JP3096353B2 - How to classify data - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for classifying data into several categories (classes) when there are cases or calculation results in which the data is given as a set of attribute-value pairs. In particular, the present invention relates to a recursive machine learning method in the case where attributes have a distribution and the distributions overlap, and particularly to a data classification method useful for pattern recognition and accident diagnosis.

[0002]

2. Description of the Related Art In the recursive machine learning method, conventionally, an attribute tree has no distribution, and a discrimination tree is created by discrete attribute values. For example, conventionally, if there is data that classifies the tastes of people in their twenties with respect to the type of car as shown in Table 1 as P if they like it, and classifies as N if they do not like, an identification tree as shown in FIG. To classify the data.

[0003]

[Table 1]

[0004] However, the attributes in Table 1 are discrete values,
For example, whether or not automatic data is present is considered, but there are cases in which normal data may be considered to be a case where attributes have continuous values or a case where discrete values are present but have attribute distributions. In Table 1, the colors are 3 for brightness, saturation, and hue.
There are two attributes, for example, generally speaking blue, lightness,
Blue is recognized within a certain distribution of saturation and hue. That is, this is a case where the attribute value is a continuous value. Also,
Considering the displacement, Table 1 shows 2000cc class, 16
Though discrete values such as 00cc class and 2800cc class are considered, 1500cc and 1600c are actually used.
c in 1600cc class, 1800cc, 2000c
c, 2200cc to 2000cc class, 2500c
Assuming that c, 2800 cc, and 3000 cc are in the 2800 cc class, there may be cases where representative discrete values are taken despite the distribution of discrete values. In this case, 1500c
It is considered that one attribute value distribution is c or more and less than 1800 cc, and similarly, it is 1800 cc or more and less than 2500 cc
More than 3000 cc may be considered as one attribute value distribution.

[0005] In addition, when tabulating or calculating actual data, it is better to consider the distribution of attribute values through statistical processing even though the data itself takes discrete values.

In addition, attribute values in actual pattern recognition and accident diagnosis, that is, actually measured values and sensor values have a range of attribute values usually due to changes in attribute values due to noise or various parameters.

[0007] In the case where the attribute value has a width or the distribution of the attribute value can be clearly obtained, the Japanese Patent Application No.
The method proposed in -269411 makes it possible to perform discrimination tree learning.

[0008]

According to the method proposed in Japanese Patent Application No. 3-269411, the category (actually the type of accident) and the category (actually, the type of accident) exist even when the attribute value has a width. Identification was possible only when there was an attribute whose attribute value distribution was completely separated (see (i) of FIG. 2).

Expert systems are useful for badly structured problems where the algorithm is unclear, and have been applied to diagnostics. However, expert systems assume that humans acquire knowledge. Even if it seems that knowledge acquisition has been completed, it is necessary to increase or correct the diagnostic knowledge in order to improve the reliability of the system. However, it is difficult to decide what kind of knowledge to add when making corrections. It is also difficult to maintain consistency with existing knowledge and verify knowledge. Therefore, acquiring, modifying, and increasing the knowledge in classifying data, and maintaining and verifying the consistency with the knowledge, require enormous human labor and development costs.

In the expert system, if-then
Another problem is that it takes a long calculation time for diagnosis because rules are used.

On the other hand, in the machine learning using the identification tree,
It is possible to automatically and efficiently create a diagnosis that does not involve human subjectivity. It is expected that a new identification tree will be developed by machine learning by using new attribute values.

Recently, a method of classifying data using a neural network has been studied. However, the classification logic of data is opaque, and in some cases, ambiguous results have been obtained, and it is difficult to estimate the accuracy of the results. It is. Increasing the number of intermediate layers improves the accuracy of the data classification result, but significantly increases the learning time, so that it is not easy to switch the nets that can be used for data classification. In the neural network, when an attribute that does not form a part of the target concept is given, it cannot be known that the attribute is irrelevant to the concept.

On the other hand, in the machine learning using the identification tree, if the identification tree becomes complicated, it becomes difficult to understand the grounds of the data classification result. You can know. If the diagnosis cannot be made, the cause can be estimated.

The present invention proposes a method of classifying non-overlapping portions of attribute values even if the distribution of the category and the attribute values of the category is not completely separated. For example, the attribute values a and b in FIG. 3 can be classified as category C _i, and the attribute values c and d can be classified as category C _j . Even for overlapping parts, that is, parts in which the categories cannot be classified, the probability distribution of the attributes is obtained to determine the occurrence probability of the overlapping parts, and the category is estimated. For example, it is possible to calculate the appearance probability by calculating the area of the portion indicated by the diagonal line in the probability distribution of FIG. 4A, and to calculate the occurrence probability when an arbitrary attribute value is obtained in the overlapping portion. it can. The appearance probability when the attribute value e is obtained from the probability distribution of FIG. 4 is indicated by a solid line.

Therefore, the problem to be solved by the present invention is to classify as much as possible even when the widths of the attribute values overlap as described above. When classification is not possible, the category and the frequency of appearance of the category that cannot be classified are given, and the situation in which classification is not possible is presented.

Further, in the method proposed in the above-mentioned prior application, when finding a set of attributes to be classified and arranging the most appropriate attribute among the set of attributes, the evaluation function is not suitable. Therefore, the effect was not sufficient. The present invention seeks to solve such a problem.

