CN111242204A - A fault feature extraction method for operation and maintenance management and control platform - Google Patents

Abstract

A fault feature extraction method for an operation and maintenance management and control platform is characterized in that it includes: principal component analysis feature extraction, secondary feature selection, and the like. Feature extraction based on principal component analysis can transform high-dimensional space samples into low-dimensional space samples, reduce the redundancy of feature attributes while reducing the feature dimension, and retain the main classification information, which greatly reduces the computational complexity of the classifier , shortens the training time; and because the secondary feature selection function is embedded in the feature extraction process, the evaluation results are sorted based on the association rule feature selection, combined with the heuristic sequence backward search strategy, and then the key features of the feature subset are determined. Making the feature attributes have the maximum correlation-minimum redundancy, that is, the correlation between attribute features and class attributes can be maximized, and the redundancy between attributes can be reduced, and the classification accuracy of management and control faults can be significantly improved. The method is scientific and reasonable, has strong applicability, and can be widely applied to various fault classification management and control platforms.

Description

A fault feature extraction method for operation and maintenance management and control platform

technical field

The invention relates to the technical field of information system operation and maintenance management and control fault feature extraction, and is a fault feature extraction method for an operation and maintenance management and control platform.

Background technique

In order to obtain information such as system operation status and operation trend, the information system management and control platform conducts real-time remote monitoring of hardware equipment and software applications. The monitoring of equipment by the management and control platform needs to be carried out in a network environment. In a network environment, data transmission usually brings corresponding characteristics to the data flow, and these characteristics are an important basis for realizing data identification. When the management and control equipment is monitored, a large amount of fault information will be collected, and feature extraction and selection technology is the basis for classifying and identifying the fault information. Feature extraction and selection technology can realize the selection of key monitoring features in a multi-attribute and highly redundant information environment.

In the information system intelligent management and control platform, in order to strengthen the centralized management and unified monitoring of the system, by realizing the whole network monitoring of the network and security equipment, it provides accurate fault judgment and processing suggestions, and improves the ability and efficiency of personnel to solve faults. In order to achieve this goal, feature extraction and selection techniques are used to determine the key features of monitoring fault data. Each fault type may contain many features, and the key features that best represent such fault types are selected. The advantage of feature extraction and selection technology is that in the process of fault type identification and classification, the accuracy of fault identification is greatly improved under the premise of reducing data redundancy. Compared with other techniques, it can more accurately select the key features that best represent this type of failure.

Through feature extraction and selection technology, it can effectively identify and classify fault types, so as to analyze and deal with faults quickly and efficiently.

The fault data of the operation and maintenance management and control platform contains data with many characteristics, which is called high-dimensional data. The fault types are automatically classified based on some features of the high-dimensional data, but some features in the fault data do not contribute much to the classification results. In addition, due to certain correlation and redundancy between features, large time and space overheads are incurred in the classification process, resulting in poor fault classification results. Redundant features of high-dimensional data greatly affect the performance of classifiers, especially standard supervised learning classification algorithms that use all data features as decision functions. Therefore, for a classifier based on supervised learning, the features of the original data are extracted or selected before classification, which reduces the redundancy of the data and can effectively improve the generalization ability of the classifier. At present, there are hundreds of fault statistical features for fault classification on the management and control platform. In order to improve the efficiency and accuracy of the classification algorithm and effectively reduce the redundancy between the scale and features of the original data, it is necessary to select and extract the features of the original high-dimensional data. Feature selection is to select an optimal feature subset from the original data features, which can represent the distribution characteristics of the original data to the greatest extent; feature extraction is to map high-dimensional data samples into low-dimensional data samples through transformation through the mapping principle dimensional samples, after mapping, a new sample feature combination is formed. This combination not only reduces the dimension, but also fully represents the original features due to the mapping transformation.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the problem of insufficient removal of redundant information by simply adopting the feature selection method when the similarity between data is strongly dependent, and to provide a scientifically reasonable and highly applicable method that can determine the feature subset in the situation In this way, the fault feature extraction method of the operation and maintenance management and control platform can more effectively remove data redundancy and achieve better classification accuracy.

