Deep fusion method based on multi-source data of transformer substation

Abstract

The invention relates to a deep fusion method based on multi-source data of a transformer substation, and the method comprises the steps: taking a direction with the maximum variance as a main feature according to a PCA algorithm, discretely marking data on a coordinate axis in each orthogonal direction, carrying out the dimension reduction of the data, carrying out the normalization processing of the data, obtaining a group of normalized data, and forming a normalized data set; carrying out dimension reduction on the normalized data to obtain data after dimension reduction; sVM multi-classification is carried out on the data after dimension reduction, then feature fusion is carried out, and screened data with prominent features are obtained. According to the method, the deep fusion speed based on the multi-source data of the transformer substation is increased, meanwhile, interference of redundant data on sample analysis can be reduced, a PCA data dimension reduction method is innovatively used, process noise and measurement noise are both considered in a data processing calculation scheme, meanwhile, an SVM multi-classification method is used for rapidly distinguishing the data, and the accuracy of the data processing calculation scheme is improved. And establishing an analysis chart and providing a strategy for sample learning.

Description

Technical Field

[0001] The present invention relates to the technical field of deep fusion of multi-source data, and in particular to a deep fusion method based on multi-source data of a substation. Background Art

[0002] The multi-source data of substation refers to various types of data involved in the operation and maintenance of substations. These data come from different systems and equipment and have different formats and characteristics. The data sources involved in the existing substation data acquisition equipment are complex and diverse, focusing on solving the problems of incomplete equipment status monitoring coverage, few mature monitoring methods, low operation stability, large maintenance workload, low application results of business systems, and low effective data utilization. In order to vigorously improve the autonomous controllability of substation systems and reduce the safety risks of system operation, deep fusion of data is a feasible solution. Deep fusion of multi-source data based on substations urgently requires a fast multi-source data processing algorithm to solve the problems of low data fusion efficiency, large restrictions, and poor analysis capabilities.

[0003] The existing deep fusion methods for multi-source data are basically single-dimensional data. They analyze the characteristics of the same type of data to obtain the characteristic performance of a basic type of data within a period of time. They have few limitations and require the data to express a single law. At the same time, the data fusion method is also very simple. Collecting data from a single source and making data statistical charts for classification and fusion often takes a lot of time and manpower. Summary of the invention

[0004] In order to solve the problems of slow processing of multi-source data and many features in the prior art, the purpose of the present invention is to provide a deep fusion method based on substation multi-source data that can speed up the deep fusion speed of substation multi-source data, reduce the interference of redundant data on sample analysis, and quickly process multidimensional data.

[0005] To achieve the above purpose, the present invention adopts the following technical solution: a deep fusion method based on multi-source data of substation, comprising:

[0006] (1) According to the PCA algorithm, the direction with the largest variance is taken as the main feature, the data is discretely marked on the coordinate axis in each orthogonal direction, the data is reduced in dimension, and then the data is normalized to obtain a set of normalized data to form a normalized data set;

[0007] (2) Perform dimensionality reduction on the normalized data to reduce data complexity and obtain dimensionality-reduced data;

[0008] (3) Perform SVM multi-classification on the data after dimensionality reduction, and then perform feature fusion to obtain the data with highlighted features after screening.

[0009] The step (1) specifically refers to: using the variance Var(x) to define the spacing between samples and calculating the variance:

[0010]

[0011] Among them, m is the number of samples, x i For the sample, is the mean;

[0012] Return the sample mean to 0, that is, subtract the sample mean from all samples;

[0013] Solve the maximum variance: find the direction of the maximum axis. After the direction vector w = (w1, w2) is mapped to the w direction, the variance maximum objective function is found. Find w so that maximum;

[0014] Where w1 represents the horizontal axis of the direction vector w, w2 represents the vertical axis of the direction vector w, and x i Represents a set of sample data; X is the sample data, and its formula is:

[0015]

[0016] Using the gradient ascent algorithm, solve w so that the distance from each variable in each set of data to the maximum plane dividing line is the largest:

[0017]

[0018] The step (2) specifically comprises the following steps in order:

[0019] (2a) Calculate the covariance matrix of the normalized data set;

[0020] (2b) Calculate the eigenvalues ​​and eigenvectors of the covariance matrix;

[0021] (2c) retain the most important k eigenvalues ​​of the key features in the fused data;

[0022] (2d) Find the eigenvectors corresponding to the k eigenvalues;

[0023] (2e) Multiply the m*n normalized data set by the feature vector n*k of k n-dimensional feature vectors to obtain the final reduced dimensionality data.

