CN109583337A - Electrical energy power quality disturbance recognition methods based on wavelet transformation - Google Patents

Abstract

The PQD recognition methods based on wavelet transformation that the present invention provides a kind of, to solve the problems, such as Power Quality Detection low efficiency in transmission and disttrbution system in the prior art.Described method includes following steps: building PQD signal model；PQD feature extraction parameter that MVU algorithm extracts wavelet transformation is chosen based on MVU and is compressed.MVU method is introduced into the PQD feature extraction based on wavelet transformation by the present invention, considers disturbance parameter randomness and influence of noise, carries out feature vector dimensionality reduction by MVU on the basis of extracting signal wavelet energy vector.Obtained low-dimensional PQD feature vector maintains the distributing edge of former data well, and information content is more concentrated.MVU algorithm reduces feature vector number and meets the constant constraint of local distance between k- nearest neighbor point, the distribution character of High Dimensional Data Set has been fully considered when choosing kernel function, the classification pressure for alleviating subsequent PQD, reduces the sort operation time, improves PQD recognition accuracy.

Description

Electric energy quality disturbance identification method based on wavelet transformation

Technical Field

The invention belongs to the field of detection and analysis of power systems, and particularly relates to a wavelet transform-based electric energy quality disturbance identification method.

Background

With the access of a large number of nonlinear loads and power electronic devices to the power grid, in order to ensure that the power quality meets the performance requirements of users, it is necessary to effectively detect and analyze the power quality in the power transmission and distribution system. Among them, a Power Quality Disturbance (PQD) signal is an important parameter for analyzing power quality, and a PQD recognition technology has become an important research direction in the field of power quality analysis. The goal of PQD identification is to quickly and accurately locate and identify PQDs from a vast amount of power quality data. The PQD recognition process includes two parts, feature extraction and pattern recognition. And the PQD feature extraction is the key point of PQD identification, and good PQD features can effectively improve the identification accuracy and reduce the calculation complexity.

The PQD feature extraction is to extract a feature quantity capable of reflecting the waveform feature of the disturbance signal by mapping. Currently, common feature extraction methods include: fourier transform, wavelet transform, hilbert-yellow transform, etc. The Fourier transform mainly reflects the overall information of the analysis signal, but the local characteristics of the signal are ignored, and the time locality is not provided for the non-stationary signal, so the time-frequency analysis requirement is not met. Although wavelet transform has been widely used for PQD feature extraction, which extracts feature vectors from wavelet decomposition coefficients of each layer, and is suitable for analyzing stationary and non-stationary signals, the wavelet transform algorithm in the prior art has high complexity.

Disclosure of Invention

In order to improve the accuracy of PQD identification and overcome the problem of low power quality detection efficiency in a power transmission and distribution system, the invention provides a power quality disturbance identification method based on wavelet transformation, which introduces a maximum variance expansion (MVU) nonlinear popular learning algorithm into the wavelet transformation, and performs feature vector dimension reduction through MVU on the basis of extracting signal wavelet energy vectors so as to well keep the distribution boundary of original data and enable the information quantity to be more concentrated, reduce the number of feature vectors and meet the constraint of unchanged local distance between k-nearest neighbors, reduce the classification operation time of subsequent PQD and improve the accuracy of PQD identification.

In order to achieve the purpose, the invention adopts the following technical scheme.

A PQD recognition method based on wavelet transform, the method comprising the steps of:

step S1, constructing a PQD signal model;

in step S2, the PQD feature vectors extracted by the wavelet transform are compressed based on MVU.

Further, the PQD signal model constructed in step S1 includes six PQD signals, namely, voltage hump, voltage sag, voltage discontinuity, harmonic, voltage pulse transient and voltage oscillation transient.

Further, the step S2 further includes:

step S21, the wavelet transformation of PQD obtains the original feature set of PQD;

step S22, introducing MVU algorithm to extract MVU features to obtain PQD feature vectors;

and step S23, selecting MVU algorithm to extract PQD characteristic parameters from the wavelet transform and compressing the parameters.

Further, the step S22 further includes:

step S221, selecting proper neighbor number k to construct adjacency matrix W_N×NIf x_iIs x_jK is adjacent, then W_ij1, otherwise W_ij0, wherein x_iIs a phase space data point, and N is the number of data points in the phase space;

and step S222, solving the optimization problem to obtain an optimal kernel matrix K.

