CN107886085A - A kind of electrical energy power quality disturbance feature extracting method based on t SNE - Google Patents
A kind of electrical energy power quality disturbance feature extracting method based on t SNE Download PDFInfo
- Publication number
- CN107886085A CN107886085A CN201711228181.3A CN201711228181A CN107886085A CN 107886085 A CN107886085 A CN 107886085A CN 201711228181 A CN201711228181 A CN 201711228181A CN 107886085 A CN107886085 A CN 107886085A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msub
- msup
- sne
- formula
- 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
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
- G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Abstract
The invention discloses the extracting method of the electrical energy power quality disturbance feature based on t SNE, methods described establishes common PQD signal mathematical modelings, and consider the influence of disturbance parameter and noise, then the wavelet energy vector construction primitive character collection of signal is obtained using wavelet decomposition, Further Feature Extraction is carried out finally by t SNE algorithms, be maintained sample higher dimensional space structure, the low-dimensional feature that susceptibility is high and cluster property is good.
Description
Technical field
The present invention relates to electric power monitoring technical field, more particularly to carrying for the electrical energy power quality disturbance feature based on t-SNE
Take method.
Background technology
Electrical energy power quality disturbance (power quality disturbance, PQD) is present in the power system monitoring number of magnanimity
In, the emphasis for turning into power workers and having studied is identified based on big data and intelligentized PQD.Wherein analysis voltage disturbs
Signal simultaneously chooses the key point that suitable characteristic vector is PQD identifications basis, and the correlation and redundancy of characteristic vector decide
The height of recognition accuracy.Visual characteristic and subjective understanding with reference to people, make PQD data minings result have interaction and
Intuitive, and then help that complicated PQD data are further analyzed from different visualization angles, it has also become in PQD researchs
A developing direction.
At present, when generally using-frequency analysis method extraction PQD primary signals in characteristic vector, wherein wavelet transformation quilt
It is widely used in PQD feature extractions, and achieves preferable effect.Because wavelet function is decayed quickly in itself, belong to a kind of temporary
State waveform, using it for power quality analysis field, particularly analysis of transient process field has Fourier transformation and in short-term Fu
In leaf transformation it is incomparable the advantages of.
Kept to obtain in the low-dimensional data of taxonomic structure in higher-dimension, and obtain preferable visualization result, epidemiology
Algorithm is practised to be introduced in PQD feature extractions.In order to handle the data of the nonlinear organization run into a large amount of daily lifes, people
The dimensionality reduction technology based on geometry is proposed, is referred to as manifold learning, i.e., it is empty to carry out son by the geometry distribution of input data
Between solve.Such as Multidimensional Scaling (Multi Dimensional Scaling, MDS), Isometric Maps algorithm
(isometric mapping, ISOMAP), it is locally linear embedding into (locally linear embedding, LLE), locally cuts
Space arrangement algorithm (local tangent space alignment, LTSA), local holding mapping (locality
Preserving projection, LPP) etc..It is empty that dimension reduction algorithm described above solves son by matrix- eigenvector-decomposition
Between, so as to reach the purpose of dimensionality reduction.Euclidean distance between high dimensional data is cleverly converted into probability tables and reaches shape by Hinton et al.
Formula, it is proposed that random neighbor is embedded in (stochastic neighbor embedding, SNE), and its object function structure criterion will
Ask subspace and the former input space that there is identical form of probability.SNE belongs to a kind of new dimensionality reduction based on probability metrics
Algorithm, its dimensionality reduction and effect of visualization are better than the most dimension-reduction algorithm based on matrix measures.Later, Laurens et al. existed
Improved on the basis of this, it is proposed that t distribution random neighbor insertions (t-distributed stochastic neighbor
Embedding, t-SNE), substitute the conditional probability in SNE with the joint probability with symmetry, and t is introduced in subspace
The similarity of two samples of Distribution Function Definition.A kind of gradient table of simpler object function can be symmetrically obtained using this
Show, therefore t-SNE algorithms are easier to optimize than SNE algorithm, so as to obtain more preferable low dimensional structures and more preferable visualization result.