The purpose of the present invention is achieved by the following technical solutions: a method for extracting fault features of an operation and maintenance management and control platform, which is characterized in that the content it includes includes:

1) Principal Component Analysis Feature Extraction

Principal Component Analysis (PCA) is to transform the sample space, determine the projection direction with the largest variance of all the original eigenvectors through projection, and perform feature extraction on the locating and discriminating vector of the projection direction. After the projection transformation, the original sample becomes as scattered as possible. while maintaining the difference of the original high-dimensional sample space before transformation, it is assumed that the original high-dimensional space contains N samples, X∈R ⁿ , and each sample is a X _i =[x _i1 ,..., x _in ] ^T ∈R ⁿ , its vector mean is M, then the corresponding eigenvector is X _i =[x _1i ,...,x _ni ]∈R ⁿ , and the corresponding covariance matrix is formula (1),

The distribution variance of the sample on the eigenvector, that is, the eigenvalue of the covariance matrix of formula (1), the orthogonal matrix obtained by diagonalizing the covariance matrix in formula (1) is formula (2),

其中M为正交矩阵Q的维数，PCA则基于Q推导出矩阵中的特征值λ₁≥λ₂≥…≥λ_n，并求出特征值对应的标准正交特征向量v₁≥v₂≥…≥v_n，通过正交矩阵Q的特征值及相应的标准正交特征向量，即得到协方差矩阵S的标准正交特征向量u₁,u₂,…u_d，如公式(3)，其中标准正交特征向量u₁,u₂,…u_d对应S的前d个最大非零特征值，Denote Q as

Where M is the dimension of the orthogonal matrix Q, PCA derives the eigenvalues λ ₁ ≥λ ₂ ≥...≥λ _n in the matrix based on Q, and obtains the standard orthogonal eigenvectors v ₁ ≥ v ₂ corresponding to the eigenvalues ≥...≥v _n , through the eigenvalues of the orthogonal matrix Q and the corresponding standard orthogonal eigenvectors, the standard orthogonal eigenvectors u ₁ , u ₂ ,... _ud of the covariance matrix S are obtained, as shown in formula (3) , where the standard orthogonal eigenvectors u ₁ , u ₂ ,… _ud correspond to the first d largest non-zero eigenvalues of S,

Set t=95%, u _i > t, then the cumulative contribution rate of the principal components of the spatial samples on the first d axes is 95% of the original data, so that for any sample _xi , it is mapped to the reduced dimension. The low-dimensional sample space U={u ₁ , u ₂ ,...u _d }, the principal component feature of x _i is y=(u ₁ , u ₂ ,... u _d ) ^T x _i , then y _i is a low-dimensional space The sample points in , through the spatial sample transformation of PCA, not only the transformed samples can represent 95% of the cumulative contribution rate of the principal components, but also the original spatial dimension is reduced from n to d, d<<n, thus greatly reducing the It has the dimension of space and played the role of feature extraction;

2) Quadratic feature selection

After the PCA feature extraction, in order to further obtain the optimal feature subset and the key features of the PCA low-dimensional space, a quadratic feature selection algorithm is embedded, which is based on the filter association rule feature selection (Correlation-based Feature Selection, CFS), When evaluating the correlation of sample features, a heuristic sequence backward search strategy is used, and the optimal feature subset is determined by sorting the correlation of features.