According to the technical scheme provided by the embodiment of the invention, the MVU method is introduced into the PQD feature extraction based on wavelet transformation, disturbance parameter randomness and noise influence are considered, signals are generated according to a common PQD model, and feature vector dimension reduction is carried out through MVU on the basis of extracting signal wavelet energy vectors. The obtained low-dimensional PQD feature vector well keeps the distribution boundary of the original data, and the information amount is more concentrated. MVU the algorithm reduces the number of the feature vectors and satisfies the constraint of invariable local distance between k-nearest neighbors, fully considers the distribution characteristic of a high-dimensional data set when selecting the kernel function, lightens the classification pressure of subsequent PQD, reduces the classification operation time and improves the PQD identification accuracy.

Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a flow chart of a PQD identification method based on wavelet transformation according to an embodiment of the present invention;

FIG. 2 is a waveform diagram of an input signal analysis time period of 10 cycles according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a 10-scale decomposition of pure sinusoidal signals using db1-10 wavelets, respectively, according to an embodiment of the present invention;

FIG. 4 is a diagram illustrating the optimization result of the parameter k according to the embodiment of the present invention;

FIG. 5 is a diagram of the results of data processing using the Isomap algorithm in accordance with the present invention;

FIG. 6 is a diagram of the results of data processing using the LLE algorithm according to the embodiment of the present invention;

FIG. 7 is a diagram illustrating the result of data processing using the LTSA algorithm according to an embodiment of the present invention;

FIG. 8 is a diagram illustrating the result of data processing performed by the LPP algorithm according to the embodiment of the present invention;

fig. 9 is a diagram illustrating the result of data processing performed by MVU according to the embodiment of the present invention.

Detailed Description

Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention.

As used herein, the singular forms "a", "an", "the" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Further, "connected" or "coupled" as used herein may include wirelessly connected or coupled. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

For the convenience of understanding of the embodiments of the present invention, the following description will be further explained by taking specific embodiments as examples with reference to the drawings, and the embodiments are not to be construed as limiting the present invention.

The embodiment of the invention provides a power quality disturbance identification method based on wavelet transformation.

Wavelet transformation is a PQD feature extraction method and is widely applied to PQD feature extraction, and in the prior art, the transformation can extract feature vectors from wavelet decomposition coefficients of each layer, so that a good signal analysis effect can be obtained, but the algorithm complexity is high.

The data dimension reduction technology is divided into linear dimension reduction and nonlinear dimension reduction, and the nonlinear dimension reduction is divided into nonlinear dimension reduction based on a kernel method and a geometric structure. The nonlinear dimensionality reduction of the geometric structure, namely the popular learning, can effectively reduce the data scale and simultaneously reserve the nonlinear manifold structure of a high-dimensional space, thereby realizing dimensionality reduction. And the manifold learning is to find unknown mapping according to the high-dimensional observation data set and find the low-dimensional manifold representation corresponding to the high-dimensional observation data set without any prior knowledge. The main manifold learning algorithms at present are: an isometric mapping (ISOMAP) algorithm, a Local Linear Embedding (LLE) algorithm, a Local Tangent Space Arrangement (LTSA) algorithm, a local preserving mapping (LPP) algorithm, and a Maximum Variance Unfolding (MVU) algorithm. The MVU algorithm is a manifold learning algorithm based on manifold local equidistant concept, and when the Euclidean distance between adjacent points is kept unchanged, the high-dimensional data manifold can be expanded in a low-dimensional space through translation, rotation and other transformations.

The embodiment of the invention combines the PQD identification problem of wavelet transformation, introduces the MVU method into the PQD feature extraction based on wavelet transformation, considers disturbance parameter randomness and noise influence, generates signals according to a common PQD model, and performs compression and dimension reduction on feature vectors through MVU on the basis of extracting signal wavelet energy vectors. The obtained low-dimensional PQD feature vector well keeps the distribution boundary of the original data, and the information amount is more concentrated. MVU the algorithm reduces the number of the feature vectors and satisfies the constraint of invariable local distance between k-nearest neighbors, fully considers the distribution characteristic of a high-dimensional data set when selecting the kernel function, lightens the classification pressure of subsequent PQD, reduces the classification operation time and improves the PQD identification accuracy.