It is desirable to have a kind of extracting method of the electrical energy power quality disturbance feature based on t-SNE to can solve the problem that in the prior art
The problem of correlation and redundancy of existing PQD characteristic vectors.
The content of the invention
It is an object of the invention to provide a kind of extracting method of electrical energy power quality disturbance feature based on t-SNE to solve low-dimensional
PQD feature space data are excessively crowded, part aliasing easily occur, lose the problem of legacy data structure.
The present invention provides a kind of extracting method of the electrical energy power quality disturbance feature based on t-SNE, and basis signal model produces
PQD signals, the PQD signals of acquisition are sampled, wavelet transform process then is carried out to sample, it is popular by t-SNE
Learning algorithm carries out dimension-reduction treatment to the PQD wavelet energies vector extracted, and it is carried out into visualization with scatterplot diagram form and shown
Show.
Preferably, the basis signal model produces PQD signals and included：Voltage projection, voltage sag, voltage are interrupted, are humorous
Ripple, impulse transients and vibration transient state；
Voltage projection formula：V (t)=A { 1+ α [u (t2)-u(t1)] sin ω t, wherein 0.1≤α≤0.8, T≤t2-t1
≤9T；
Voltage sag formula：V (t)=A { 1- α [u (t2)-u(t1)] sin ω t, wherein 0.1≤α≤0.8, T≤t2-t1
≤9T；
Voltage is interrupted formula：V (t)=A { 1- α [u (t2)-u(t1)] sin ω t, wherein 0.9≤α≤1, T≤t2-t1≤
9T, 0.05≤α3(α5,α7)≤0.15；
Harmonic wave formula：V (t)=A [α1sinωt+α3sin3ωt+α1sin5ωt+α7Sin ω t], wherein 0.05≤α3
(α5,α7)≤0.15, ∑ αi 2=1；
Impulse transients formula：V (t)={ 1- [u (t-t1)-u(t-t2)] sin ω t, wherein 0.05T≤t2-t1≤0.1T；
Vibrate transient state formula：V (t)=sin ω t+ae-c(t-t1)sinbωt[u(t2)-u(t1)], wherein 0≤t2-t1≤
2T；
Wherein, α is amplitude；t1And t2Respectively disturb start time and finish time；T is signal cycle.
Preferably, the wavelet transform process is realized and decomposed by reusing low pass filter and high-pass filter；Filter
Low frequency component and high fdrequency component that ripple device obtains respectively account for the 1/2 of signal band, and above-mentioned point is repeated to obtained new low frequency component
Solution preocess, obtain next layer of high fdrequency component and low frequency component；
According to Parseval theorems, energy wavelet coefficient formula is as follows：
∫[f(t)]2Dt=∑s [aj(k)]2+∑[dj(k)]2 (1)
In formula, f (t) is signal to be decomposed, aj(k) it is the approximation coefficient of wavelet decomposition jth layer, dj(k) it is wavelet decomposition the
The detail coefficients of j layers；
J layer decomposition is carried out to PQD, the wavelet transformation approximate energy distribution of f (t) signals and details Energy distribution define respectively
For：
In formula, j=1,2 ..., J, by J layer wavelet decompositions, characteristic vector is obtained
Wavelet decomposition is carried out using above formula, is counted under same yardstickComposition of vector
Set of eigenvectors as PQD signals.
Preferably due to the orthogonality, compact sup-port, the sensitivity to means of chaotic signals of Daubechies (db) wavelets
Property, with reference to PQD signals, the J=7, db4 small echos carry out 7 layers of decomposition and obtain Ed7 dimensional vectors, the data as t-SNE computings
Sample.