属性间的相关度均值则用

表示，由公式(4)可知，由关联规则特征选择算法确定的候选特征子集能够使特征属性具有最大关联—最小冗余，即可以最大程度提高属性特征与类属性关联度，且降低属性与属性之间的冗余，即公式(4)中评价值M_s越高，特征属性与类的相关度均值为

越大，属性间的相关度均值则用

越小，CFS takes the correlation of features as the evaluation standard, and is a filter-type feature selection algorithm. Under the corresponding search strategy, it aims to reduce the redundancy between attributes and improve the association between attribute features and class attributes. To achieve the screening of attributes with high redundancy and attributes that are not related to categories, formula (4) is its evaluation standard, and the evaluation of k features of feature subset S is represented by M _s , where the correlation between feature attributes and classes mean is

The mean correlation between attributes is used

It can be seen from formula (4) that the candidate feature subset determined by the association rule feature selection algorithm can make the feature attributes have the maximum correlation-minimum redundancy, that is, it can maximize the degree of association between attribute features and class attributes, and reduce the relationship between attributes and attributes. The redundancy between attributes, that is, the higher the evaluation value M _s in formula (4), the mean of the correlation between the feature attribute and the class is

The larger the value, the mean of the correlation between attributes is used

the smaller the

In the feature selection of association rules, the information gain algorithm is used to evaluate the correlation between attributes, and the calculation method of information gain is the measurement method of symmetry, so when two high-order related features in the feature subset S exist, such as feature W _i , W _j , the symmetric uncertainty method of formula (5) can be used, the entropy of the feature is H(W), and the feature correlation is U, so formula (6) is the evaluation of the feature subset based on the correlation between attributes function, when the evaluation value H _s increases _, the correlation between the feature W _j and Wi in the feature subset S decreases, and the correlation with the class attribute increases,

Using the CFS algorithm, the secondary feature selection function is embedded in the PCA, and then based on the heuristic sequence backward search strategy, the evaluation results of the CFS are calculated, and the optimal feature subsets are filtered out after sorting.

A fault feature extraction method for an operation and maintenance management and control platform of the present invention is a feature extraction method embedded with a secondary feature selection function, because based on PCA feature extraction, the high-dimensional space samples are transformed into low-dimensional space samples, and in the case of reduced feature dimensions At the same time, the redundancy of feature attributes is reduced, and the main classification information is retained, which greatly reduces the computational complexity of the classifier and shortens the training time; and because the secondary feature selection function is embedded in the feature extraction process, based on CFS Combine the heuristic sequence backward search strategy to sort the evaluation results, and then determine the key features of the feature subset, so that the feature attributes have the maximum correlation-minimum redundancy, that is, the correlation between attribute features and class attributes can be maximized, and the attributes can be reduced. The redundancy between attributes significantly improves the classification accuracy of management and control faults. The method is scientific and reasonable, has strong applicability, and can be widely applied to various fault classification management and control platforms.

Description of drawings

1 is a functional schematic diagram of a method for extracting fault features of an operation and maintenance management and control platform of the present invention;

Fig. 2 is a flow chart of a feature backward search strategy embedded with a secondary feature selection function;

Figure 3 is a comparison chart of fault classification performance before and after feature extraction based on the initial PCA;

Figure 4 is a performance comparison diagram of the PCA feature extraction method embedded with the secondary feature selection function and the traditional feature extraction method.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

A method for extracting fault features of an operation and maintenance management and control platform of the present invention includes the following contents:

1) Principal Component Analysis Feature Extraction

Principal Component Analysis (PCA) is to transform the sample space, determine the projection direction with the largest variance of all the original eigenvectors through projection, and perform feature extraction on the locating and discriminating vector of the projection direction. After the projection transformation, the original sample becomes as scattered as possible. while maintaining the difference of the original high-dimensional sample space before transformation, it is assumed that the original high-dimensional space contains N samples, X∈R ⁿ , and each sample is a X _i =[x _i1 ,..., x _in ] ^T ∈R ⁿ , its vector mean is M, then the corresponding eigenvector is X _i =[x _1i ,...,x _ni ]∈R ⁿ , and the corresponding covariance matrix is formula (1),

The distribution variance of the sample on the eigenvector, that is, the eigenvalue of the covariance matrix of formula (1), the orthogonal matrix obtained by diagonalizing the covariance matrix in formula (1) is formula (2),

