Feature extraction method based on neighborhood division

Abstract

The invention provides a feature extraction method based on neighborhood division, which relates to the technical field of data classification, and comprises the following steps: dividing a neighborhood, and judging whether the neighborhood is a consistent neighborhood or an inconsistent neighborhood; constructing a neighborhood division classification model; classifying consistent neighborhood objects to a decision positive domain through a neighborhood division classification model, and classifying inconsistent neighborhood objects to a decision boundary; constructing an unbalanced binary tree algorithm, and calculating consistent neighborhood objects and inconsistent neighborhood objects in the discourse domain; constructing a neighborhood positive domain certainty degree algorithm, and determining the sum of contribution degrees of all objects in the discourse domain to the decision positive domain; the contribution degree of a single object to the decision positive domain is [0, 1]; and constructing a feature extraction algorithm, and determining extraction attributes. Neighborhood particles are calculated by adopting an unbalanced binary tree algorithm, so that the calculation efficiency of a feature extraction algorithm is improved; and the attributes are evaluated by adopting a neighborhood and positive domain certainty degree algorithm, so that the extracted attributes have better classification quality in industrial data application.

Description

Technical Field

[0001] The present invention relates to the field of data classification technology, and in particular to a feature extraction method based on neighborhood division. The present application claims priority, and the application number of the prior application is: 2024108350454, and the name is: A feature extraction method based on neighborhood division, and the priority date is 2024-06-26. Background Art

[0002] Industrial data generally contains multiple attributes and is usually recorded as numerical data. It is impossible to directly use the Pawlak model to classify industrial data. Indirect and direct methods are usually used. The indirect method is to first use a discretization algorithm to convert industrial data into symbolic data, and then use the Pawlak model to process it. Although it can solve the application problem of the Pawlak model on numerical data to a certain extent, it ignores the degree of membership information from the original numerical value to the discrete value, resulting in information loss and adverse effects on the classification results; for example, industrial data is data for predictive maintenance. If there is an inaccurate classification result, the status of the equipment may be misjudged, resulting in unnecessary maintenance or ignoring potential failures, thereby affecting production efficiency and safety; for example, industrial data is a parameter for quality presentation. The reduction in its classification accuracy may lead to misjudgment of product quality, thereby affecting the product's pass rate and consistency.

[0003] The direct method directly processes the original data by extending the Pawlak model. Considering that the typical fuzzy rough set model establishes multiple fuzzy concepts covering the entire value range on numerical attributes to construct fuzzy similarity relations on the domain, it is worth exploring how to construct effective fuzzy similarity relations to improve the classification quality according to the classification needs of industrial data. Summary of the invention

[0004] The purpose of the present invention is to provide a feature extraction method based on neighborhood division, and the technical problem to be solved is how to improve the classification quality.

[0005] The present invention is achieved through the following technical solutions:

[0006] A feature extraction method based on neighborhood division is used to analyze industrial data and extract key attributes. The feature extraction method includes the following steps:

[0007] S100, dividing the neighborhood, and determining whether the neighborhood is a consistent neighborhood or an inconsistent neighborhood;

[0008] S200, constructing a neighborhood partition classification model; classifying objects through the neighborhood partition classification model to obtain consistent neighborhood objects and inconsistent neighborhood objects; classifying consistent neighborhood objects into the decision positive domain, and classifying inconsistent neighborhood objects into the decision boundary;

[0009] S300, constructing an unbalanced binary tree algorithm; calculating consistent neighbor objects and inconsistent neighbor objects in the domain through the unbalanced binary tree algorithm;

[0010] The domain can be an industrial data set called from a database, and the industrial data in the industrial data set is taken as an object;

[0011] S400, constructing a neighborhood positive domain certainty algorithm; determining the sum of contributions of all objects in the domain to the decision positive domain through the neighborhood positive domain certainty algorithm; the contribution of a single object to the decision positive domain is [0, 1];

[0012] S500: construct a feature extraction algorithm, determine the extraction attributes through the feature extraction algorithm and use them for classification.

[0013] The classification is performed on the domain through the extracted attributes.

[0014] Through the above neighborhood division, consistent neighborhood and inconsistent neighborhood are defined. The neighborhood division classification model classifies objects into decision positive domain and decision boundary. Consistent neighborhood objects are classified into decision positive domain, and inconsistent neighborhood objects are classified into decision boundary. The decision positive domain contains objects that are highly related to a certain category, and the decision boundary contains objects with relatively fuzzy classification. The unbalanced binary tree can quickly locate objects that are closely related to a specific category, improve classification efficiency, and gradually refine classification decisions through the hierarchical structure of the tree, thereby improving classification accuracy and classification quality. Through the neighborhood positive domain certainty algorithm, the contribution of each object in the domain to the decision positive domain is quantified. The value range of this contribution is [0,1], which provides a quantitative indicator of the degree of association with a specific category for the object. Compared with the contribution of the object to the decision positive domain {0,1} in the prior art, it can better reflect the classification ability of the attribute. The quantitative indicator helps to identify the attributes that play a key role in the classification decision, thereby improving the classification quality.