Preferably, the number that the t-SNE prevalences learning algorithm is introduced into the former space of joint probability expression with symmetry
According to similarity, using the similarity between t distribution expression images in subspace；
Given n d dimension sample vector X={ x1, x2..., xn, data xiAnd xjBetween similarity by joint probability pij
Expression, represents x in former spaceiSelect xjAs the probability of neighbour, i.e.,：
Wherein λ be Gaussian function variance, pijSimilarity probability between=0 and data and be 1；
Choose n r (r ＜ ＜ d) dimensional vector Y={ y1, y2..., ynIt is used as subspace data corresponding to X；It is distributed using t
Probability between expressor spatial data, qijRepresent subspace yiAnd yjBetween similarity：
T-SNE is obtained by formula (6) and is minimized object function：
The vector table of optimal subspace reaches, that is, minimizes the Kullback- of former two probability distribution in space and subspace
Leibler divergences, essence are farthest to match pijAnd qij, then pass through the optimal of gradient descent method solution formula (6)
Value；
To improve existing oscillatory occurences during optimization in solution procedure, and accelerate optimization process, in formula (6)
On the basis of plus a momentum term, obtain the gradient with momentum：
Wherein,For the m times iterative vectorized Y value, η is learning rate, and β (m) is
The momentum value of the m times iteration.
Preferably, the t-SNE prevalences learning algorithm comprises the following steps：
Step 1：Determine sample matrix X={ x1, x2..., xn, setting variance parameter λ；
Step 2：Euclidean distance two-by-two between input sample is calculated according to X；According to formula (4) design conditions joint probability pij；
Step 3：First pass through formula (5) and calculate joint probability qij；Then gradient distribution is calculated according to formula (6), finally by
The optimal value of gradient descent method solution formula (6)；
Step 4：Output.
The invention discloses a kind of extracting method of the electrical energy power quality disturbance feature based on t-SNE, is considering disturbance parameter
Under conditions of randomness and influence of noise, establish sinusoidal signal and 6 kinds of common PQD signals, to the PQD wavelet energies that extract to
Amount carries out t-SNE algorithm dimensionality reductions, obtains 3-dimensional PQD characteristic vectors, and its visualization of 3 d figure clearly can effectively distinguish PQD letters
Number.Show that t-SNE algorithms solve low-dimensional PQD feature spaces data and excessively gathered around with the contrast experiment of other popular learning algorithms
Squeeze, part aliasing easily occur, lose the problem of legacy data structure.
Brief description of the drawings
Fig. 1 is the flow chart of the extracting method of the electrical energy power quality disturbance feature based on t-SNE.
Fig. 2 is that basis signal model produces PQD signal waveforms.
Fig. 3 is SNE visualization of 3 d scatter diagrams.
Fig. 4 is t-SNE visualization of 3 d scatter diagrams.
Embodiment
To make the purpose, technical scheme and advantage that the present invention is implemented clearer, below in conjunction with the embodiment of the present invention
Accompanying drawing, the technical scheme in the embodiment of the present invention is further described in more detail.In the accompanying drawings, identical from beginning to end or class
As label represent same or similar element or the element with same or like function.Described embodiment is the present invention
Part of the embodiment, rather than whole embodiments.The embodiments described below with reference to the accompanying drawings are exemplary, it is intended to uses
It is of the invention in explaining, and be not considered as limiting the invention.Based on the embodiment in the present invention, ordinary skill people
The every other embodiment that member is obtained under the premise of creative work is not made, belongs to the scope of protection of the invention.
Data Dimensionality Reduction, i.e., high dimension vector (being more than 3-dimensional) is converted into 2 dimensions or 3-dimensional data vector, it is brief after data can be with
The classification of display data in 2 dimensions or 3-dimensional coordinate diagram.Excellent Method of Data with Adding Windows can be to greatest extent by high dimensional data
Portion's structural relation is showed in a manner of visual.
SNE methods define a kind of new probability that can keep the distant relationships before and after dimensionality reduction between data point, so as to
Keep the internal structure of data.The core concept of this method is to select the sample of phase neighbour by complexity factors first；Its
It is secondary, the Euclidean distance between neighbour's sample is converted into Probability Forms, that is, the similarity of sample；Finally, pass through
Kullback-Leibler divergence object functions obtain the data representation after dimensionality reduction.Wherein data xiAnd xjBetween similarity by
Conditional probability is expressed, and illustrates xiSelect xjProbability as neighbour；Data y in its subspace being embedded iniAnd yjBetween it is similar
Degree have selected similar probability expression way.