其中M为正交矩阵Q的维数，PCA则基于Q推导出矩阵中的特征值λ₁≥λ₂≥…≥λ_n，并求出特征值对应的标准正交特征向量v₁≥v₂≥…≥v_n，通过正交矩阵Q的特征值及相应的标准正交特征向量，即得到协方差矩阵S的标准正交特征向量u₁,u₂,…u_d，如公式(3)，其中标准正交特征向量u₁,u₂,…u_d对应S的前d个最大非零特征值，Denote Q as

Where M is the dimension of the orthogonal matrix Q, PCA derives the eigenvalues λ ₁ ≥λ ₂ ≥...≥λ _n in the matrix based on Q, and obtains the standard orthogonal eigenvectors v ₁ ≥ v ₂ corresponding to the eigenvalues ≥...≥v _n , through the eigenvalues of the orthogonal matrix Q and the corresponding standard orthogonal eigenvectors, the standard orthogonal eigenvectors u ₁ , u ₂ ,... _ud of the covariance matrix S are obtained, as shown in formula (3) , where the standard orthogonal eigenvectors u ₁ , u ₂ ,… _ud correspond to the first d largest non-zero eigenvalues of S,

Set t=95%, u _i > t, then the cumulative contribution rate of the principal components of the spatial samples on the first d axes is 95% of the original data, so that for any sample _xi , it is mapped to the reduced dimension. The low-dimensional sample space U={u ₁ , u ₂ ,...u _d }, the principal component feature of x _i is y=(u ₁ , u ₂ ,... u _d ) ^T x _i , then y _i is a low-dimensional space The sample points in , through the spatial sample transformation of PCA, not only the transformed samples can represent 95% of the cumulative contribution rate of the principal components, but also the original spatial dimension is reduced from n to d, d<<n, thus greatly reducing the It has the dimension of space and played the role of feature extraction;

2. Quadratic Feature Selection

After the PCA feature extraction, in order to further obtain the optimal feature subset and the key features of the PCA low-dimensional space, a quadratic feature selection algorithm is embedded, which is based on the filter association rule feature selection (Correlation-based Feature Selection, CFS), When evaluating the correlation of sample features, a heuristic sequence backward search strategy is used, and the optimal feature subset is determined by sorting the correlation of features.

属性间的相关度均值则用

表示，由公式(4)可知，由关联规则特征选择算法确定的候选特征子集能够使特征属性具有最大关联—最小冗余，即可以最大程度提高属性特征与类属性关联度，且降低属性与属性之间的冗余，即公式(4)中评价值M_s越高，特征属性与类的相关度均值为

越大，属性间的相关度均值则用

越小，CFS takes the correlation of features as the evaluation standard, and is a filter-type feature selection algorithm. Under the corresponding search strategy, it aims to reduce the redundancy between attributes and improve the association between attribute features and class attributes. To achieve the screening of attributes with high redundancy and attributes that are not related to categories, formula (4) is its evaluation standard, and the evaluation of k features of feature subset S is represented by M _s , where the correlation between feature attributes and classes mean is

The mean correlation between attributes is used

It can be seen from formula (4) that the candidate feature subset determined by the association rule feature selection algorithm can make the feature attributes have the maximum correlation-minimum redundancy, that is, it can maximize the degree of association between attribute features and class attributes, and reduce the relationship between attributes and attributes. The redundancy between attributes, that is, the higher the evaluation value M _s in formula (4), the mean of the correlation between the feature attribute and the class is

The larger the value, the mean of the correlation between attributes is used

the smaller the

In the feature selection of association rules, the information gain algorithm is used to evaluate the correlation between attributes, and the calculation method of information gain is the measurement method of symmetry, so when two high-order related features in the feature subset S exist, such as feature W _i , W _j , the symmetric uncertainty method of formula (5) can be used, the feature correlation is U, and the feature entropy is H(W), so formula (6) is the evaluation of feature subset based on the correlation between attributes function, when the evaluation value H _s increases _, the correlation between the feature W _j and Wi in the feature subset S decreases, and the correlation with the class attribute increases,

Using the CFS algorithm, the secondary feature selection function is embedded in the PCA, and then based on the heuristic sequence backward search strategy, the evaluation results of the CFS are calculated, and the optimal feature subsets are filtered out after sorting.