In the PQD recognition method based on wavelet transformation, the PQD signal is firstly subjected to wavelet decomposition to obtain the wavelet energy of the signal as an original feature set, then the original feature set is compressed through MVU algorithm, and since the kernel function is introduced into the algorithm to convert non-convex quadratic programming into the convex-semi positive optimization problem, the low-dimensional PQD feature with more concentrated information content and well maintained training data distribution boundary is obtained, and finally the PQD recognition is completed by combining with the classifier algorithm.

The nonlinear popular learning theory shows that MVU popular learning algorithm can well find the potential in R when the original PQD feature set needs to compress data and classify signal classes^dInner R^r. K is subjected to spectral decomposition to obtain characteristic values and corresponding characteristic vectors, the original vector space can be represented, and r is selected<d, i.e., the r-dimensional feature vector, retains all the useful features of the d-dimensional original PQD signal X and the amount of information is more focused. The above analysis shows that the features extracted based on the MVU algorithm are better than the original feature set with dispersed features in the classification identification process. Therefore, the method for identifying PQD based on MVU algorithm to extract PQD feature vector as the input of classifier is an improved PQD identification method.

Fig. 1 is a flowchart illustrating a PQD identification method based on wavelet transformation according to an embodiment of the present invention. As shown in fig. 1, according to an embodiment of the present invention, the method for identifying PQD based on wavelet transformation comprises the steps of:

step S1, constructing a PQD signal model;

in step S2, the PQD feature vectors extracted by the wavelet transform are compressed based on MVU.

In the step S1:

this example establishes a simulation model of the standard signal and 6 disturbance signals as shown in table 1.

TABLE 1 PQD Signal model

As shown in table 1, the mathematical formula model includes six common PQD signals (voltage hump, voltage sag, voltage discontinuity, harmonic wave, and voltage transient, where the voltage transient includes impulse transient and oscillation transient), where u (T) is a step function, α is amplitude, T1 and T2 are disturbance start time and disturbance end time, T is a signal cycle, ω is a fundamental wave angular frequency, the fundamental wave frequency is set to 50 Hz., and the analysis time length of the input signal is 10 cycles, i.e., 0.2s, and fig. 2 shows a waveform diagram of the input signal analysis time length of 10 cycles in this embodiment.

The step S2 further includes the following steps:

and step S21, the PQD original feature set is obtained through the wavelet transformation of the PQD.

Specifically, the method comprises the following steps: and (3) decomposing the original PQD signal by selecting a proper wavelet function to obtain an m-dimensional (m is 1,2,3, …, N) phase space.

At present, wavelet transform is a common method for extracting feature vectors in PQD original signals. The multiresolution analysis is the basic theory of wavelet analysis, and the Mallat algorithm can quickly realize wavelet transformation. Which achieves decomposition by reusing a low-pass filter and a high-pass filter. The low frequency component and the high frequency component obtained by the filter occupy 1/2 of the signal band. And repeating the decomposition process on the obtained new low-frequency component to obtain the high-frequency component and the low-frequency component of the next layer.

According to Parseval's theorem, the energy wavelet coefficient formula is as follows:

∫[f(t)]²dt＝∑[a_j(k)]²+∑[d_j(k)]²(1)

in the formula (1), f (t) is the signal to be decomposed, a_j(k) For wavelet decomposition of approximation coefficients of the j-th layer, d_j(k) The detail coefficients of the j-th layer are decomposed for wavelets. Performing J-layer decomposition on the PQD, wherein the wavelet transform approximate energy distribution and the detail energy distribution of the f (t) signal are respectively defined as:

in the expressions (2) and (3), J is 1,2, … and J, J +1 characteristic quantities are obtained through J-layer wavelet decomposition, and the characteristic quantities form an energy function

Compared with normal signals, the energy of the wavelet transformation coefficient at each scale corresponding to each PQD is different, so that the disturbance can be subjected to multi-scale decomposition, and then a feature vector is constructed according to the energy characteristics. Because of the orthogonality and the compactness of the Daubechies (db) series wavelets, the method is sensitive to irregular signals, is very suitable for transforming PQD signals, and can obtain the best db4 wavelet processing capacity. A pure sinusoidal signal (without noise) with a sampling rate of 6.4khz was analyzed with the db orthogonal wavelet basis mentioned above. FIG. 3 is a schematic diagram of a 10-scale decomposition of pure sinusoidal signals using db1-10 wavelets, respectively. As shown in fig. 3, the abscissa is the decomposition scale and the ordinate is the energy, and the maximum energy value of the signal decomposition energy occurs at the 7 th decomposition scale. Using the above formula to perform db4 wavelet 7-layer decomposition, only extracting the minutiaeThe energy-saving part is composed of 7 characteristic quantities to form a vectorThe original vector set of the PQD signal is constructed as data samples for the next operation MVU.