[0017]

Means for Solving the Problems To solve these problems, the data classification method of the present invention comprises the steps of: (a) setting categories C _{1 to} C _{i to} C _m for classifying data;
And the attributes T _{1 to} T _{j to} T _{n of} each category
Data is aggregated or calculated by calculation
Rate and divide the results into corresponding categories
Attribute that represents the distribution represented by the upper and lower limits for each attribute
Storing in a storage device as a value distribution table
With reference to the attribute value table stored in (b) the storage device
Then, for each attribute, a certain category C _i and another category C _i
-Analysis of the state of overlap of the attribute value distribution with C _j
The state of attribute value distribution for at least one attribute is
Attributes of the category C _j from the distribution of the attribute values of category C _i
Whether the distribution of values can be completely identified (i) or category
The distribution of the attribute values of the distribution of category C _j of attribute values of C _i and one
Partial overlap (ii) or category C
state distribution of the attribute values of _j are included in the distribution of the attribute values of C _i
The process of determining which state (iii) it belongs to
And performing, 1 if (c) each attribute is in the state (i), other
In the case of, define a coefficient of 0 and logically define each attribute.
As variables, and the category C _i and other categories C _j
The set of attributes that can classify
The logical expression of the logical variable with the same attribute
Step of selecting a formula as a set of attributes
If, on the combination of; (d) Category C _i and C _j is the state (i)
, The category C _i and all other categories
In step (c ), a set of attribute sets that can be classified is determined.
Performing the process of calculating the logical product of the set of logical expressions
If, on the combination of (e) the category C _i and C _j is the state (i)
To make all categories classifiable with each other.
A set of attribute sets for the logical expression obtained in step (d)
(F) performing the process of obtaining the identification tree from the set of the attribute sets;
In order to select a rational set of attributes,
Evaluation function based on the state of appearance and the appearance frequency of category C _i
Process to evaluate more and select the most efficient attribute set
(G) setting the attribute set selected in the step (f).
Of the attributes that have the highest rating
-And placed as parent node, and (g-1) parent node
The distribution of attribute values of attributes in a category included
Overlap with the distribution of attribute values of the attribute in other categories
If not, assign the category to the parent node.
Classification is completed by placing it as a child node to
(G-2) When overlapping, use another category and classification
The set of unsuccessful categories is assigned to the child node for the parent node.
(G-3) The category of the child node
The steps (b) to (e) are performed between
The evaluation in the set of attributes selected in step (f)
Excludes attributes used for classification at parent node for child nodes.
The attribute that is the largest of the attributes
Placed as attributes used to classify child nodes as nodes
(G-4) The processing of the above (g-1) to (g-3)
Between the category C _j in Tegori C _i and state (i)
Process until there are no more child nodes to be classified in
Cormorants and steps not completed classification in (h) wherein step (g)
For each child node, the category is divided into an attribute T
The distribution of attribute values for _k is s categories
C _1, ..., if C _i, are overlapping in ... C _s,
The distribution of attribute values for a certain attribute T _k is
Part where Lee C _i does not overlap with all other categories
Min, any category C _i and any other one category
-Overlap, any category C _i and any other 2
Where the categories overlap, ..., any category
-Part where C _i and any other s-1 categories overlap
To divide, for these to split the category attribute T _k
Except for the case of the empty set Te, split a new category
Performing a process to make Lee, using the probability distribution in the range of attribute values for each (i) attribute, stearyl
New category C _i and category created in top (h)
-Calculate the attribute value probabilities of areas where C _j does not overlap at all
Step a, the attribute stored in the storage device (j) said step (a) to
From the attribute values in the value table, find the frequency of the category,
The frequency of occurrence and the distribution of attribute values obtained in step (i)
Evaluation is performed using an evaluation function that uses the probability of overlap, and
Steps to select the attributes that are most effective for classifying the code
When could not be classified in (k) the step (c) ~ (g)
For the child node, the attribute that is most effective in classifying the child node
Of the category obtained in step (h) using the property
Perform processing to classify in a new category created by splitting
And (l) the knowledge created by the steps (b) to (k).
Create and store data classification flowcharts from different trees
Storing the data in an apparatus, and classifying the data according to the flowchart.

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be specifically described below.

FIG. 5 shows the concept of the present invention. FIG. 6 shows an overall flowchart of the present invention.

In the present invention, a method of classifying data in the case of transmission / distribution line fault diagnosis will be specifically described as an embodiment.

Normally, when data is obtained as continuous values by calculation or tabulation, within the range given by the continuous values,
It is given as the distribution of attribute values of each attribute. When the data is given as discrete values, a range of values having a standard deviation of 3σ is given as an attribute value distribution by, for example, statistical processing.

In this embodiment, the distribution of attribute values is obtained under the following conditions.