Embodiments of the invention are described in detail below in conjunction with the accompanying drawings.
As shown in figure 1, producing PQD signals using Matlab simulation softwares basis signal model, carried out to obtaining PQD signals
Sampling, wavelet transformation then is carried out to sample, dimension-reduction treatment is carried out to the vector by t-SNE prevalences learning algorithm, and
It is carried out into visualization with scatterplot diagram form to show.7 layers of decomposition are carried out to signal from db4 small echos in experimentation and obtain small echo
Energy vectors, one 7 dimension sample vector is constructed, then uses popular learning algorithm to carry out at Data Dimensionality Reduction and visualization to it
Reason.
The mathematical formulae model provided according to table 1, establishes six kinds of common PQD signals (voltage projection, voltage sag, electricity
Interruption, harmonic wave, voltage transient are pressed, wherein voltage transient includes impulse transients and vibration transient state).
The signal model of table 1
Tab.1Signal model
PQD oscillograms as shown in Figure 2, wherein, α is amplitude；t1And t2Respectively disturb start time and finish time；T is
Signal cycle, the analysis time length of input signal are taken as 10 cycles i.e. 0.2s, electric voltage frequency 50Hz.
Take voltage sinusoidal signal and voltage projection, voltage sag, voltage interruption, harmonic wave, wherein voltage transient, voltage transient
Including impulse transients and each 100 of the class disturbing signal of transient state 6 is vibrated, totally 700 samples；The analysis time length of input signal takes
10 cycles (0.2s), sample rate 6.4kHz, the sampled point of signal is 1280, and each cycle adopts 128 points；Preferably
The various situations of simulation in practice, ensure the reliability of analysis result, make the parameter of every kind of disturbance such as disturb beginning and ending time, width
Value, duration etc. change at random in allowed band, and add the random white noise that signal to noise ratio is 30dB to disturbing signal.It is right
PQD signals are sampled to obtain 7 group 100 × 1280 of sample data.
Euclidean distance between high dimensional data is converted into likelihood probability and represented by t-SNE algorithms, substitutes SNE algorithms with joint probability
In conditional probability, so as to effectively alleviate SNE mapping point congested problems present in lower dimensional space.In order to verify t-SNE side
Method to greatest extent can show the internal structure relation of PQD data, and experiment passes through the energy of 7 layers of decomposition of db4 small echos first
Value construction sample vector data, are then respectively adopted SNE, t-SNE algorithm and carry out data processing again, so as to extract PQD 3-dimensional spy
Sign vector.
As shown in figure 3, the three-dimensional scatter diagram of Data Dimensionality Reduction is carried out using SNE methods, it can be seen that the embedded knot of SNE methods
The scattered chaos disorder distribution of different classes of sample in fruit, and data are very crowded.SNE algorithms are based on probability metrics
Dimension-reduction algorithm, the insertion of its low-dimensional maintain the distant relationships between former data point.But its there is also value equation optimization it is difficult and
Low-dimensional data " congested problem ".
As shown in figure 4, the three-dimensional scatter diagram of PQD Data Dimensionality Reductions is carried out using t-SNE methods, the embedded knot of t-SNE methods
Fruit shows that 7 kinds of different classes of sample datas are distributed in seven Different Planes in three dimensions, normal signal and 6 kinds of PQD
Signal can be distinguished substantially, have certain Clustering Effect t-SNE algorithms on the basis of SNE algorithms while dimensionality reduction is completed
On propose joint probability with symmetry and represent data similarity in former space, represented in subspace using t distributions
Similarity between image.So as to solve the value equation of SNE algorithms optimization difficulty and low-dimensional data congested problem.Due to t-
T distributions in SNE methods are a kind of typical heavytailed distributions so that the distance between Various types of data after dimensionality reduction is bigger, alleviates
" congested problem " in nonlinear reductive dimension algorithm.This result shows that the PQD feature extracting methods based on t-SNE can keep sample
The low dimensional structures hidden in this higher dimensional space, and this potential manifold is embedded into lower dimensional space.Each dimensionality reduction is calculated more than
The three-dimensional visualization scatter diagram of method can be seen that t-SNE methods are better than other epidemiologies for the dimensionality reduction classifying quality of PQD signals
Habit dimension reduction method.