And step S22, introducing MVU algorithm to extract the feature of MVU to obtain the PQD feature vector.

The method specifically comprises the following steps:

step S221, selecting proper neighbor number k to construct adjacency matrix W_N×NIf x_iIs x_jK is adjacent, then W_ij1, otherwise W_ij0, wherein x_iIs the phase space data point, and N is the number of data points in the phase space.

Suppose a high dimensional space R^dX ═ X (X) in the observation data set (c)₁,x₂,...,x_n)^TIs obtained by up-sampling an R-dimensional manifold M embedded in a d-dimensional space, and the manifold learning is to find an unknown mapping f: R according to a high-dimensional observation data set X under the condition of no prior knowledge about M and R^d→R^r(r < d), and find the low-dimensional manifold representation Y ═ Y (Y) in one-to-one correspondence with the high-dimensional observation dataset X₁,y₂,...y_n)^T,y∈R^r。

MVU the algorithm first constructs an n adjacency matrix W based on X, if X_iIs x_jK is adjacent, then W_ij1, otherwise W_ij0. The MVU algorithm can be expressed as an optimization problem as follows:

s.t.||y_i-y_j||²W_ij＝||x_i-x_j||²W_ij,

in the equation (4), the first constraint function is a local equidistant constraint, which ensures that when a data point in a high-dimensional space is mapped to a low-dimensional space, the euclidean distance between adjacent points remains unchanged, and thus the local structure of the data set is preserved. The second is a centering constraint to eliminate translational degrees of freedom.

The method is a non-convex quadratic programming problem under the constraint of a quadratic equation and is easy to fall into a local optimal solution. The method can be converted into a convex-semi positive definite optimization problem by introducing a kernel function, and a kernel matrix for defining the data set Y is K ═ K_ij]_n×nIts element is K_ij＝<y_i,y_j>Wherein<·,·>This indicates that the inner product is obtained. At this point, the above optimization problem can be transformed into:

max trace(K)

s.t.K≥0,

K_ii-2K_ij+K_jj＝||x_i-x_j||²if W is_ij＝1 (5)

In equation (5), trace (. cndot.) represents the trace of the matrix, and the first added constraint K ≧ 0 indicates that K is a semi-positive definite matrix, which is used to ensure that the data comes from the convex set.

And step S222, solving an optimization problem formula (5) to obtain an optimal kernel matrix K.

The solution of the convex-half positive definite optimization problem (5) has a mature rapid algorithm, and is not described in detail herein.

And step S223, performing spectral decomposition on the K, and calculating the coordinate representation of the high-dimensional manifold in an r-dimensional space by formulas (7) and (8) according to the K characteristic value and the corresponding characteristic vector.

Let K^*For the optimal solution of the optimization problem, for K^*Performing spectrum decomposition to obtain

Wherein λ_αIs a matrix K^*α th characteristic value, V_αiIs the ith element of the corresponding feature vector. Thus, the high dimensional data x_iN-dimensional mapping ofα th element of

So far, r-dimensional embedding of d-dimensional observation data xi can be represented by equation (8):

MVU is a popular algorithm that can automatically learn the training data kernel matrix. And solving a semi-positive definite plan to obtain an optimal kernel matrix of the training data. The variance of the training data mapped into the kernel feature space is maximized under the constraint that the local distance between k-nearest neighbors is constant. Thus, non-linear structures in potentially high-dimensional data are mapped into the corresponding kernel feature space and make the mapping linear. Since the MVU algorithm only needs to keep the dataset locally equidistant, MVU can handle non-convex manifolds efficiently and with a more focused amount of information. The distribution characteristic of the high-dimensional data set is fully considered when the MVU algorithm takes the kernel function, so that the nonlinear distribution data set can meet linear distribution after being mapped to the feature space, and the distribution boundary of the training data set is well maintained.

And step S23, selecting MVU algorithm to extract PQD characteristic parameters from the wavelet transform and compressing the parameters.