In the embodiment of the present invention, 3 shown in FIG.
Assuming a line model of a line distribution line, it is assumed that an accident has occurred on the line c. The capacitance of each distribution line is as shown. Further, the impedance of the secondary bus of the substation of the c distribution line is 0.362Ω and the line impedance is 0.536.
+ J 1.407Ω. The load is considered to be a uniform load, the ground voltage at the power supply end is 3810 V, and the current is 200 A. The standard deviation of the relative error is 3.8 based on the load prediction.
% And 3σ, 177A to 223A. In the event of a disconnection accident, the load was considered to be only a three-phase load, and the power factor of the load was considered to be 100%. The accident was considered to have occurred between the power supply end and the receiving end of the c distribution line, and the absolute values at the power supply end were calculated in a steady state for the 1-line ground fault, the 2-line short-circuit, and the 1-line disconnection. . The distribution line shall be ungrounded, and the fault point resistance in case of ground fault
6000Ω.

(A) First Embodiment 1) Method of Creating Identification Tree and Application of Data Classification in Distribution Line Fault Diagnosis In this embodiment, zero-phase current, zero-phase voltage, each phase current, Based on sensor information such as relative ground-to-ground voltage, an algorithm for detecting an accident such as a ground fault, short-circuit, or disconnection is created by a data classification tree. here,
The sensor information is used as the attribute value, and the categories of normal and short circuit accidents, ground fault accidents, and disconnection accidents are classified.

Here, m categories to be selected are represented by C
And ₁ ··· C _i ··· C _m, these categories of n attributes to T ₁ ··· T _j ··· T _n having individually.

The categories of data to be selected and the normal values in the classification of data in the distribution line accident diagnosis are as _follows : C _N : normal C _bc : bc line 2 line short circuit accident C _ca : ca line 2 line short circuit accident C _ab : ab line 2 Line short circuit accident C _a : a line ground fault accident C _b : b line ground fault accident Cc: c line ground fault accident C _Da : a line disconnection accident _CDB : b line disconnection accident C _Dc : c line disconnection accident

The attributes of each of the above categories are T _V0 : zero-phase voltage T _I0 : zero-phase current T _Ia : a-phase current T _Ib : b-phase current T _Ic : c-phase current T _Va : a relative ground Voltage T _Vb : b relative ground-to-ground voltage T _Vc : c relative ground-to-ground voltage.

Table 2 shows the attribute values calculated by the distribution line model.

[0029]

[Table 2] Here, the unit of the attribute value of T _V0 , T _Va , T _Vb , and T _Vc is V, and the unit of the attribute value of T _I0 , T _Ia , T _Ib , and T _Ic is A.

2) 2-1) Selection of Attributes Required for Classification of Arbitrary Two Categories (corresponding to 3 in the flowchart of FIG. 6) First, in order to find the attributes required for classifying all the categories, Attention is paid to one arbitrary category, and an attribute necessary for classifying it is obtained. The category of interest is C _i, and any one other than C _i C
Consider a relative distribution relationship with _j on the attribute value distribution diagram.
Relative distribution relationship C _j as viewed from the C _i in the distribution map in the attribute T _k, as shown in FIG. 2, can be considered the following three states.

State (i) The distribution of C _{i and} the distribution of C _j do not overlap. State (ii) The distribution of C _i all overlaps with the distribution of C _j . The distribution of the state (iii) C _i partially overlaps with the distribution of C _j .

Of these three states, the only state in which C _i and C _j can be completely classified is state (i). That is, in order for C _i and C _{j to} be categorized by an arbitrary attribute T _k , a necessary condition is that the state of the attribute value distribution of the two categories is state (i). Therefore, a coefficient _aik as shown in Expression (1) is defined to indicate whether the attribute _Tk is in the state (i).

[0033] a _ik = 1 T _k is the state (i) 0 Other (1)

_Considering that T _k is a logical variable, a binary coefficient of 1 is used when it is used for classification, and 0 when it is not used. The attribute value that allows classification of C _i and C _j is expressed by the following equation when expressed in the form of a logical sum using a coefficient a _ik .

[0035] _{f (C i, C j)} = a i1 T 1 + ··· + a ik T k + ···· a in T n (2)

That is, in equation (2), C _i and C _j are f (C
_i , C _j ) = 1, classification is possible, and f (C _i ,
If at least one attribute of the term C _j ) is used, C _i and C _j
Can be classified.

The result of selecting the attributes necessary for classifying the category C _N from other categories is shown below.

F (C _N , C _bc ) = T _Ib + T _Ic + T _Vb + T _Vc (3) f (C _N , C _ca ) = T _Ia + T _Ic + T _Va + T _Vc (4) f (C _N , C _ab) ) = T _Ia + T _Ib + T _Va + T _Vb (5) f (C _N , C _a ) = T _V0 + T _I0 + T _Va + T _Vb + T _Vc (6) f (C _N , C _b ) = T _V0 + T _I0 + T _Va + T _Vb + T _Vc (7) f (C _N , C _c ) = T _V0 + T _I0 + T _Va + T _Vb + T _Vc (8)

However, f (C _N , C _Da ), f
(C _N , C _Db ) and f (C _N , C _Dc ) have no attribute whose attribute T _k is state (i). With the method proposed earlier, it was impossible to create an identification tree in this case.

2-2) Selection of Attributes Necessary for Classification of Attention Category (corresponding to 4 in the flowchart of FIG. 6) Here, the category C _{i of} interest and a certain attribute T _k are in state (i). A set of attributes that can classify all the categories C _j in the state is obtained.