Data visualization result, which shows compared to other epidemic algorithms can more reflect based on t-SNE methods, is hidden in PQD
Low dimensional structures in higher-dimension sample data, it is found that the internal distribution of high dimensional data is regular.Low-dimensional after t-SNE Dimensionality Reductions is special
It is more preferable to levy the Clustering Effect of vector, follow-up PQD classification pressure can be mitigated while dimensionality reduction.
It is last it is to be noted that：The above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations.To the greatest extent
The present invention is described in detail with reference to the foregoing embodiments for pipe, it will be understood by those within the art that：It is still
Technical scheme described in foregoing embodiments can be modified, or which part technical characteristic is equally replaced
Change；And these modifications or replacement, the essence of appropriate technical solution is departed from the essence of various embodiments of the present invention technical scheme
God and scope.
Claims (6)
1. a kind of electrical energy power quality disturbance feature extracting method based on t-SNE, it is characterised in that basis signal model produces PQD
Signal, the PQD signals of acquisition are sampled, wavelet transform process then is carried out to sample, learnt by the way that t-SNE is popular
Algorithm carries out dimension-reduction treatment to the PQD wavelet energies vector extracted, and it is carried out into visualization with scatterplot diagram form and shown.
2. the extracting method of the electrical energy power quality disturbance feature according to claim 1 based on t-SNE, it is characterised in that：Institute
Stating basis signal model generation PQD signals includes：Voltage projection, voltage sag, voltage interruption, harmonic wave, impulse transients and vibration
Transient state；
Voltage projection formula：V (t)=A { 1+ α [u (t2)-u(t1)] sin ω t, wherein 0.1≤α≤0.8, T≤t2-t1≤9T；
Voltage sag formula：V (t)=A { 1- α [u (t2)-u(t1)] sin ω t, wherein 0.1≤α≤0.8, T≤t2-t1≤9T；
Voltage is interrupted formula：V (t)=A { 1- α [u (t2)-u(t1)] sin ω t, wherein 0.9≤α≤1, T≤t2-t1≤ 9T,
0.05≤α3(α5,α7)≤0.15；
Harmonic wave formula：V (t)=A [α1sinωt+α3sin3ωt+α1sin5ωt+α7Sin ω t], wherein 0.05≤α3(α5,α7)
≤ 0.15, ∑ αi 2=1；
Impulse transients formula：V (t)={ 1- [u (t-t1)-u(t-t2)] sin ω t, wherein 0.05T≤t2-t1≤0.1T；
Vibrate transient state formula：V (t)=sin ω t+ae-c(t-t1)sinbωt[u(t2)-u(t1)], wherein 0≤t2-t1≤2T；
Wherein, α is amplitude；t1And t2Respectively disturb start time and finish time；T is signal cycle.
3. the extracting method of the electrical energy power quality disturbance feature according to claim 2 based on t-SNE, it is characterised in that：Institute
State wavelet transform process and realize and decompose by reusing low pass filter and high-pass filter；The low frequency component that wave filter obtains
The 1/2 of signal band is respectively accounted for high fdrequency component, above-mentioned decomposable process is repeated to obtained new low frequency component, obtains next layer
High fdrequency component and low frequency component；
According to Parseval theorems, energy wavelet coefficient formula is as follows：
∫[f(t)]2Dt=∑s [aj(k)]2+∑[dj(k)]2 (1)
In formula, f (t) is signal to be decomposed, aj(k) it is the approximation coefficient of wavelet decomposition jth layer, dj(k) it is wavelet decomposition jth layer
Detail coefficients；
J layer decomposition is carried out to PQD, the wavelet transformation approximate energy distribution of f (t) signals and details Energy distribution are respectively defined as：
<mrow>
<msub>
<mi>E</mi>
<msub>
<mi>a</mi>
<mi>j</mi>
</msub>
</msub>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mi>n</mi>
</munder>
<msup>
<mrow>
<mo>(</mo>
<msub>
<mi>a</mi>
<mi>j</mi>
</msub>
<mo>(</mo>
<mi>n</mi>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<msub>
<mi>E</mi>
<msub>
<mi>d</mi>
<mi>j</mi>
</msub>
</msub>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mi>n</mi>
</munder>
<msup>
<mrow>
<mo>(</mo>
<msub>
<mi>d</mi>
<mi>j</mi>
</msub>
<mo>(</mo>
<mi>n</mi>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>3</mn>
<mo>)</mo>
</mrow>
</mrow>
In formula, j=1,2 ..., J, by J layer wavelet decompositions, characteristic vector is obtained
Wavelet decomposition is carried out using above formula, is counted under same yardstickComposition of vectorAs
The set of eigenvectors of PQD signals.
