagnosis method based on DCNN model | |
| CN108444696A (en) | A kind of gearbox fault analysis method | |
| CN114152825B (en) | Transformer fault diagnosis method and device and transformer fault diagnosis system | |
| CN110110809A (en) | The construction method of fuzzy automata based on Machine Fault Diagnosis | |
| CN111680788A (en) | Equipment fault diagnosis method based on deep learning | |
| CN110060368A (en) | Mechanical method for detecting abnormality based on potential feature coding | |
| CN113203914A (en) | Underground cable early fault detection and identification method based on DAE-CNN | |
| CN114263621A (en) | Test method and system for diagnosing and simulating cavitation fault of centrifugal pump | |
| CN115910097A (en) | Audible signal identification method and system for latent fault of high-voltage circuit breaker | |
| JPH0392795A (en) | Diagnostic method of nuclear power plant |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP03 | Change of name, title or address |
Address after: Beilin District in Shaanxi province Xi'an City friendship road 710054 No. 308 Patentee after: STATE GRID SHAANXI ELECTRIC POWER Research Institute Country or region after: China Patentee after: STATE GRID SHAANXI ELECTRIC POWER Co. Patentee after: National Network (Xi'an) Environmental Protection Technology Center Co.,Ltd. Address before: Beilin District in Shaanxi province Xi'an City friendship road 710054 No. 308 Patentee before: STATE GRID SHAANXI ELECTRIC POWER Research Institute Country or region before: China Patentee before: STATE GRID SHAANXI ELECTRIC POWER Co. Patentee before: XI'AN POWER TRANSMISSION AND TRANSFORMATION PROJECT ENVIRONMENTAL IMPACT CONTROL TECHNIQUE CENTER Co.,Ltd. |
|
| CP03 | Change of name, title or address | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20240403 Address after: No. 669, Hangtian Middle Road, Xi'an national civil aerospace industry base, Shaanxi 710100 Patentee after: Electric Power Research Institute of State Grid Shaanxi Electric Power Co.,Ltd. Country or region after: China Patentee after: State Grid Shaanxi Electric Power Co.,Ltd. Patentee after: National Network (Xi'an) Environmental Protection Technology Center Co.,Ltd. Address before: Beilin District in Shaanxi province Xi'an City friendship road 710054 No. 308 Patentee before: STATE GRID SHAANXI ELECTRIC POWER Research Institute Country or region before: China Patentee before: STATE GRID SHAANXI ELECTRIC POWER Co. Patentee before: National Network (Xi'an) Environmental Protection Technology Center Co.,Ltd. |
|
| TR01 | Transfer of patent right |