The attributes C _i and attribute T _k is to be classified and one C _j state is the state (i) is been determined by equation (2). Thus, C _i and attribute T _k is f with respect to the state (i) a is all categories and the C _i is to allow classify other each category of state (C _i, C _j)
The logical product of (j = 1,..., M, i ≠ j) is performed as in Expression (9).

F (C _i ) = f (C _i , C ₁ ) ·· f (C _i , C _j ) ·· f (C _i , C _m ) where i ≠ j (9)

That is, the set of attributes given in the form of a logical product in the operation result of f (C _i ) is independent of all the categories of the state in which C _i and the attribute T _k are in state (i). Is a set of attributes that can be classified.

[0044] That is, for any of C _j category C _i and attributes T _k is not in the state of the state _(i) f (C i,
C _j ) is not included as a term of the logical product of f (C _i ). However, if f (C _i , C _j ) = 0 in C _i and all other categories, f (C _i ) is not obtained, and
Move on to the next calculation step.

According to the above equation, the attribute necessary for classifying C _N and all the categories C _j in the state where the attribute T _k is the state (i) is expressed by the following equation (3) as in the following equation (10). )
To (8).

F (C _N ) = f (C _N , C _bc ) f (C _N , C _ca ) f (C _N , C _ab ) f (C _N , C _a ) f (C _N , C _b ) f _{(C N, C c) =} T Ic T Ia T V0 + T Ic T Va T I0 + T Ic T Va T V0 + T Vb T Ia T I0 + T Vb T Ia T V0 + T Vb T Ic T I0 + T Vb T Ic T V0 + T _Vb T _Va + T _Vb T _Vc + T _Vc T _Ia T _I0 + T _Vc T _Ia T _V0 + T _Vc T _Va (10) In the expression (10), the _CN has the attribute T _k according to each attribute set of the 12 terms. All categories _Cj in the state (i) can be classified.

2-3) Selection of attributes that can classify all categories in which attribute T _k is in state (i) (corresponding to 5 in the flowchart of FIG. 6) Each category determined by equation (9) is an attribute T _k is at least 1 from the set of attributes required to classify the category of state in state (i).
By taking out each set and using a set of attributes including all of them, the category of the state where the attribute T _k is state (i) can be classified. That is, a set of attributes necessary for the attribute T _k to be able to classify the state category of the state (i) is:
For each category, f (C _i ) = 1 (i = 1,.
.., m) are obtained by finding an attribute that can be expressed as: (m), and their logical product is performed as in Expression (11).

E = f (C ₁ )... F (C _i )... F (C _m ) (11) The operation result is a product-sum form of a logical expression and can be expressed as follows.

[0049] _{E = T 1 · T 2 ·} T 3 ··· + ··· + T a · T b · T c ··· + ··· + T p · T q · T r ··· where A _{1 = T 1 · T 2 · T} 3 ···, a x = T a T b T c ···, When _{a p = T p · T q} · T r ···, E = a 1 + ·· · + a _x + a ··· + a _p (12).

Therefore, A ₁ ,..., A _{x ,.}
Is a set of attributes necessary for the attribute T _k to classify the category of the state of the state (i).

Similarly, f (C _bc ), f (C _ca ) and f (C _ca )
(C _ab ), f (C _a ), f (C _b ), f (C _c ), f (C
A set of attributes that allows the attribute T _k for _{obtaining Da} ), f (C _Db ), and f (C _Dc ) to classify the category of the state of the state (i) is:

E = f (C _N ) f (C _bc ) f (C _ca ) f (C _ab ) f (C _a ) f (C _b ) f (C _c ) f (C _Da ) f (C _Db ) f (C _Dc ) = T _Ia T _Ib T _Va T _Vb T _Vc + T _Ia T _Ic T _Va T _Vb T _Vc + T _Ib T _Ic T _Va T _Vb T _Vc (13) Replace this with:

A ₁ = T _Ia T _Ib T _Va T _Vb T _Vc , A ₂ = T _Ia T _Ic T _Va T _Vb T _Vc , A ₃ = T _Ib T _Ic T _Va T _Vb T _Vc (14)

In other words, these three sets are attribute sets that enable classification of a state category in which at least one attribute T _k is state (i). Also,
If you choose a set of attributes of _{A 3, T Ib, T Ic} , T Va, T
Attributes other than _Vb and _TVc are attributes that are not required for classification.

Considering the procedures so far, the categories connected by the solid line in FIG. 8 can be classified by using the attributes of the three sets of attributes A ₁ , A ₂ , and A ₃ obtained. . Therefore, for each of the three sets of attributes A ₁ , A ₂ , and A ₃ , the category connected by the solid line can be classified, but the category connected by the broken line cannot be completely classified.

3) Arrangement of attributes in each node of the identification tree (corresponding to 6 in the flowchart of FIG. 6) First, an optimal set of attributes selected to classify the categories connected by solid lines is selected. And how to arrange them most efficiently.

3-1) Evaluation method for selecting an optimal attribute (1) Degree of non-overlapping a _k (i, j) In an attribute value distribution, an attribute value having many areas having no overlap with other distributions is classified. Contributes more. Constructing an identification tree using a set of attributes including a large number of such attribute values increases the probability of completion of classification at a higher-level node, leading to a reduction in classification time. Therefore, a non-overlapping degree a _k (i, j) indicating how much a region that does not overlap at all with the attribute value distribution of a certain category C _i occupies the attribute value distribution of C _j is expressed by the following equation. This shows how T _k is the contribution degree C _j with respect to the classification of C _i.