[0024] The step (3) specifically refers to: hashing the data vectors in a two-dimensional plane, and using a support vector machine (SVM) classifier to solve the best hyperplane by optimizing a convex quadratic programming problem;

[0025] Assume that the given feature space training set is:

[0026] T={(x1,y1),(x2,y2)...(x N ,y N )}

[0027] x i ∈R n ,y n ∈{+1,-1},i=1,2,...,N

[0028] Among them, (x i ,y i ) is the sample point

[0029] x i is the i-th sample

[0030] y i For x i Mark: y i =1, x i is a positive example; i = -1, x i is a negative example;

[0031] The training set minimizes the complexity of the plane mapping, that is, minimizes the sum of squares of weights, while limiting the misclassification of training samples;

[0032] Optimize a convex quadratic programming problem to solve the optimal hyperplane, using the Lagrange multiplier method; the separating plane is obtained by maximizing the margin:

[0033] y(x)=w T φ(x)+b

[0034] The classification decision function is:

[0035] f(x)=sign(w T φ(x)+b)

[0036] In the formula, sign is the kernel method, w T is the direction vector, φ(x) is the Lagrange multiplier method, and b is the y-axis offset;

[0037] For the nonlinearly separable case, the SVM support vector machine maps the input features to a high-dimensional space through a kernel function, so that the originally linearly inseparable data becomes linearly separable in the high-dimensional space. The kernel function is any one of a linear kernel, a polynomial kernel, and a Gaussian kernel.

[0038] It can be seen from the above technical scheme that the beneficial effects of the present invention are: first, the present invention accelerates the speed of deep fusion of multi-source data based on substations, while reducing the interference of redundant data on sample analysis. The present invention innovatively uses the PCA data dimensionality reduction method, taking both process noise and measurement noise into account in the data processing calculation scheme, and uses the SVM multi-classification method to quickly distinguish data, establish analysis diagrams and provide strategies for sample learning; second, the original variables are converted into new variables, namely principal components, through linear transformation. These principal components are independent of each other and can maximize the retention of variation information in the original data, reduce the dimension of the data set, simplify the data structure, and highlight the relationship between variables; third, the kernel function can be used to map to a high-dimensional space, and the kernel function can be used to solve nonlinear classification. The classification idea is very simple, which is to maximize the interval between the sample and the decision surface. The classification effect is good and multidimensional data can be processed quickly. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is an example diagram of the dimension reduction sample in the present invention;

[0040] Figure 2 is an example diagram of a hyperplane in the present invention;

[0041] Figure 3 It is the PCA algorithm flow chart of the present invention;

[0042] Figure 4 This is a flow chart of the SVM multi-classification algorithm in the present invention. DETAILED DESCRIPTION

[0043] A deep fusion method based on multi-source data of substation, comprising:

[0044] (1) According to the PCA algorithm, the direction with the largest variance is taken as the main feature, the data is discretely marked on the coordinate axis in each orthogonal direction, the data is reduced in dimension, and then the data is normalized to obtain a set of normalized data to form a normalized data set;

[0045] (2) Perform dimensionality reduction on the normalized data to reduce data complexity and obtain dimensionality-reduced data;

[0046] (3) Perform SVM multi-classification on the data after dimensionality reduction, and then perform feature fusion to obtain the data with highlighted features after screening.

[0047] The step (1) specifically refers to: using the PCA algorithm principle, taking the direction with the largest variance as the main feature, and “decorrelating” the data in each orthogonal direction, such as Figure 1 As shown, that is, they have no correlation in different orthogonal directions. Figure 3As shown, the variance Var(x) is used to define the spacing between samples and the variance is calculated:

[0048]

[0049] Among them, m is the number of samples, x i For the sample, is the mean;

[0050] Return the sample mean to 0, that is, subtract the sample mean from all samples;

[0051] Solve the maximum variance: find the direction of the maximum axis. After the direction vector w = (w1, w2) is mapped to the w direction, the variance maximum objective function is found. Find w so that maximum;

[0052] Where w1 represents the horizontal axis of the direction vector w, w2 represents the vertical axis of the direction vector w, and x i Represents a set of sample data; X is the sample data, and its formula is:

[0053]

[0054] Using the gradient ascent algorithm, solve w so that the distance from each variable in each set of data to the maximum plane dividing line is the largest:

[0055]

[0056] The step (2) specifically comprises the following steps in order:

[0057] (2a) Calculate the covariance matrix of the normalized data set;

[0058] (2b) Calculate the eigenvalues ​​and eigenvectors of the covariance matrix;

[0059] (2c) retain the most important k eigenvalues ​​of the key features in the fused data;

[0060] (2d) Find the eigenvectors corresponding to the k eigenvalues;

[0061] (2e) Multiply the m*n normalized data set by the feature vector n*k of k n-dimensional feature vectors to obtain the final reduced dimensionality data.

[0062] The step (3) specifically refers to: hashing the data vectors in a two-dimensional plane, such as Figure 2 As shown, the support vector machine SVM classifier solves the optimal hyperplane by optimizing a convex quadratic programming problem;

[0063] Assume that the given feature space training set is:

[0064] T={(x1,y1),(x2,y2)...(x N ,y N )}

[0065] x i ∈R n ,y n ∈{+1,-1},i=1,2,...,N

[0066] Among them, (x i ,y i ) is the sample point

[0067] x i is the i-th sample

[0068] y i For x i Mark: y i =1, x i is a positive example; i = -1, x i is a negative example;

[0069] The training set minimizes the complexity of the plane mapping, that is, minimizes the sum of squares of weights, while limiting the misclassification of training samples;

[0070] Optimize a convex quadratic programming problem to solve the best hyperplane, using the Lagrange multiplier method;

[0071] The separation plane is obtained by maximizing the margin:

[0072] y(x)=w T φ(x)+b

[0073] The classification decision function is:

[0074] f(x)=sign(w T φ(x)+b)

[0075] In the formula, sign is the kernel method, w T is the direction vector, φ(x) is the Lagrange multiplier method, and b is the y-axis offset;

[0076] The multi-classification algorithm of support vector machine can quickly classify the characteristics of multi-source data from discrete nonlinear distribution to a category with specific characteristics, making complex data observable and analyzable, distinguishing data according to the observation targets required by the substation, increasing the depth that can be represented by single-source data, and integrating the data change laws of different features to derive the required fusion attribute description, thereby completing the deep fusion research of multi-source data in substations.