4. the extracting method of the electrical energy power quality disturbance feature according to claim 3 based on t-SNE, it is characterised in that：By
In the orthogonality of Daubechies (db) wavelets, compact sup-port and the sensitiveness to means of chaotic signals, with reference to PQD signals, institute
J=7 is stated, db4 small echos carry out 7 layers of decomposition and obtain Ed7 dimensional vectors, the data sample as t-SNE computings.
5. the extracting method of the electrical energy power quality disturbance feature according to claim 4 based on t-SNE, it is characterised in that：Institute
The data similarity that t-SNE prevalence learning algorithms are introduced into the former space of joint probability expression with symmetry is stated, in subspace
It is middle that the similarity represented between image is distributed using t；
Given n d dimension sample vector X={ x1, x2..., xn, data xiAnd xjBetween similarity by joint probability pijExpression,
Represent x in former spaceiSelect xjAs the probability of neighbour, i.e.,：
<mrow>
<msub>
<mi>p</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
</msub>
<mo>=</mo>
<mfrac>
<mrow>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msup>
<mi>&lambda;</mi>
<mn>2</mn>
</msup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>k</mi>
<mo>&NotEqual;</mo>
<mi>l</mi>
</mrow>
</munder>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>k</mi>
</msub>
<mo>-</mo>
<msub>
<mi>x</mi>
<mi>l</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msup>
<mi>&lambda;</mi>
<mn>2</mn>
</msup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>4</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein λ be Gaussian function variance, pijSimilarity probability between=0 and data and be 1；
Choose n r (r ＜ ＜ d) dimensional vector Y={ y1, y2..., ynIt is used as subspace data corresponding to X；Using t distribution and expressions
Probability between the data of subspace, qijRepresent subspace yiAnd yjBetween similarity：
<mrow>
<msub>
<mi>q</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
</msub>
<mo>=</mo>
<mfrac>
<msup>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>+</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>)</mo>
</mrow>
<mrow>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
<mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>k</mi>
<mo>&NotEqual;</mo>
<mi>l</mi>
</mrow>
</munder>
<msup>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>+</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>y</mi>
<mi>k</mi>
</msub>
<mo>-</mo>
<msub>
<mi>y</mi>
<mi>l</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>)</mo>
</mrow>
<mrow>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>5</mn>
<mo>)</mo>
</mrow>
</mrow>
T-SNE is obtained by formula (6) and is minimized object function：
<mrow>
<mi>C</mi>
<mrow>
<mo>(</mo>
<mi>Y</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>&NotEqual;</mo>
<mi>j</mi>
</mrow>
</munder>
<msub>
<mi>p</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
</msub>
<mi>l</mi>
<mi>g</mi>
<mfrac>
<msub>
<mi>p</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
</msub>
<msub>
<mi>q</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
</msub>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>6</mn>
<mo>)</mo>
</mrow>
</mrow>
The vector table of optimal subspace reaches, that is, minimizes the Kullback-Leibler of former two probability distribution in space and subspace
Divergence, essence are farthest to match pijAnd qij, then pass through the optimal value of gradient descent method solution formula (6)；
To improve existing oscillatory occurences during optimization in solution procedure, and accelerate optimization process, the base in formula (6)
Add a momentum term on plinth, obtain the gradient with momentum：
<mrow>
<msup>
<mi>Y</mi>
<mrow>
<mo>(</mo>
<mi>m</mi>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<msup>
<mi>Y</mi>
<mrow>
<mo>(</mo>
<mi>m</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>+</mo>