Referring to FIG. 1, the functional framework of a fault feature extraction method for an operation and maintenance management and control platform of the present invention

The data redundancy is more effectively removed after sample space transformation based on PCA feature extraction. The feature extraction process: 1) Feature extraction of the preprocessed dataset S ₀ based on PCA. According to the PCA principle, the covariance matrix S of the high-dimensional sample space X is obtained; the orthogonal matrix Q of S and its eigenvalues λ ₁ ≥λ ₂ ≥…≥λ _n are derived; the cumulative contribution is set according to the actual requirements of the control fault feature extraction the threshold value of the rate t, so as to obtain its standard orthogonal vector u _i , and the low-dimensional sample space U={u ₁ , u ₂ , ... u _d } after feature extraction, and obtain the principal component of the original sample _xi after space transformation Feature y=(u ₁ , u ₂ ,... _ud ) ^T x _i , forming a new candidate feature subset F ₁ . 2) Adaptive quadratic feature selection based on PCA. ① If the key features _of the feature subset F1 need to be locked after the PCA feature extraction of the management and control faults, enter the secondary feature selection function module. The secondary feature selection uses the association rule feature selection CFS algorithm to calculate the feature correlation of the extracted feature set, so that the feature attributes have the maximum correlation-minimum redundancy, that is, it can maximize the correlation between attribute features and class attributes, and reduce the relationship between attributes and attributes. Redundancy between attributes, and at the same time can lock the key feature subset F ₂ after PCA feature extraction. While improving the fault classification accuracy, the function module can enhance the maximum correlation-minimum redundancy of its features, and lock the key features based on PCA feature extraction. ② When only fault classification is required, and key features need not be analyzed, this function module can be skipped to quickly classify management and control faults. 3) On the basis of the feature extraction of the above-mentioned embedded adaptive secondary feature selection function, the formed optimal feature data set of management and control faults is trained, and the fault classification result of the management and control platform is obtained on the test set.

2. The algorithm framework of an operation and maintenance management and control platform fault feature extraction method of the present invention

The algorithm performs feature extraction on the original data set based on principal component analysis to form a feature set F ₁ , and measures the correlation U (W _j ,S) between the feature W _j in F ₁ and the class attribute S, sorts U in descending order, and calculates the CFS The feature entropy evaluation value H _s1 of . The search strategy used in the calculation is the heuristic sequence backward search. The backward search strategy flow is shown in Figure 2. Each time, the feature with a small correlation evaluation value with the class attribute is deleted, and the feature after this feature is deleted is calculated again. The entropy evaluation value H _s2 . Loop evaluation H _s when it is not less than the threshold value, if H _s2 ≥ H _s1 , the feature subset F ₁ will be updated, if H _s2 <H _s1 , the feature subset F ₁ will not be updated, and jump out of the loop output when H _s is less than the threshold value The optimal feature subset F ₂ . The secondary feature selection function module can further lock the key features of the optimal feature subset through association rule feature selection based on PCA feature extraction. The pseudocode of its quadratic feature selection algorithm is as follows:

Input: feature set F ₁ after PCA feature extraction, output: optimal feature set F ₂ ,

1. Select all features after PCA feature extraction to form feature subset F ₁ ,

2. Calculate the correlation U(W _j ,S) between each feature attribute W _j and the class attribute S in F ₁ ,

3. Calculate the feature entropy evaluation value H _s ,

4. Arrange in descending order the U(W _j ,S) value of the correlation between each feature and class attribute, H _s1 ←H _s ,