[0015] Furthermore, the neighborhood is divided to determine whether it is a consistent neighborhood or an inconsistent neighborhood. The specific steps are as follows:

[0016] S110, obtain neighborhood decision space NDS = (U, N, D), neighborhood relationship N, condition attribute set C, decision attribute set D, decision value range V D , attribute set P and object u;

[0017] S120, setting preconditions: the above attribute set P is included in the condition attribute set C, and the object u belongs to the domain U;

[0018] S130, based on the decision value d i For the neighborhood δ P (u) is divided to obtain the neighborhood sub-area, the formula is as follows:

[0019]

[0020] Among them, δ P (u) represents the neighborhood of object u under the attribute set P; Represents the neighborhood δ P (u) The internal decision value is d i The object collection, d i neighborhood subregion; v represents the object v in the domain; f(v,d) represents the decision value of object v; d i represents the decision value;

[0021] S140, through the neighborhood δ P (u) and the neighborhood subregion Determine the sub-area probability, the formula is as follows:

[0022]

[0023] in, Represents a neighborhood subarea The probability of a sub-area; Represents a neighborhood subarea The number of objects in |δ P (u)| represents the neighborhood δ P (u) The number of objects in it;

[0024] S150, through the neighborhood δ P (u) Each neighborhood sub-area The decision value set of the neighborhood δ P (u) Whether it is a consistent neighborhood;

[0025] S151, if the neighborhood δ P Any two neighboring sub-regions in (u) The decision value sets of are equal, then the neighborhood δ P (u) is the consistent neighborhood;

[0026] S152, if the neighborhood δ P (u) There are two unequal neighborhood sub-regions The decision value set of the neighborhood δ P (u) is the inconsistent neighborhood;

[0027] Furthermore, the above steps of constructing the neighborhood partition classification model include:

[0028] S210, by |λ(δ P (u))| is equal to 1, which determines whether the object u belongs to the decision positive domain or the decision boundary; where |λ(δ P (u))| represents the neighborhood δ P(u) the number of decision values ​​in the decision value set of the contained object;

[0029] S211, if the above |λ(δ P (u))|=1, then the object u belongs to the decision positive domain;

[0030] S212, if the above |λ(δ P (u))|>1, then the object u belongs to the decision boundary.

[0031] Furthermore, the steps of constructing an unbalanced binary tree algorithm include:

[0032] S310, taking the above domain as the root node

[0033] S320, through Calculate the left child node of the unbalanced binary tree algorithm;

[0034] pass Calculate the right child node of the unbalanced binary tree algorithm;

[0035] in, represents the kth left child node; ψ con represents the consistent neighborhood dimension-increasing operator; represents the k-1th left child node; a k represents the kth attribute, a k Belongs to the attribute set P; represents the kth right child node; ψ inc represents the inconsistent neighborhood dimension increasing operator; Represents the k-1th right child node;

[0036] S330, the kth right child node As the parent node, execute S320 to obtain the k+1th left child node and the k+1th right child

[0037] The left child node of this unbalanced binary tree algorithm is a leaf node, and the right child node keeps growing.

[0038] Furthermore, the formula of the above neighborhood positive domain certainty algorithm is as follows:

[0039]

[0040] Among them, NPRC P represents the certainty of the neighborhood positive domain under the attribute set P; U represents the domain; σ NPRC (u) represents the contribution of object u to the decision positive domain; represents the focal sub-region of object u; represents the probability of the focal sub-region of object u; d F represents the decision value of the focal sub-area; NPU(u) represents the uncertainty of the neighborhood partition of object u under the attribute set P.

[0041] Furthermore, a feature extraction algorithm is constructed, and the extraction attributes are determined by the feature extraction algorithm. The specific steps include:

[0042] S510, setting an additional condition: attribute b belongs to the relative complement of attribute set P in condition attribute set C;

[0043] S520, calculating the importance of attribute b relative to the decision attribute set D, the formula is as follows:

[0044] Sig(b,P,D)=NPRC P∪ { b} -NPRC P

[0045] Among them, Sig(b,P,D) represents the importance of attribute b relative to the decision attribute set D; NPRC P∪ { b} represents the neighborhood positive domain certainty under the attribute set P∪{b}; NPRC P It represents the certainty of the neighborhood positive domain under the attribute set P;

[0046] S530: Select the attribute b with the greatest importance as the extracted attribute.