Two important parameters, namely, the estimated dimension parameter r and the number k of neighboring points, can be obtained from steps S21 and S22, and the two parameters are selected respectively to relate to the eigen-dimension estimation and the neighborhood parameter selection based on the PQD signal. Wherein,

the eigen-dimension estimation selection process is as follows:

on the premise of ensuring no information loss, free variables of data are represented by the minimum variable number, which is called intrinsic dimensionality of the data set and is an inherent property of the data set. The distribution property in the neighborhood of points in the phase space is determined by the eigen-dimension of the signal, and therefore the eigen-dimension of the signal can be determined by analyzing the subspace spanned by the neighborhood of points. In the reconstruction phase space, the noise is distributed over the entire high-dimensional phase space, while the useful signal is distributed essentially in a feature space of the size of the eigen dimension. Therefore, when the manifold learning is adopted for dimensionality reduction, if the reduced target dimension is selected too large, the noise distribution is possibly not ideal to eliminate; if the reduced target dimension is chosen too small, the useful signal may be eliminated.

The Maximum Likelihood Estimation (MLE) method for estimating the intrinsic dimension of data is to estimate the intrinsic dimension of data by establishing a likelihood function of the distance between a pair of neighboring points. Let x₁，x₂，…，x_nConstructing a binomial random process for high-dimensional sample data, and establishing a likelihood function of the distance between adjacent points. For a given k, traverse the estimate of the eigen-dimension through x₁，x₂，…，x_nThereby obtaining n estimated values m of local intrinsic dimensions_k(x_i) And taking the average value as an estimated value of the global intrinsic dimension:

e for each PQD signal_DTo carry outMLE eigen-dimension estimation, and the result is that the estimated dimension r is 3.

The neighborhood parameter selection process is as follows:

the neighborhood selection parameters and the number k of the neighboring points have great influence on the performance of the manifold learning algorithm. This is because a large number of nearest neighbors can contribute to the elimination of manifold small-scale structures and the smoothing of the entire manifold. Conversely, too few neighborhoods may misclassify a continuous manifold into disjointed sub-manifolds. Because the parameters of the algorithm are less and independent of each other, and the searching complexity is not high, the number k of all adjacent points is traversed in the searching range, and the most appropriate parameter value is found under a certain constraint condition, so that the optimal neighborhood parameter k is searched.

E for PQD signal_DWhen MVU popular algorithm feature extraction calculation is carried out, when r is 3, the parameter k traverses the adjacent parameter k in the search range of 10-20, the search step length is 2, and the constraint condition is that the training accuracy and the testing accuracy are highest when an SVM classifier is used for classification, so that the optimal neighborhood parameter k is searched. The optimization results for parameter k are shown in fig. 4. As shown in fig. 4, when the k value varies from 10 to 20, the model training accuracy and the test accuracy are the highest, respectively 100% and 99.52%, when k is 12 and k is 14.

In order to avoid excessive consumption of computing resources, it is sufficient that the k value at least well maintains the characteristics of the local low-dimensional manifold of the dataset. Table 2 compares the feature extraction running times of 700 PQD signals when k is 12 and k is 14, and it can be seen that the time consumption is less when k is 12. Therefore, the algorithm parameters r-3 and k-12 were chosen MVU during the experiment.

TABLE 2 run time comparison

Tab.2 Run time comparison

In order to verify the effectiveness and robustness of the feature Vector extracted by the wavelet transform-based PQD recognition method, a classical classification algorithm K Nearest Neighbor method (KNN, K-Nearest Neighbor) and a Support Vector Machine (SVM) method are respectively selected for classification verification. The KNN classification algorithm is one of the simplest methods in data mining classification technology. The method has the advantages of simple algorithm, easy understanding, easy realization and no need of parameter estimation and classification training. The SVM is a new pattern recognition method developed on the basis of a statistical learning theory. In the embodiment, a one-to-one SVM is selected to construct a plurality of two classifiers to realize PQD classification, wherein a Gaussian kernel function in a radial basis function is selected as a kernel function of the SVM. As the SVM algorithm needs pre-learning, training samples and test samples are respectively input for experimental classification.