<mi>&eta;</mi>
<mfrac>
<mrow>
<mi>d</mi>
<mi>C</mi>
<mrow>
<mo>(</mo>
<mi>Y</mi>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mi>d</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<msub>
<mo>|</mo>
<mrow>
<mi>Y</mi>
<mo>=</mo>
<msup>
<mi>Y</mi>
<mrow>
<mo>(</mo>
<mi>m</mi>
<mo>)</mo>
</mrow>
</msup>
</mrow>
</msub>
<mo>+</mo>
<mi>&beta;</mi>
<mrow>
<mo>(</mo>
<mi>m</mi>
<mo>)</mo>
</mrow>
<mrow>
<mo>(</mo>
<msup>
<mi>Y</mi>
<mrow>
<mo>(</mo>
<mi>m</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>-</mo>
<msup>
<mi>Y</mi>
<mrow>
<mo>(</mo>
<mi>m</mi>
<mo>-</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>7</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein,For the m times iterative vectorized Y value, η is learning rate, and β (m) is m
The momentum value of secondary iteration.
6. the extracting method of the electrical energy power quality disturbance feature according to claim 5 based on t-SNE, it is characterised in that：Institute
T-SNE prevalence learning algorithms are stated to comprise the following steps：
Step 1：Determine sample matrix X={ x1, x2..., xn, setting variance parameter λ；
Step 2：Euclidean distance two-by-two between input sample is calculated according to X；According to formula (4) design conditions joint probability pij；
Step 3：First pass through formula (5) and calculate joint probability qij；Then gradient distribution is calculated according to formula (6), finally by gradient
The optimal value of descent method solution formula (6)；
Step 4：Output.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711228181.3A CN107886085A (en) | 2017-11-29 | 2017-11-29 | A kind of electrical energy power quality disturbance feature extracting method based on t SNE |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711228181.3A CN107886085A (en) | 2017-11-29 | 2017-11-29 | A kind of electrical energy power quality disturbance feature extracting method based on t SNE |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107886085A true CN107886085A (en) | 2018-04-06 |
Family
ID=61775965
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711228181.3A Pending CN107886085A (en) | 2017-11-29 | 2017-11-29 | A kind of electrical energy power quality disturbance feature extracting method based on t SNE |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107886085A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109101890A (en) * | 2018-07-16 | 2018-12-28 | 中国科学院自动化研究所 | Electrical energy power quality disturbance recognition methods and device based on wavelet transformation |
| CN109255313A (en) * | 2018-08-30 | 2019-01-22 | 中国科学院国家授时中心 | A kind of method of promotion signal recognition correct rate |
| CN109583337A (en) * | 2018-11-16 | 2019-04-05 | 华北电力大学（保定） | Electrical energy power quality disturbance recognition methods based on wavelet transformation |
| CN109918162A (en) * | 2019-02-28 | 2019-06-21 | 集智学园（北京）科技有限公司 | A kind of massive information higher-dimension graphic interaction methods of exhibiting learnt |
| CN110363227A (en) * | 2019-06-24 | 2019-10-22 | 广东工业大学 | A kind of LED classification method based on manifold learning |
| CN111950696A (en) * | 2020-06-29 | 2020-11-17 | 燕山大学 | Short-term power load prediction method based on dimension reduction and improved neural network |
| CN112888008A (en) * | 2021-01-08 | 2021-06-01 | 南京中兴力维软件有限公司 | Base station abnormity detection method, device, equipment and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102510044A (en) * | 2011-11-04 | 2012-06-20 | 上海电力学院 | Excitation inrush current identification method based on wavelet transformation and probabilistic neural network (PNN) |
| CN106443335A (en) * | 2016-09-20 | 2017-02-22 | 广西电网有限责任公司钦州供电局 | Lightning stroke fault identification method and system |
-
2017
- 2017-11-29 CN CN201711228181.3A patent/CN107886085A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102510044A (en) * | 2011-11-04 | 2012-06-20 | 上海电力学院 | Excitation inrush current identification method based on wavelet transformation and probabilistic neural network (PNN) |