[0047] Furthermore, the calculation formula for the uncertainty of neighborhood partition of object u under the above attribute set P is as follows:

[0048]

[0049] Among them, NPU(u) represents the uncertainty of neighborhood partitioning of object u under the attribute set P.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] The present invention defines consistent neighborhoods and inconsistent neighborhoods through the above-mentioned neighborhood division, and can better analyze the types of industrial data. The objects are classified into decision positive domains and decision boundaries through the neighborhood division classification model. Consistent neighborhood objects are classified into decision positive domains, and inconsistent neighborhood objects are classified into decision boundaries. This method can obtain better rationality and accuracy when extracting key attributes of industrial data; at the same time, it can effectively isolate or reduce the interference of data noise on the feature extraction process, ensure the reliability of data analysis results, and avoid the influence of noise in industrial data on analysis conclusions.

[0052] The decision positive domain contains objects that are highly related to a certain category, and the decision boundary contains objects with relatively vague classifications; the decision boundary is conducive to handling complex or uncertain classification situations, ensuring that the classification model can accurately handle both deterministic objects and fuzzy or boundary situations, thereby improving the overall classification performance. This is especially obvious when the amount of industrial data is large and messy.

[0053] The present invention can also quickly locate objects closely related to a specific category through an unbalanced binary tree, thereby improving classification efficiency, and gradually refine classification decisions through the hierarchical structure of the tree, thereby improving classification accuracy and classification quality; through the neighborhood positive domain certainty algorithm, the contribution of each object in the domain to the decision positive domain is quantified, and the value range of the contribution is [0,1], which provides a quantitative indicator of the degree of association with a specific category for the object. Compared with the contribution of the object to the decision positive domain {0,1} in the prior art, it can better reflect the classification ability of the attribute; the quantitative indicator helps to identify the attributes that play a key role in the classification decision, thereby improving the classification quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0055] Figure 1 As the main flow chart;

[0056] Figure 2 is the neighborhood δ in Example 1 P (x) Schematic diagram of particle (object) distribution;

[0057] Figure 3 is the neighborhood δ in Example 1 P (y) Schematic diagram of particle (object) distribution;

[0058] Figure 4 Schematic diagram of object distribution in the two-dimensional neighborhood decision space in Example 3. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0060] Embodiment 1:

[0061] Combination Figure 1This embodiment 1 provides a feature extraction method based on neighborhood division, which is used to analyze industrial data and extract key attributes; the feature extraction method includes the following steps:

[0062] S100, calling an industrial data set from a database, taking the industrial data set as a neighborhood, and the industrial data in the industrial data set as an object; dividing the neighborhood, and determining whether the neighborhood is a consistent neighborhood or an inconsistent neighborhood;

[0063] S200, constructing a neighborhood partition classification model; classifying objects through the neighborhood partition classification model to obtain consistent neighborhood objects and inconsistent neighborhood objects; classifying consistent neighborhood objects into the decision positive domain, and classifying inconsistent neighborhood objects into the decision boundary;

[0064] S300, constructing an unbalanced binary tree algorithm; calculating consistent neighbor objects and inconsistent neighbor objects in the domain through the unbalanced binary tree algorithm;

[0065] The domain can be an industrial dataset called from a database;

[0066] S400, constructing a neighborhood positive domain certainty algorithm; determining the sum of contributions of all objects in the domain to the decision positive domain through the neighborhood positive domain certainty algorithm; the contribution of a single object to the decision positive domain is [0, 1];

[0067] S500: construct a feature extraction algorithm, and determine the extraction attributes and use them for classification through the feature extraction algorithm.

[0068] Through the above neighborhood division, consistent neighborhood and inconsistent neighborhood are defined. The neighborhood division classification model classifies objects into decision positive domain and decision boundary. Consistent neighborhood objects are classified into decision positive domain, and inconsistent neighborhood objects are classified into decision boundary. The decision positive domain contains objects that are highly related to a certain category, and the decision boundary contains objects with relatively fuzzy classification. The unbalanced binary tree can quickly locate objects that are closely related to a specific category, improve classification efficiency, and gradually refine classification decisions through the hierarchical structure of the tree, thereby improving classification accuracy and classification quality. Through the neighborhood positive domain certainty algorithm, the contribution of each object in the domain to the decision positive domain is quantified. The value range of this contribution is [0,1], which provides a quantitative indicator of the degree of association with a specific category for the object. Compared with the contribution of the object to the decision positive domain {0,1} in the prior art, it can better reflect the classification ability of the attribute. The quantitative indicator helps to identify the attributes that play a key role in the classification decision, thereby improving the classification quality.