PQD data was generated according to the signal model provided in table 1. Taking voltage sinusoidal signals, voltage bulges, voltage recesses, voltage discontinuities, harmonics and voltage transient states, wherein the voltage transient states comprise 100 disturbance signals of 6 types of pulse transient states and oscillation transient states respectively, and 700 samples are obtained in total; the analysis time length of the input signal is 10 cycles (0.2s), the sampling rate is 6.4kHz, the sampling points of the signal are 1280, and each cycle is 128 points; in order to better simulate various actual conditions and ensure the reliability of an analysis result, parameters of each disturbance such as disturbance starting and stopping time, amplitude and duration are randomly changed within an allowable range, and random white noise with the signal-to-noise ratio of 30dB is added to a disturbance signal. Sampling the PQD signal yields 7 sets of sample data of 100 × 1280. And (3) realizing the PQD feature extraction based on MVU algorithm by applying the data, and respectively selecting KNN and SVM classifiers for algorithm verification. When the SVM classifier is used for classification, 70 samples of each disturbance are respectively taken as a training set, and 30 samples are taken as a test set.

In order to verify that the MVU method can effectively solve the problems of relevance and redundancy of PQD feature extraction, MVU algorithm is adopted for E_DData processing is performed to extract a 3-dimensional feature vector of PQD. And ISOMAP, LLE, LTSA and LPP algorithms are compared with the algorithm in the text for feature extraction. First, the energy value structure by db4 wavelet 7-layer decompositionAnd (3) making sample vector data, and then performing data processing by respectively adopting Isomap, LLE, LTSA, LPP and MVU algorithms so as to extract the PQD 3-dimensional feature vector. The visual inset diagrams are shown in fig. 5, 6, 7, 8, and 9, respectively. The sample representation method of different classes of PQD is shown in the figure.

Fig. 5, 6, and 7 are visual embedding diagrams for PQD feature extraction using Isomap, LLE, and LTSA algorithms, respectively. It can be seen that the embedding results of the three dimension reduction methods are not ideal, the serious data congestion problem occurs in the graph, and different types of samples in the embedding results are mutually overlapped and can hardly show the mutual separation condition.

Fig. 8 is a three-dimensional scattergram for PQD feature extraction using the LPP method. The figure shows that the LPP method divides the PQD signal into two types of aliasing, i.e., the voltage sine, convex and concave signals are crowded and mixed together and cannot be distinguished; other signals are crowded together and indistinguishable.

Fig. 9 is an embedded diagram of the PQD feature vector extraction result of the MVU algorithm. Since the MVU algorithm better maintains the distribution boundary of the training data set, the separation degree of the 3-dimensional scatter point dimensionality reduction result is better, and the 3-dimensional result can be projected to a 2-dimensional plane for displaying the result more clearly. From fig. 9, the spatial distribution of 7 different types of sample data after dimensionality reduction can be obtained, the normal signal and the 6 PQD signals can be distinguished obviously, and the distribution boundary of training data is well maintained while dimensionality reduction is completed.

Compared with other classical nonlinear popular learning algorithms, the MVU algorithm not only completes dimension reduction of the PQD feature vector, but also well solves the data 'crowding problem' of the popular algorithm in the PQD analysis. This is because the MVU algorithm attempts to solve the problem of how to select the kernel function by learning the kernel matrix. The MVU algorithm defines a neighborhood map on the data and keeps pairwise distances in the result map to learn the kernel matrix. Prevalence expansion is achieved by maximizing the euclidean distance between data points, under the constraint that the distance in the neighborhood graph remains constant (i.e., under the constraint that the local geometry of the data prevalence is constant). The resulting optimization problem can be solved using semi-positive planning.

In conclusion, the PQD feature extraction method based on the MVU algorithm can maintain the hidden low-dimensional structure in the sample high-dimensional space and embed the potential manifold into the low-dimensional space. And MVU the low-dimensional data after dimension reduction well maintains the distribution boundary of the training data. From the analysis of the experimental results, it can be seen that the dimension reduction effect of the MVU algorithm on the PQD feature extraction is superior to that of other popular learning algorithms.

In order to verify the effectiveness and robustness of the PQD identification method based on wavelet transformation, the KNN and the SVM classifiers are respectively adopted to carry out classification verification on the obtained PQD characteristic vectors.

The recognition effect of PQD feature extraction based on the MVU algorithm was verified using the KNN classification algorithm. In the experimental process, any k value is taken in a given range, and the recognition results of the PQD feature vectors obtained based on the MVU algorithm are all 100%. The above results show that the feature vector has a definite classification boundary, so that the requirement for the subsequent classification algorithm is extremely low, the classifier parameter selection is not sensitive, and the feature vector is a strong and robust PQD feature.