| CN106443335A (en) * | 2016-09-20 | 2017-02-22 | 广西电网有限责任公司钦州供电局 | Lightning stroke fault identification method and system |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109101890A (en) * | 2018-07-16 | 2018-12-28 | 中国科学院自动化研究所 | Electrical energy power quality disturbance recognition methods and device based on wavelet transformation |
| CN109255313A (en) * | 2018-08-30 | 2019-01-22 | 中国科学院国家授时中心 | A kind of method of promotion signal recognition correct rate |
| CN109583337A (en) * | 2018-11-16 | 2019-04-05 | 华北电力大学（保定） | Electrical energy power quality disturbance recognition methods based on wavelet transformation |
| CN109918162A (en) * | 2019-02-28 | 2019-06-21 | 集智学园（北京）科技有限公司 | A kind of massive information higher-dimension graphic interaction methods of exhibiting learnt |
| CN109918162B (en) * | 2019-02-28 | 2021-11-02 | 集智学园（北京）科技有限公司 | High-dimensional graph interactive display method for learnable mass information |
| CN110363227A (en) * | 2019-06-24 | 2019-10-22 | 广东工业大学 | A kind of LED classification method based on manifold learning |
| CN111950696A (en) * | 2020-06-29 | 2020-11-17 | 燕山大学 | Short-term power load prediction method based on dimension reduction and improved neural network |
| CN112888008A (en) * | 2021-01-08 | 2021-06-01 | 南京中兴力维软件有限公司 | Base station abnormity detection method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107886085A (en) | A kind of electrical energy power quality disturbance feature extracting method based on t SNE | |
| Cao et al. | Introduction to text visualization | |
| Kantz et al. | Nonlinear time series analysis | |
| Feichtinger | Advances in Gabor analysis | |
| CN105335496B (en) | Customer service based on cosine similarity text mining algorithm repeats call processing method | |
| CN105389341B (en) | A kind of service calls repeat the text cluster and analysis method of incoming call work order | |
| Wang et al. | Computing linear transforms of symbolic signals | |
| US20050192805A1 (en) | Voice analysis device, voice analysis method and voice analysis program | |
| Zhang | Music feature extraction and classification algorithm based on deep learning | |
| CN108052863A (en) | Electrical energy power quality disturbance recognition methods based on the maximum variance method of development | |
| CN109446328A (en) | A kind of text recognition method, device and its storage medium | |
| CN109394258A (en) | A kind of classification method, device and the terminal device of lung's breath sound | |
| CN104200240A (en) | Sketch retrieval method based on content adaptive Hash encoding | |
| Shneiderman et al. | Interactive network exploration to derive insights: Filtering, clustering, grouping, and simplification | |
| CN106294331A (en) | Audio information retrieval method and device | |
| Lapidus et al. | Dirac operators and geodesic metric on the harmonic Sierpinski gasket and other fractal sets | |
| Byrnes et al. | Wavelets and their applications | |
| CN103892821A (en) | Emotion recognition model generating device based on electrocardiosignals and method thereof | |
| CN112151071A (en) | Speech emotion recognition method based on mixed wavelet packet feature deep learning | |
| CN102680838B (en) | Electric energy quality monitoring and distinguishing method and electric energy quality monitoring and distinguishing system based on dual-tree complex wavelet transform | |
| Zhang et al. | Temporal Transformer Networks for Acoustic Scene Classification. | |
| Chen et al. | Audio style transfer using shallow convolutional networks and random filters | |
| Dutta et al. | Pointwise information guided visual analysis of time-varying multi-fields | |
| CN114626424B (en) | Data enhancement-based silent speech recognition method and device | |
| Al-Sultany et al. | Enriching tweets for topic modeling via linking to the wikipedia |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180406 |