[0069] A reference application scenario is that this method is used for industrial data processing in factories producing steel wire. The industrial data is the recorded data of the factory monitoring the production process during the production of steel wire, and a data set is obtained; the data set records the temperature, tensile force, material composition (attributes) and steel wire quality (decision value) and other information in the production process. By performing neighborhood division using the above method, it is possible to identify whether the quality of steel wire produced under a specific temperature and tensile force is consistent. If it is consistent, the corresponding data belongs to the consistent neighborhood; and when the quality of the steel wire shows a large difference under certain combinations of temperature and tensile force, the corresponding data belongs to the inconsistent neighborhood.

[0070] Since the positive domain of the decision includes objects that can be clearly classified into a certain category, while the decision boundary includes objects with fuzzy or complex classification; therefore, the positive domain of the decision may include high-quality steel wire data produced under specific temperature and tensile force conditions; while the decision boundary may include steel wire data produced under different temperatures and tensile forces;

[0071] Subsequently, the unbalanced binary tree algorithm is used to quickly locate production conditions that have a significant impact on the quality of steel wire, such as temperature, tensile force and other attributes, so as to quickly screen out key data points from a large amount of industrial data for further analysis. The neighborhood positive domain certainty algorithm can quantify the impact of each data point (such as temperature and tensile force under a single production condition) on steel wire production. For example, different combinations of temperature and tensile force may quantify the contribution of high-quality steel wire production, and the importance of different production conditions to the production of high-quality steel wire can be better identified through quantitative methods. Finally, the most influential attributes and classifications are determined through the feature extraction algorithm. Due to its reliable classification quality, it objectively provides an important reference for the subsequent adjustment of production parameters.

[0072] It is worth noting that the data set may also be industrial data from production of other types of factories. The industrial data contains influencing factors (ie, attributes). Important influencing factors are extracted through feature extraction algorithms, and the influencing factors are then used to classify the industrial data.

[0073] Embodiment 2:

[0074] On the basis of Example 1, the neighborhood is divided to determine whether the neighborhood is a consistent neighborhood or an inconsistent neighborhood. The specific steps are as follows:

[0075] S110, obtain neighborhood decision space NDS = (U, N, D), neighborhood relationship N, condition attribute set C, decision attribute set D, decision value range V D , attribute set P and object u;

[0076] S120, set precondition: NDS = (U, N, D), Neighborhood relationship N,D={d},V D={d1,d2,…,d j}, u∈U (U is the domain);

[0077] S130, based on the decision value d i For the neighborhood δ P (u) is divided to obtain the neighborhood sub-area, the formula is as follows:

[0078]

[0079] Among them, δ P (u) represents the neighborhood of object u under the attribute set P; Represents the neighborhood δ P (u) The internal decision value is d i The object collection, d i neighborhood subregion; v represents the object v in the domain; f(v,d) represents the decision value of object v; d represents the decision attribute; d i represents the decision value;

[0080] S140, through the neighborhood δ P (u) and the neighborhood subregion Determine the sub-area probability, the formula is as follows:

[0081]

[0082] in, Represents a neighborhood subarea The probability of a sub-area; Represents a neighborhood subarea The number of objects in |δ P (u)| represents the neighborhood δ P (u) The number of objects in it;

[0083] For example: If a neighborhood subarea If f(u,d)=d F (d F ∈V D ),but is the probability of the focal sub-area, f(u,d) represents the decision value of object u; d F represents the decision value of the focal sub-area;

[0084] S150, given condition: NDS=(U,N,D), D={d},u∈U,

[0085] V D ={d1,d2,…,d j}, neighborhood division

[0086] Through the neighborhood δ P (u) Each neighborhood sub-area The decision value set of the neighborhood δ P (u) Whether it is a consistent neighborhood;

[0087] S151, if the neighborhood δ P Any two neighboring sub-regions in (u) The decision value sets of are equal, then the neighborhood δ P (u) is the consistent neighborhood;

[0088] S152, if the neighborhood δ P (u) There are two unequal neighborhood sub-regions The decision value set of the neighborhood δ P (u) is an inconsistent neighborhood.

[0089] For example: From the neighborhood δ P (u) Take any two neighborhood subregions when i=m and i=n and like Then the neighborhood δ P (u) is a consistent neighborhood; if Then the neighborhood δ P (u) is an inconsistent neighborhood; where Represents a neighborhood subarea The set of decision values ​​for the contained objects; Represents a neighborhood subarea A collection of decision values ​​for the contained objects.