5. For H _s1 ≥ δdo,

6. Delete a feature in F ₁ to form a new feature subset F ₂ , calculate the feature entropy evaluation value H _s2 ,

7. If H _s2 ≥ H _s1 , then F ₁ =F ₂ ,

_8.else , F1 unchanged,

9.End if,

10. H _s1 =H _s2 ,

11. End For.

The inventor adopts a fault feature extraction method for an operation and maintenance management and control platform of the present invention, and conducts a comparative analysis on the fault identification performance of the management and control platform after the feature extraction. First, the feature extraction is carried out through PCA, and the cumulative contribution rate of the principal components is determined to be 94%. This is because when the threshold t (threshold=94%), the feature dimension is reduced to 18 dimensions, and the average fault identification accuracy rate reaches more than 98%. As shown in Figure 3. It should be noted that the threshold t determines the cumulative contribution rate of the PCA principal components. Although the cumulative contribution rate is the largest when the threshold=100%, it has a high recognition accuracy, but at the same time, the dimension of the feature also increases sharply. Therefore, the higher the threshold t is, the better. Only when the dimension and classification accuracy are balanced can the performance of the classifier be optimized. After PCA feature extraction, the secondary selection of 18-dimensional features is performed again. The results show that the 1, 2, 5, 6, 7, and 12-dimensional features have the smallest redundancy and the strongest correlation with class attributes. After screening, they are the key feature subsets after feature extraction. Table 1 shows the cross-validation results after secondary feature selection. The comparison effect of the 6-dimensional key feature subset based on secondary feature selection is shown in Figure 4. The classification accuracy differs by less than 1.1%. Since the feature dimension is reduced to 6 dimensions, the feature dimension is reduced by 65% compared with the simple PCA dimension reduction; the average execution time of the classifier model is reduced by 31.3%. During the fault classification process of the management and control platform, adaptive feature extraction and selection can be performed according to specific needs. When only fault classification is required, the classification accuracy is high, and key features do not need to be analyzed, the secondary feature selection module can be skipped to classify management and control faults. When the key features need to be locked and the feature dimension is required to be high, the secondary feature selection module can be adaptively entered to further lock the key features, and at the same time, the fault classification results of the management and control platform can be obtained on the test set. The above proves the feasibility and effectiveness of a fault feature extraction method for an operation and maintenance management and control platform proposed by the present invention.

Table 1 PCA-based secondary fault feature selection (ten-fold cross-validation)

Feature dimension after PCA extraction Cross-validation(%) Feature dimension after PCA extraction Cross-validation(%) 1 9 (90%) 11 1 (10%) 2 10 (100%) 12 10 (100%) 3 5 (50%) 13 0 (0%) 4 4 (40%) 14 0 (0%) 5 10 (100%) 15 0 (0%) 6 10 (100%) 16 0 (0%) 7 9 (90%) 17 0 (0%) 8 7 (70%) 18 0 (0%) 9 1 (10%) 10 0 (0%)

To sum up, the method for extracting fault features of an operation and maintenance management and control platform of the present invention reduces the feature dimension of each fault sample space, shortens the training time, and improves the classification accuracy of the learning classifier. Because it performs PCA feature extraction first, it greatly reduces the feature dimension of management and control fault classification and reduces the computational complexity. At the same time, because it performs adaptive secondary feature selection after feature extraction, it overcomes the problem that a single feature extraction method cannot lock key features, and reduces the redundancy between features, enhances the correlation between features and class attributes, and greatly improves faults. Classification accuracy.

The software program of the present invention is compiled according to automation and computer processing technology, and is a technology familiar to those skilled in the art.

The embodiments of the present invention are not exhaustive, and those skilled in the art can simply copy and improve without creative work, and still fall within the scope of the protection of the present invention.