A _k (i, j) = l _ik / L (C _i ) (15) Here, in the attribute value distribution of l _ik : T _k , there is no overlap with the distribution of C _{i due to} the distribution of C _j. Range of region (see FIG. 9) L (C _i ): range of distribution of C _i

The non-overlapping degree a _k (i, j) is calculated for T _V0 based on the measurement data in Table 2 above, and Table 3
become that way. In Table 3, i represents a column and j represents a row.

[0060]

[Table 3]

(2) Appearance Frequency P ₁ Next, the category, that is, the appearance frequency of the type of accident is obtained. Table 4 shows the results.

[0062]

[Table 4]

Naturally, the appearance frequency P ₁ is normal time≫ground fault>
Short circuit accident> Disconnection accident.

(3) Evaluation value F (T _k ) Using the two parameters a _k (i, j) and P ₁ described above, an evaluation function as shown in the following equation is determined for each attribute.

[0065]

(Equation 1)

[0066] The F (T _k) is larger T _k, there is a high possibility of classification for large category of frequency of occurrence.

In the case of the above example, the evaluation values are as follows.

[0068]

[Table 5]

Hereinafter, the process proceeds with definition 2.

A set of attributes A _eff effective for creating an identification tree is
This is a set that maximizes the product G (A _x ) of the evaluation values F (T _k ) of the respective attributes of A ₁ , A ₂ , and A ₃ . Therefore, G (A _x ) is obtained for each set.

G (A ₁ ) = F (T _Ia ) F (T _Ib ) F (T _Va ) F (T _Vb ) F (T _Vc ) = 5.364 × 10 ⁻¹² (18) G (A ₂ ) = F (T _Ia ) F (T _Ic ) F (T _Va ) F (T _Vb ) F (T _Vc ) = 5.364 × 10 ^-12 (19) G (A ₃ ) = F (T _Ib ) F ( T _Ic ) F (T _Va ) F (T _Vb ) F (T _Vc ) = 8.648 × 10 ⁻¹² (20) Since the value of G (A ₃ ) is the maximum, G (A ₃ ) is taken.

4) Attributes to each node of the identification tree (corresponding to 7 in the flowchart of FIG. 6) The arrangement of the identification tree to each node is as follows. First, regarding the root node, the one having the largest evaluation value F (T _k ) of A _eff is considered as the root node. Here, F (T _Va ) = F (T
_Vc), so the T _Va. Depending on the state of the attribute overlap, the area is divided into an area where the attribute distribution does not overlap and an area where the attribute distribution overlaps.

When the attribute becomes the value of the non-overlapping area, the classification is completed at the root node. Overlapping regions cannot be classified between categories, and are re-classified with other attributes. That is, the former is the leaf node N _e,
The latter is a re-classification node _Nc . Set in N _c N _c '
For example, regarding the area 1 shown in FIG. 10, (C _ca , C
_ab , C _a ).

Next, the attribute to be arranged in the re-classification node is selected as follows. Taking region 1 as an example, it is possible to classify two categories of elements of S _c ′. The attributes are as follows: Here, f (C _i , C _j ) = f (C _j ,
C _i ).

F (C _ca , C _ab ) = T _Ib + T _Ic + T _Vb + T _Vc (21) f (C _ab , C _a ) = T _V0 + T _I0 + T _Ia + T _Ib + T _Vb + T _Vc (22) f (C _a , C _ca ) = T _V0 + T _I0 + T _Ia + T _Ic + T _Vc (23)

An attribute that allows all elements of S _c ′ to be classified is
The logical product of these results in the following equation.

F (C _ca , C _ab ) f (C _ab , C _a ) f (C _a , C _ca ) = T _V0 T _Ib + T _V0 T _Ic + T _I0 T _Ib + T _I0 T _Ic + T _Ia T _Ib + T _Ia T _Ic + T _Ib T _Ic + T _Vc + T _V0 T _Vb + T _I0 T _Vb + T _Ia T _Vb + T _Ic T _Vb (24)

As a result, the attributes that are a subset of A _eff are T _Vc , T _Ib T _Ic , and T _Ic T _Vb . The set of three attributes, the set of attributes of the product G (A _X) Request the T _Vc of the evaluation value F for each attribute (T _k) is to place the T _Vc since the maximum.

The above operation is performed using the attributes of the subset of A _eff . FIG. 11 shows a part of the result. However, nodes marked with * cannot be separated by the conventional procedures.

[0080] 5) divided categories (corresponding to 8 in the flowchart of FIG. 6) s number of category C _1, which is in the attribute distribution T _k with,
.., C _i ,..., C _s overlap, ie, s
All combinations of the categories are in the state (i
In the case of i) or state (iii), the category is divided by the following method.

For any attribute T _k , any category C _i
Is a portion that does not overlap with all other categories, a portion where any category C _i overlaps with any other one category, a portion where any category C _i overlaps with any other two categories, ... , An arbitrary category C _i and another arbitrary s−2 categories overlap, and an arbitrary category C _i and another arbitrary s−1 categories overlap. With the above division, a new divided category can be created. In addition, the number of combinations of a portion where an arbitrary category C _i and another arbitrary sn category overlap with each other is given by _s C _{s−n + 1} . If the divided category is an empty set for all the attributes T _k , no new category is created.