[0090] The neighborhood is determined to be a consistent neighborhood. The application scenario for reference is that the factory monitors and records the corresponding data and decision values ​​of each steel wire regarding temperature, tensile force, and material composition (attributes) during the production of steel wire. The corresponding data of temperature, tensile force, material composition, and decision attributes constitute an industrial data (i.e., an object), and the recorded steel wire data constitute an industrial data set. The industrial data set is used as the neighborhood δ P (x), the objects contained in the industrial dataset are {x1,x2,…,x 11}, in a given neighborhood decision space, for the attribute set P (the attribute set P includes temperature, tensile force and material composition), consider the neighborhood δ of object x P (x), such as Figure 2 As shown, the decision range V D ={d1,d2,d3,d4,d5}, Figure 2 ●, ○, ▲, Δ, □ represent decision values ​​d1, d2, d3, d4, d5 respectively;

[0091] For the neighborhood δ P(x), the objects contained in each neighborhood sub-area are: Since f(x,d)=d2, we know is the focal sub-region, d2 is the decision value of the focal sub-region, x1, x2,…, x 11 is the supporting object of x;

[0092] The neighborhood is divided into The probability distribution of neighborhood partition NP(x) is NPP(x) = {(0,d1), (1,d2), (0,d3), (0,d4), (0,d5)}; due to λ(δ P (x)) = {d2}, |λ(δ P (x))|=1, so δ P (x) is a consistent neighborhood. P (x)) represents the neighborhood δ P (x) is the set of decision values ​​of the objects contained; |λ(δ P (x))| represents the neighborhood δ P (x) The number of decision values ​​in the decision value set of the object contained.

[0093] The neighborhood is determined to be an inconsistent neighborhood; the application scenario for reference is that the factory monitors and records the corresponding data and decision values ​​of each nut regarding the outer diameter, inner diameter, pitch, lead, production temperature and material composition during the production of nuts. The corresponding data and decision values ​​of the outer diameter, inner diameter, pitch, lead, production temperature and material composition constitute an industrial data (i.e., an object), and the recorded data of several nuts constitute an industrial data set. The industrial data set is used as the neighborhood δ P (y), the objects contained in the industrial dataset are {y1,y2,…,y 15}, in a given neighborhood decision space, for the attribute set P (the attribute set P includes outer diameter, inner diameter, pitch, lead, production temperature and material composition), consider the neighborhood δ of object y P (y), such as Figure 3 As shown; decision range V D ={d1,d2,d3,d4,d5}, where d1,d2,d3,d4,d5 are represented by ●,○,▲,Δ,□ respectively;

[0094] For the neighborhood δ P (y), the objects contained in each neighborhood sub-area are:

[0095] Since f(y,d)=d2, we know is the focus sub-area, d2 is the decision value of the focus sub-area, y1, y2, y3, y4 are the support objects of y;

[0096] The neighborhood is divided into:

[0097] NP(y)={({y5,y6,y7},d1),({y1,y2,y3,y4},d2),({y 11 ,y 12 ,y 13 ,y 14},d3),({y8},d4),

[0098] ({y9,y 10},d5)}; The probability distribution of neighborhood partition NP(y) is:

[0099] NPP(y)={(3 / 14,d1),(4 / 14,d2),(4 / 14,d3),(1 / 14,d4),(2 / 14,d5)};

[0100] λ(δ P (y))={d1,d2,d3,d4,d5}, |λ(δ P (y))|=5, so δ P (y) is an inconsistent neighborhood. P (y)) represents the neighborhood δ P (y) The set of decision values ​​of the objects contained; |λ(δ P (y))| represents λ(δ P The number of decision values ​​in (y).

[0101] In a specific embodiment, the step of constructing a neighborhood partitioning classification model includes:

[0102] S210, given condition: NDS=(U,N,D), Neighborhood relationship N,D={d},

[0103] V D ={d1,d2,…,d j},u∈U;

[0104] By |λ(δ P (u))| is equal to 1, which determines whether the object u belongs to the decision positive domain or the decision boundary; where |λ(δ P (u))| represents the neighborhood δ P (u) the number of decision values ​​in the decision value set of the contained object;

[0105] S211, if {u∈U||λ(δ P (u))|=1}, then the object u belongs to the decision positive domain;

[0106] S212, if {u∈U||λ(δP (u))|>1} and Then the object u belongs to the decision boundary.

[0107] The above neighborhood partition classification model classifies consistent neighborhood objects as decision-positive domain objects and inconsistent neighborhood objects as decision-boundary objects, so that the classification of an object is independent of the decision class and is only related to the properties of the neighborhood itself. The model is more intuitive and concise, and has higher computational efficiency.