Claims (1)

1. A fault feature extraction method for an operation and maintenance management and control platform is characterized by comprising the following contents:

1) principal component analysis feature extraction

Principal Component Analysis (PCA) is to perform sample space transformation, determine the projection direction with the largest variance of all original feature vectors by projection, perform feature extraction on the projection direction positioning discrimination vectors, and change the original samples into low-dimensional samples dispersed as much as possible after projection transformation while maintaining the original samplesBefore transformation, the difference of original high-dimensional sample space is set to contain N samples, X belongs to RⁿEach sample is an X_i＝[x_i1,...,x_in]^T∈RⁿIf the mean vector is M, the corresponding feature vector is X_i＝[x_1i,...,x_ni]∈RⁿAnd the corresponding covariance matrix is formula (1),

the distribution variance of the sample on the feature vector, namely the feature value of the covariance matrix of formula (1), the orthogonal matrix obtained by diagonalizing the covariance matrix of formula (1) is formula (2),

denotes Q as

Where M is the dimension of the orthogonal matrix Q, and PCA derives the eigenvalues λ in the matrix based on Q₁≥λ₂≥…≥λ_nAnd calculating the orthonormal eigenvector v corresponding to the eigenvalue₁≥v₂≥…≥v_nObtaining the orthonormal eigenvector u of the covariance matrix S through the eigenvalue of the orthonormal matrix Q and the corresponding orthonormal eigenvector₁,u₂,…u_dAs in equation (3) where the orthonormal eigenvector u₁,u₂,…u_dCorresponding to the first d largest non-zero eigenvalues of S,

setting t to 95%, u_iT, the cumulative contribution of the principal components of the spatial samples on the first d axes is 95% of the original data, thus, for any sample x_iMapping the space to a reduced-dimension low-dimension sample space U-U₁,u₂,…u_dX of_iIs characterized by y ═ u (u)₁,u₂,…u_d)^Tx_iThen y is_iFor sample points in a low-dimensional space, through the spatial sample transformation of PCA, the transformed samples can represent 95% of the accumulated contribution rate of the principal components, and the original spatial dimension is reduced from n to d, wherein d is smaller than n, so that the spatial dimension is greatly reduced, and the function of feature extraction is played;

2) quadratic feature selection

After PCA (principal component analysis) feature extraction, embedding a quadratic feature selection algorithm for further obtaining an optimal feature subset and key features of a PCA low-dimensional space, wherein the algorithm is based on Filter-based (Filter) association rule feature selection (CFS), adopts a heuristic sequence back search strategy when carrying out Correlation evaluation on sample features, determines the optimal feature subset through the Correlation sorting of the features,

CFS uses the correlation of the characteristics as an evaluation standard, is a Filter type characteristic selection algorithm, aims to reduce the redundancy between attributes and improve the correlation degree of the attribute characteristics and the class attributes under the corresponding search strategy, achieves the screening of the attributes with high redundancy and the attributes irrelevant to the classes, uses a formula (4) as the evaluation standard, and uses M for evaluating the k characteristics of the characteristic subset S_sRepresentation in which the mean of the correlation of the feature attributes with the classes is

The mean of the correlation between attributes is used

As shown in formula (4), the candidate feature subset determined by the association rule feature selection algorithm can make the feature attribute have the maximum association-minimum redundancy, that is, the association degree between the attribute feature and the class attribute can be improved to the maximum extent, and the redundancy between the attribute and the attribute can be reduced, that is, the evaluation value M in formula (4)_sHigher, characteristic property and classMean value of correlation

The larger the correlation mean between the attributes is

The smaller the size of the tube is,

the correlation between the attributes is evaluated by using an information gain algorithm in the association rule feature selection, and the information gain calculation method is a symmetry measurement method, so when two high-order associated features exist in the feature subset S, such as the feature W_i、W_jThe method of symmetry uncertainty of formula (5) can be used, the entropy of the features is H (W), the feature association is U, and thus formula (6) is an evaluation function of a subset of features based on the correlation between attributes, when evaluating the value H_sWhen raised, the features W in the feature subset S_jAnd W_iThe correlation decreases, and the correlation with the class attribute increases,

and embedding a secondary feature selection function in the PCA by adopting a CFS algorithm, then calculating an evaluation result of the CFS based on a heuristic sequence backward search strategy, and screening out an optimal feature subset after sorting.

Images