More specifically, consider a case in FIG. 12 in which categories C ₁ , C ₂ and C ₃ having three child nodes cannot be distinguished. The attributes here are T ₁ and T ₂ .

Here, a category newly created by a portion where the category C _i does not overlap with all other categories is defined as C _{1 *} . For example, C _{1 *} attributes T ₁ of the FIG. 12, there is a case where the distribution of the attribute as in this example are separated.
The newly created category is represented by C _ij by the portion where any category C _i overlaps with any one other C _j category.
And Hereinafter, a newly created category is similarly defined by a portion where an arbitrary category C _i and another arbitrary two categories overlap. C _{2 *} in attribute T ₁ in FIG.
And C _{3 *} does not create a new category due to the empty set. At this time, the newly created category satisfies state (1) in all arbitrary two combinations. The attribute T _k can be divided into categories by the above method.

With the above method, the category of the node is divided. Usually, two categories cannot be classified by the mark * in FIG. 11, but there are cases where the four categories C _N , C _Da , C _Db , and C _Dc cannot be recognized as shown by * 1.

Considering * 1, an arbitrary category C _i in the attribute T _k is a category of a portion that does not overlap with all other categories, in this case, C _{N *} , C _{Da *} ,
Although C _{Db *} and C _{Dc *} can be created, category C _{N *} cannot be created because C _{N *} is an empty set for all attributes. Similarly, the category of a portion where any category C _i overlaps with any other one category, the category of a portion where C _i overlaps with any other two categories, and the category where C _i overlaps with any three other categories Depending on the category of the part, eight new categories C _{Da *} , C
_{Db *} , C _{Dc *} , C _{DaDb *} , C _{DbDc *} , C _{DbDc *} , C _{DaDbDc *} ,
C _{NDaDbDc *} can be made. Table 6 shows the distribution of the attribute values of the category.

[0086]

[Table 6]

6) Category Classification without Separated Attributes A new category is generated by dividing a category into categories whose attribute value distributions are not completely separated. To consider which attributes are used to classify them, consider the probability distribution of the attributes.

6-1) Attribute probability distribution (corresponding to 9 in the flowchart of FIG. 6) Attribute values are calculated using several parameters. For example, in the case of a current value of a disconnection accident, the current before the accident and the accident point are parameters. If the probability distribution of those parameters is known, the probability distribution of the attribute can be known. If the distribution of the attribute overlaps in any two categories, the probability of the overlapping part of each category does not overlap. You can know the probability of the part.

Here, the concept will be described using a discrete two-dimensional probability distribution as an example. The two random variables X and Y are independent of each other, and given arbitrary values x _i and y _i , respectively, the probabilities p _i and q _j are given by

P (X = x _i ) = p _i (25) P (Y = y _j ) = q _j (26), and P (X = x _i , _{Y = y j) = P (} X = x i) P (Y = y j) (27) i.e. _{p ij = p i q j (} 28) holds.

Now, assuming that the attribute T _k can be expressed as a function h of the variables X and Y, it can be expressed as in equation (29), and T _k = h (X, Y) (29)

[0092] Thus for example, in the case of the current value of accidental disconnection, before the accident two parameters current value x _i and the fault point y _j
By _calculating the sum of h (X, Y) and p _ij from the probability and the probabilities p _i and p _j , the probability distribution of the attribute T _k can be obtained. Therefore, _{assuming that} the probability distribution of the attribute T _k is Z and the probability of the sum of p _ij is z _k , the probability distribution of the attribute T _k can be expressed as P (Z = z _k ) (30). _In the attribute distribution of _k , the probability in the range of a ≦ Z ≦ b is as follows: P (a ≦ Z ≦ b) = Σ * _pr (31) where Σ ^* _pr is the sum of the probability that a ≦ Z ≦ b. Represent.
Therefore, the probability of any attribute T _k at Z and the attribute T _k
Can be obtained in the attribute distribution of a ≦ Z ≦ b.

6-2) Attribute Selection Using the probability distribution obtained in the previous section, an attribute is selected which is effective for classifying child nodes. In the attribute value distribution, an attribute value having a high probability of a portion that does not overlap with other distributions has a high contribution for classification. Therefore, a probability distribution of attribute values of a region of a certain attribute T _k that does not overlap C _j of the category C _i at all is obtained, and the probability is defined as Σ ^* _pr (i, j). Actually, it is the probability other than the shaded portion in FIG. This shows how T _k is the effect of extent C _j with respect to the classification of C _i. Probability _{^{Σ * p r (i, j}} ) and use the frequency of occurrence P _i of the category, define the following evaluation function.

[0094]

(Equation 2)

Here, a _k (i, j) determined by equation (17)
Is Σ ^* when the probability distribution of the attribute T _k is uniformly distributed
p _r (i, j). Using the attribute T _k having a large F ^* (T _k ), the classification of the child node in which the distribution of the attribute values is not completely separated for all the attributes is a category newly formed by the category division described in 8). Classify according to

6-3) Embodiment of attribute probability distribution Here, the current value of each phase when one line is broken is taken as an example. When the current value is a three-phase load and the power factor is 100%, the current at the time of the fault at the fault point is 0 A in the disconnection phase and the currents of the other two phases are half the current before the fault. The load before the accident point does not fluctuate. For example, assuming that the disconnection was made exactly at the center of the c-distribution line in FIG. 7 and the current at the power supply end in a normal state was 200 A, the load was an equal load. 100A of the load to the center of the current flows. The other two phase currents are 100% of the load to the center of c.
A and a total of 150A of 50A after the disconnection point flow.