[0108] Embodiment 3:

[0109] Based on Example 2, the steps of constructing an unbalanced binary tree algorithm include:

[0110] S310, taking the above domain as the root node

[0111] S320, through Calculate the left child node of the unbalanced binary tree algorithm;

[0112] pass Calculate the right child node of the unbalanced binary tree algorithm;

[0113] in, represents the kth left child node; ψ con represents the consistent neighborhood dimension-increasing operator; represents the k-1th left child node; a k represents the kth attribute, a k Belongs to the attribute set P; represents the kth right child node; ψ inc represents the inconsistent neighborhood dimension increasing operator; Represents the k-1th right child node;

[0114] S330, the kth right child node As the parent node, execute S320 to obtain the k+1th left child node and the k+1th right child

[0115] The left child node of this unbalanced binary tree algorithm is a leaf node, and the right child node keeps growing.

[0116] The application scenario for reference is that the factory monitors the entire production process when producing glass plates, and records the corresponding data and decision values ​​of each glass plate on production temperature, production time, fuel consumption and material composition. The corresponding data and decision values ​​of production temperature, production time, fuel consumption and material composition constitute an industrial data (i.e., an object), and the recorded data of several glass plates constitute an industrial data set. The industrial data set is taken as the domain U. Given the neighborhood decision space NDS = (U, N, D), And P = {a1, a2, ..., a m}, the consistent neighborhood particle set and the inconsistent neighborhood particle set are regarded as the left child node and the right child node of the binary tree respectively; the specific construction method is as follows:

[0117] (1) Take the domain U as the root node

[0118] (2) Calculate the nodes under a single attribute: P1 = {a1}, then the left child node under P1 and right child They are The new neighborhood particle set calculated by the consistent neighborhood particle set under the dimension-increasing operator ψ is also a consistent neighborhood particle set. Only the inconsistent neighborhood particle set may generate a new inconsistent neighborhood particle set under the dimension-increasing operator ψ. Therefore, The neighborhood particle sets obtained under the dimension-increasing operator ψ are all consistent neighborhood particle sets. As a leaf node, it is called a consistent leaf node and no longer grows; Under the dimensionality increasing operator ψ, consistent neighborhood particle sets and inconsistent neighborhood particle sets can be generated. As a parent node, it is called an inconsistent node and participates in the subsequent dimension increase operation;

[0119] (3) Calculate the nodes under the two attributes: P2 = {a1, a2}, the left child node under P2 and right child They are Similarly, As a consistent leaf node, As a parent node, continue to grow;

[0120] (4) Calculate the nodes under multiple attributes. k ={a1,a2,…,a k}, in P k The left child node below and right child They are

[0121] Embodiment 4:

[0122] Based on Example 3, given the neighborhood decision space NDS=(U,N,D), Neighborhood relationship N,D={d},V D ={d1,d2,…,d j}, u∈U, neighborhood partition NP(u), probability distribution NPP(u), determine the uncertainty of neighborhood partition of object u under attribute set P, the calculation formula is as follows:

[0123]

[0124] Among them, NPU(u) represents the uncertainty of neighborhood partitioning of object u under the attribute set P.

[0125] NPU(u) reflects the detailed information of neighborhood partitioning. If the objects in a neighborhood are more concentrated, the uncertainty of neighborhood partitioning will be smaller and the ability to distinguish attributes will be better. P When (u) is a consistent neighborhood, objects are concentrated in the focal sub-area, and NPU(u) = 0, and the attribute discrimination ability is the best; when δ P The objects (u) are evenly distributed in |V D | neighborhood sub-areas, that is, NPU(u)=1, the attribute has the worst ability to distinguish; therefore, NPU(u)∈[0,1].

[0126] In a specific embodiment, given a neighborhood decision space NDS=(U, N, D), Neighborhood relationship N,D={d},V D ={d1,d2,…,d j}, u∈U, neighborhood partition NP(u), probability distribution NPP(u), determine the above neighborhood positive domain certainty algorithm, the formula is as follows:

[0127]

[0128] Among them, NPRC P represents the certainty of the neighborhood positive domain under the attribute set P; U represents the domain; σ NPRC (u) represents the contribution of object u to the decision positive domain; represents the focal sub-region of object u; represents the probability of the focal sub-region of object u; d F represents the decision value of the focal sub-area; NPU(u) represents the uncertainty of the neighborhood partition of object u under the attribute set P.