[0077] like Figure 4As shown, for the case of nonlinear separability, the SVM support vector machine maps the input features to a high-dimensional space through a kernel function, so that the originally linearly inseparable data becomes linearly separable in the high-dimensional space. The kernel function is any one of a linear kernel, a polynomial kernel, and a Gaussian kernel.

[0078] In summary, the present invention accelerates the speed of deep fusion of multi-source data based on substations, while reducing the interference of redundant data on sample analysis. The present invention innovatively uses the PCA data dimensionality reduction method, taking both process noise and measurement noise into account in the data processing calculation scheme, and uses the SVM multi-classification method to quickly distinguish data, establish analysis graphs and provide strategies for sample learning; the original variables are converted into new variables, namely principal components, through linear transformation. These principal components are independent of each other and can retain the variation information in the original data to the maximum extent, reduce the dimension of the data set, simplify the data structure, and highlight the relationship between variables; the kernel function can be used to map to a high-dimensional space, and the kernel function can be used to solve nonlinear classification. The classification idea is very simple, which is to maximize the interval between the sample and the decision surface. The classification effect is good and multidimensional data can be processed quickly.

Claims

1. A deep fusion method based on multi-source data of substation, characterized by: The method comprises the following steps in order: (1) According to the PCA algorithm, the direction with the largest variance is taken as the main feature, the data is discretely marked on the coordinate axis in each orthogonal direction, the data is reduced in dimension, and then the data is normalized to obtain a set of normalized data to form a normalized data set; (2) Perform dimensionality reduction on the normalized data to reduce data complexity and obtain dimensionality-reduced data; (3) Perform SVM multi-classification on the data after dimensionality reduction, and then perform feature fusion to obtain the data with highlighted features after screening.

2. The deep fusion method based on substation multi-source data according to claim 1 is characterized by: The step (1) specifically refers to: using the variance Var(x) to define the spacing between samples and calculating the variance: Among them, m is the number of samples, x i For the sample, is the mean; Return the sample mean to 0, that is, subtract the sample mean from all samples; Solve the maximum variance: find the direction of the maximum axis. After the direction vector w = (w1, w2) is mapped to the w direction, the variance maximum objective function is found. Find w so that maximum; Where w1 represents the horizontal axis of the direction vector w, w2 represents the vertical axis of the direction vector w, and x i Represents a set of sample data; X is the sample data, and its formula is: Using the gradient ascent algorithm, solve w so that the distance from each variable in each set of data to the maximum plane dividing line is the largest:

3. The deep fusion method based on substation multi-source data according to claim 1 is characterized in that: The step (2) specifically comprises the following steps in order: (2a) Calculate the covariance matrix of the normalized data set; (2b) Calculate the eigenvalues ​​and eigenvectors of the covariance matrix; (2c) retain the most important k eigenvalues ​​of the key features in the fused data; (2d) Find the eigenvectors corresponding to the k eigenvalues; (2e) Multiply the m*n normalized data set by the feature vector n*k of k n-dimensional feature vectors to obtain the final reduced dimensionality data.

4. The deep fusion method based on substation multi-source data according to claim 1 is characterized in that: The step (3) specifically refers to: hashing the data vectors in a two-dimensional plane, and using a support vector machine (SVM) classifier to solve the best hyperplane by optimizing a convex quadratic programming problem; Assume that the given feature space training set is: T={(x1,y1),(x2,y2)...(x N ,y N )} x i ∈R n ,y n ∈{+1,-1},i=1,2,...,N Among them, (x i ,y i ) is the sample point x i is the i-th sample y i For x i Mark: y i =1, x i is a positive example; i = -1, x i is a negative example; The training set minimizes the complexity of the plane mapping, that is, minimizes the sum of squares of weights, while limiting the misclassification of training samples; Optimize a convex quadratic programming problem to solve the best hyperplane, using the Lagrange multiplier method; The separation plane is obtained by maximizing the margin: y(x)=w T φ(x)+b The classification decision function is: f(x)=sign(w T φ(x)+b) In the formula, sign is the kernel method, w T is the direction vector, φ(x) is the Lagrange multiplier method, and b is the y-axis offset; For the nonlinearly separable case, the SVM support vector machine maps the input features to a high-dimensional space through a kernel function, so that the originally linearly inseparable data becomes linearly separable in the high-dimensional space. The kernel function is any one of a linear kernel, a polynomial kernel, and a Gaussian kernel.

Images