Since the probability distribution of the current value in the load prediction before the accident is a normal distribution, and the accident is considered to occur uniformly on the distribution line, the probability distribution of the accident point may be considered to be a uniform distribution. . Therefore, the probability distribution of the current value of each phase at the time of disconnection can be calculated from the two-dimensional probability distribution of the above two random variables by the equation (30).

Here, the probability is considered as a continuous value, and FIG.
The probability density functions of the currents of the a, b, and c phases at the time of phase disconnection and the probability density functions of the normal times are shown. Attribute T _{Ia of} category C _Da
The probability of the portion where the attribute distribution does not overlap with the attribute distribution of the attribute T _Ia of the category C _N is 0 to 176 in the graph of FIG.
A can be calculated from the area of the probability density function up to A, the probability is 0.88, and the probability of the overlapping part is 0.12.
Can be requested. Attribute value of the attribute T _Ia is obtained and 180A, the probability when the classification categories C _Da or category C _N is not possible with other attributes than the probability density function of FIG. 13, C _Da is 0.005, C _N Is 0.0016
Is obtained. However, considering the category occurrence frequency P _i from Table 4, the probability of a disconnection accident is small, so C _Da is 1.05 × 10 ⁻⁴ and C _N is 1.58 × 10 ^−4.
And the probability of normality is high.

When the attribute value of the attribute T _Ia is obtained as 150 A, the probability becomes 0.00 from the probability density function shown in FIG.
5. Considering the appearance frequency P _i of the category in Table 4,
The probability of being C _Da can be obtained as 1.05 × 10 ⁻⁴ .

As described above, the most rudimentary method of the conventional disconnection detection is the current value detection at the power supply end. However,
The biggest disadvantage of this method is that if a disconnection accident occurs near the end, it is not known whether the disconnection accident has occurred or the load has simply decreased. Therefore, there is a method of detecting a disconnection accident using sensor information, that is, an attribute value at the end of a distribution line, but a communication line to a substation must be installed, which increases costs. It has been found in principle that it is impossible in principle to completely detect a disconnection accident to the end by detecting the power supply end.

Therefore, in order to classify a disconnection accident or normal with higher accuracy, it is necessary to increase the accuracy of load prediction. For example, it is considered that the load prediction one hour ago has a smaller relative error in the prediction than the load prediction five hours ago used as the current data. That probability distribution of the normal distribution of FIG. 13, by gathering near the 200A, the C _N T
The range of attribute distributions taken by _Ia , T _Ib , and T _Ic becomes narrow. Therefore, the probability of detecting a disconnection accident can be increased. As a result, short-time load prediction improves the accuracy of load prediction and takes into account load fluctuations that change with time.
By making the diagnosis algorithm at that time by discriminating tree learning, it becomes possible to detect a disconnection accident closer to the end than was previously impossible.

[0102] 6-4) Table four categories _{C N, C Da, C Db} , * the evaluation function F at C _Dc to (T _k), where classification calculation Narabiniko node evaluation function F ^* (T _k) 4, the appearance frequency P _i and (31)
The evaluation function F ^* (T _k ) is calculated by Expression (32) using Σ ^* _pr (i, j) obtained from the expression. As for Σ ^* _pr (i, j), for example, the probability of a region not overlapping at all with C _N of C _Da of the attribute T _Ia is obtained as 0.88 as obtained in the previous section. Table 7 shows the results. In this case, F ^* (T _Va ) = F ^* (T
_{Since Vb} ) = F ^* (T _Vc ), classification is based on the attribute T _Va . Categories C _{Da *} , C _{Db *} ,
C _{DbDc *} and C _{NDaDbDc *} can be classified. The result is shown in FIG.

[0103]

[Table 7]

FIGS. 14 to 19 show flowcharts for classifying data for concretely diagnosing an accident using the numerical values used in the above-described embodiment.

The first embodiment has been described above. Although the first embodiment is the most efficient method for creating an identification tree and a flowchart, similar classification can be performed even if efficiency is sacrificed to some extent. An example is shown below.

(B) Second Embodiment In this embodiment, in a combination of categories in which at least one attribute is state (i), a set of attributes for classifying category C _i and all other categories is obtained. To find a set of attributes that classify categories
Although the steps of the steps are performed, in the present embodiment, the above steps are performed at once. The description up to 2-1) of the first embodiment is omitted because it is the same.

2-2 ') Select an attribute capable of classifying all categories in which the attribute T _k is in the state (i).

A set of attribute sets that can identify a combination of categories in which the distribution of at least one or more attribute values is completely separated corresponds to all combinations where f (C _i , C _j ) = 1. f (C _i , C _j ) (i = 1,..., n,
j = 1,..., m, i ≠ j), and can be obtained by equation (33).