[0129] The neighborhood positive domain certainty not only takes into account the support of the focal sub-region for the central object, but also takes into account the structural information of the neighborhood partition. It takes the probability distribution value of the neighborhood partition certainty on the focal sub-region as the contribution of the central object to the decision positive domain σ NPRC (u),σ NPRC The larger (u) is, the stronger the classification ability of the attribute is; NPU(u)∈[0,1], therefore, σ NPRC (u)∈[0,1]; when δ P When (u) is a consistent neighborhood, σ NPRC (u) = 1, indicating that the central object u contributes the most to the decision positive domain (equal to 1); when δP (u) is an inconsistent neighborhood and objects are evenly distributed in |V D | neighborhood sub-regions, σ NPRC (u) = 0, indicating that the contribution of the central object u to the decision positive domain is the smallest (equal to 0).

[0130] The application scenario for reference is that the factory monitors the entire production process when producing steel wire, and records the corresponding data and decision values ​​of each steel wire on temperature, tensile force and material composition. The corresponding data and decision values ​​of temperature, tensile force and material composition constitute an industrial data (i.e., an object), and the recorded steel wire data constitute an industrial data set. Given a neighborhood decision space, the neighborhood decision space is a two-dimensional space. The objects in the space can be divided into three categories. "+" represents an object with a decision value of d1, "*" represents an object with a decision value of d2, and "o" represents an object with a decision value of d3. In order to facilitate comparative analysis, we only examined the neighborhood conditions of five objects x1, x2, x3, x4, and x5, such as Figure 4 shown.

[0131] The neighborhood partition probability distribution calculation results of the five objects are as follows:

[0132]

[0133]

[0134]

[0135]

[0136]

[0137] The contribution of the five objects to the decision positive domain is calculated using the neighborhood positive domain certainty algorithm;

[0138] For x1: |λ(δ P (x1))|=1, we get σ NPRC (x1) = 1.000;

[0139] For x2: |λ(δ P (x2))|>1, we get σ NPRC (x2) = 0.0035;

[0140] For x3: |λ(δ P (x3))|>1, we get σ NPRC (x3) = 0.0565;

[0141] For x4: |λ(δ P (x4))|>1, we get σ NPRC (x4) = 0.1301;

[0142] For x5: |λ(δ P (x5))|>1, we get σ NPRC (x5) = 0.8460;

[0143] Neighborhood Positive Region Certainty NPRC P =2.0361, NPRC uses neighborhood structure information to characterize the contribution of the neighborhood to the decision positive domain, and its value is [0,1], which truly reflects the classification ability of the attribute. Although x2 and x3 have the same focal sub-area probability However, since the neighborhood division of x2 is more uniform than that of x3, the uncertainty of the neighborhood division of x2 is greater than that of x3, so σ NPRC (x2)<σ NPRC (x3); The neighborhood divisions of x4 and x5 are the same, but due to the focal sub-area probability of x4 Probability of focal sub-area less than x5 Therefore, σ NPRC (x4)<σ NPRC (x5).

[0144] Embodiment 5:

[0145] A feature extraction algorithm is constructed based on any of the above embodiments, and the extraction attribute is determined by the feature extraction algorithm. The specific steps include:

[0146] S510, setting an additional condition: attribute b belongs to the relative complement of attribute set P in condition attribute set C (ie, b∈CP);

[0147] S520, calculating the importance of attribute b relative to the decision attribute set D, the formula is as follows:

[0148] Sig(b,P,D)=NPRC P∪{b} -NPRC P

[0149] Among them, Sig(b,P,D) represents the importance of attribute b relative to the decision attribute set D; NPRC P∪{b} represents the neighborhood positive domain certainty under the attribute set P∪{b}; NPRC P It represents the certainty of the neighborhood positive domain under the attribute set P;

[0150] S530: Select the attribute b with the greatest importance as the extraction attribute, and extract the object using the extraction attribute.

[0151] The industrial data of the present invention is not limited to steel wire production data, but can also be case backup within a hospital, defect data detected by a certain product, etc.

[0152] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A feature extraction method based on neighborhood division, which is used to analyze data and extract key attribute indicators; characterized in that: The following steps are involved: S100, dividing the neighborhood, and determining whether the neighborhood is a consistent neighborhood or an inconsistent neighborhood; S200, constructing a neighborhood partition classification model; classifying objects using the neighborhood partition classification model to obtain consistent neighborhood objects and inconsistent neighborhood objects; classifying consistent neighborhood objects into a decision positive domain, and classifying inconsistent neighborhood objects into a decision boundary; S300, constructing an unbalanced binary tree algorithm; calculating consistent neighbor objects and inconsistent neighbor objects in the domain of discourse by using the unbalanced binary tree algorithm; S400, constructing a neighborhood positive domain determination algorithm; determining the sum of contributions of all objects in the domain to the decision positive domain through the neighborhood positive domain determination algorithm; the contribution of a single object to the decision positive domain is [0, 1]; S500: construct a feature extraction algorithm, and determine extraction attributes through the feature extraction algorithm.