A multi-classification SVM classifier is constructed by adopting a one-to-one method to verify the recognition effect of the PQD feature extraction based on MVU. The kernel function of the SVM adopts a Gaussian radial basis kernel function. Wherein, the scale parameter of the kernel function is 0.1, and the regularization parameter is 100. The first 70 data for each perturbation were used for training of the SVM classifier. The latter 30 data were used to test the accuracy of the classification. The training set classification result of the SVM classifier is 100%. The test set classification result for the SVM classifier was 99.52%, as shown in table 3.

TABLE 3 SVM test set Classification results based on MVU Algorithm

Tab.3 Classification results of SVM test sets based on MVU

As can be seen from the PQD feature extraction, sampling the PQD signal results in 7 sets of sample data of 100 × 1280. After 7-layer wavelet decomposition, the wavelet energy data of each disturbance is a 100 × 7 matrix, and dimension reduction is performed by an MVU algorithm to obtain a 100 × 3 matrix. Thereby achieving the purpose of reducing the dimension of the sample. Respectively selecting MVU algorithm to reduce dimension X_100×7Reduced dimension Y for eigenvector sum MVU algorithm_100×3The results of identifying the two feature vectors are shown in table 4.

TABLE 4 MVU PQD recognition results before and after algorithm dimensionality reduction

Tab.4 Comparison of PQD Recognition Results

According to the experimental results, before dimension reduction of MVU algorithm, the wavelet energy ED is used as a feature vector, an SVM classifier is used for classifying PQD, the training and testing accuracy is lower than that of MVU algorithm, and the training and testing running time is higher than that of MVU algorithm. The experimental result shows that the dimensionality of the PQD feature vector processed by the MVU algorithm is reduced, so that the running time is reduced in the identification process; the PQD feature vector processed by the MVU algorithm well keeps the distribution boundary of training data, and the information quantity is more concentrated, so that the classification accuracy is improved.

Therefore, the PQD identification method based on wavelet transformation in the embodiment of the invention well completes the low-dimensional embedded representation of the PQD signal under the consideration of the influence of random parameters and noise, and is an ideal feature extraction method. By constructing a universal classifier, the effectiveness of the PQD feature vectors extracted by MVU algorithm is verified, which can reduce the classification running time and improve the PQD recognition rate. MVU the low-dimensional PQD feature vector after dimensionality reduction well keeps the distribution boundary of training data, and the information quantity is more concentrated than other popular algorithms, and the classification pressure of subsequent PQD can be reduced while dimensionality reduction is carried out.

Those of ordinary skill in the art will understand that: the figures are merely schematic representations of one embodiment, and the blocks or flow diagrams in the figures are not necessarily required to practice the present invention.

The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for apparatus or system embodiments, since they are substantially similar to method embodiments, they are described in relative terms, as long as they are described in partial descriptions of method embodiments. The above-described embodiments of the apparatus and system are merely illustrative, and the units described as separate parts may or may not be physically separate, and the parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

Those of ordinary skill in the art will understand that: the components in the devices in the embodiments may be distributed in the devices in the embodiments according to the description of the embodiments, or may be correspondingly changed in one or more devices different from the embodiments. The components of the above embodiments may be combined into one component, or may be further divided into a plurality of sub-components.

The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (4)

1. A PQD identification method based on wavelet transformation is characterized by comprising the following steps:

step S1, constructing a PQD signal model;

in step S2, the PQD feature vectors extracted by the wavelet transform are compressed based on MVU.

2. The PQD recognition method according to claim 1, wherein the PQD signal model constructed in step S1 includes six PQD signals, i.e., voltage humps, voltage notches, voltage discontinuities, harmonics, voltage pulse transients and voltage oscillation transients.

3. The PQD recognition method according to claim 1 or 2, wherein the step S2 further includes:

step S21, the wavelet transformation of PQD obtains the original feature set of PQD;

step S22, introducing MVU algorithm to extract MVU features to obtain PQD feature vectors;

and step S23, selecting MVU algorithm to extract PQD characteristic parameters from the wavelet transform and compressing the parameters.

4. The PQD recognition method according to claim 3, wherein the step S22 further includes:

step S221, selecting proper neighbor number k to construct adjacency matrix W_N×NIf x_iIs x_jK is adjacent, then W_ij1, otherwise W_ij0, wherein x_iIs a phase space data point, and N is the number of data points in the phase space;

and step S222, solving the optimization problem to obtain an optimal kernel matrix K.