2. A feature extraction method based on neighborhood division according to claim 1, characterized in that: Divide the neighborhood and determine whether it is a consistent neighborhood or an inconsistent neighborhood. The specific steps are as follows: S110, obtain neighborhood decision space NDS = (U, N, D), neighborhood relationship N, condition attribute set C, decision attribute set D, decision value range V D , attribute set P and object u; S120, setting preconditions: the attribute set P is included in the condition attribute set C, and the object u belongs to the domain U; S130, based on the decision value d i For the neighborhood δ P (u) is divided to obtain the neighborhood sub-area, the formula is as follows: Among them, δ P (u) represents the neighborhood of object u under the attribute set P; Represents the neighborhood δ P (u) The internal decision value is d i The object collection, d i neighborhood subregion; v represents the object v in the domain; d represents the decision attribute; f(v,d) represents the decision value of object v; d i represents the decision value; S140, through the neighborhood δ P (u) and the neighborhood subregion Determine the sub-area probability, the formula is as follows: in, Represents a neighborhood subarea The probability of a sub-area; Represents a neighborhood subarea The number of objects in |δ P (u)| represents the neighborhood δ P (u) The number of objects in it; S150, through the neighborhood δ P (u) Each neighborhood sub-area The decision value set of the neighborhood δ P (u) Whether it is a consistent neighborhood; S151, if the neighborhood δ P Any two neighboring sub-regions in (u) The decision value sets of are equal, then the neighborhood δ P (u) is the consistent neighborhood; S152, if the neighborhood δ P (u) There are two unequal neighborhood sub-regions The decision value set of the neighborhood δ P (u) is an inconsistent neighborhood.

3. A feature extraction method based on neighborhood division according to claim 2, characterized in that: The step of constructing a neighborhood partition classification model comprises: S210, through Is it equal to 1, determine whether the object u belongs to the decision positive domain or the decision boundary; where |λ(δ P (u))| represents the neighborhood δ P (u) the number of decision values ​​in the decision value set of the contained object; S211, if the |λ(δ P (u))|=1, then the object u belongs to the decision positive domain; S212, if the |λ(δ P (u))|>1, then the object u belongs to the decision boundary.

4. The feature extraction method based on neighborhood division according to claim 1, characterized in that: The steps of constructing an unbalanced binary tree algorithm include: S310: Using the domain as a root node S320, through Calculate the left child node of the unbalanced binary tree algorithm; pass Calculate the right child node of the unbalanced binary tree algorithm; in, represents the kth left child node; ψ con represents the consistent neighborhood dimension-increasing operator; represents the k-1th left child node; a k represents the kth attribute, a k Belongs to the attribute set P; represents the kth right child node; ψ inc represents the inconsistent neighborhood dimension increasing operator; Represents the k-1th right child node; S330, the kth right child node As the parent node, execute S320 to obtain the k+1th left child node and the k+1th right child 5. The feature extraction method based on neighborhood division according to claim 3 is characterized in that: The formula of the neighborhood positive domain certainty algorithm is as follows: Among them, NPRC P represents the certainty of the neighborhood positive domain under the attribute set P; U represents the domain; σ NPRC (u) represents the contribution of object u to the decision positive domain; represents the focal sub-region of object u; represents the probability of the focal sub-region of object u; d F represents the decision value of the focal sub-area; NPU(u) represents the uncertainty of the neighborhood partition of object u under the attribute set P.

6. The feature extraction method based on neighborhood division according to claim 5, characterized in that: Construct a feature extraction algorithm and determine the extraction attributes through the feature extraction algorithm. The specific steps include: S510, setting an additional condition: attribute b belongs to the relative complement of attribute set P in condition attribute set C; S520, calculate the importance of attribute b relative to the decision attribute set D, the formula is as follows: Sig(b,P,D)=NPRC P∪ { b} - NPRC P , Among them, Sig(b,P,D) represents the importance of attribute b relative to the decision attribute set D; NPRC P∪ { b} represents the neighborhood positive domain certainty under the attribute set P∪{b}; NPRC P It represents the certainty of the neighborhood positive domain under the attribute set P; S530: Select the attribute b with the greatest importance as the extracted attribute.

7. The feature extraction method based on neighborhood division according to claim 5, characterized in that: The calculation formula of the neighborhood partition uncertainty of object u under the attribute set P is as follows: Among them, NPU(u) represents the uncertainty of neighborhood partitioning of object u under the attribute set P.

Images