E = f (C ₁ , C ₁ ) ·· f (C ₁ , C _j ) ··· f (C ₁ , C _m ) ······· f (C _i , C ₁ ) ·· f (C _i , C _j ) f (C _i , C _m ) f (C _n , C ₁ ) f (C _n , C _j ) f (C _n , C _m) ) Where i ≠ j (33)

The result of the operation is a product-sum form of a logical expression and can be expressed as follows. E = T ₁ · T ₂ · T ₃ ··· + + T _a · T _b · T _c ··· + + T _p · T _q · T _r ··· where A ₁ = T ₁ ··· T ₂ · T ₃ ···, A _x = T _a T _b T _c ···, A _p = T _p · T _q · T _r ··· E = A ₁ +... + A _x + ... + A _p (34) .

Therefore, A ₁ ,..., A _x ,.
A _p is the attribute set that can classify the combination of category distribution of the attribute values of the at least one attribute T _k are completely separated.

According to equation (33), a set of attribute sets that can classify all combinations of categories in which the distribution of at least one or more attribute values is completely separated can be selected.

E = T _Ia T _Ib T _Va T _Vb T _Vc + T _Ia T _Ic T _Va T _Vb T _Vc + T _Ib T _Ic T _Va T _Vb T _Vc (35) Replace this with: A ₁ = T _Ia T _Ib T _Va T _Vb T _Vc , A ₂ = T _Ia T _Ic T _Va T _Vb T _Vc A ₃ = T _Ib T _Ic T _Va T _Vb T _Vc (36)

That is, these three sets are sets of attributes that allow the category of the state in which the attribute T _k is state (i) to be classified independently.

Considering the procedure so far, the categories connected by the solid line in FIG. 8 can be classified by using the attributes of the three sets of attributes A ₁ , A ₂ , and A ₃ obtained. . Therefore, for each of the three sets of attributes A ₁ , A ₂ , and A ₃ , the category connected by the solid line can be classified, but the category connected by the broken line cannot be completely classified.

The following is the same as 3) and subsequent steps in the first embodiment, and a description thereof will not be repeated.

( C ) Third Embodiment In this embodiment, in a combination of categories in which at least one attribute is state (i), a set of attributes to be classified is obtained, and one set is arbitrarily selected from the obtained sets. When arranging a selected and further set of selected attributes in a node, an arbitrary attribute is arranged. This is because it does not take into account the point of selecting an efficient attribute set, and furthermore, when arranging the attribute set in the node, from the efficient attribute.
It differs from the first embodiment in that it is arbitrarily selected and arranged.

Since the steps up to 2-3) of the first embodiment are the same, the description will be omitted. The following items are added after section 2-3).

An arbitrary set of A ₁ , A ₂ and A _{3 in} the equation (14) is selected. Here, it is assumed to select the A _3. Any attribute in the set of attributes of A ₃ considered the root node. Here, _let it be T _Va .

Depending on the state of the attribute overlap, the area is divided into an area where the attribute distribution does not overlap and an area where the attribute distribution overlaps.

When the attribute becomes the value of these non-overlapping areas, the classification is completed at the root node. Overlapping regions cannot be classified between categories, and are re-classified with other attributes. That is, the former is the leaf node N _e,
The latter is a re-classification node _Nc . Set in N _c N _c '
For example, regarding the area 1 shown in FIG. 10, (C _ca , C
_ab , C _a ).

Next, the attribute to be arranged in the re-classification node is selected as follows. Taking region 1 as an example, it is possible to classify two categories of elements of S _c ′. The attributes are as follows: Here, f (C _i , C _j ) = f (C _j ,
C _i ).

F (C _ca , C _ab ) = T _Ib + T _Ic + T _Vb + T _Vc (21) f (C _ab , C _a ) = T _V0 + T _I0 + T _Ia + T _Ib + T _Vb + T _Vc (22) f (C _a , C _ca ) = T _V0 + T _I0 + T _Ia + T _Ic + T _Vc (23)

An attribute that allows all elements of S _c ′ to be classified is
The logical product of these results in the following equation.

F (C _ca , C _ab ) f (C _ab , C _a ) f (C _a , C _ca ) = T _V0 T _Ib + T _V0 T _Ic + T _I0 T _Ib + T _I0 T _Ic + T _Ia T _Ib + T _Ia T _Ic + T _Ib T _Ic + T _Vc + T _V0 T _Vb + T _I0 T _Vb + T _Ia T _Vb + T _Ic T _Vb (24)

[0126] As a result, the attribute that is a subset of A ₃ is a _{T Vc, T Ib T Ic,} T Ic T Vb. For this set of three attributes, an arbitrary _TVc is placed here.

[0127] or more of such an operation performed using the attributes of a subset of A _3. FIG. 11 shows a part of the result. However,
Nodes marked with * cannot be separated by the previous procedure.

Hereinafter, since it is the same as 5) and subsequent steps of the first embodiment, the description is omitted.

( D ) Fourth Embodiment In this embodiment, a set of attributes in which at least one attribute is classified as a combination of categories of state (i) is obtained, and the most efficient attribute set is obtained from the obtained sets. Is obtained, and when the attribute set is arranged in the node, the attribute is arbitrarily selected and arranged in the node. Therefore, the first and second
The embodiment is the same in consideration of efficiency up to the point of selecting a set of attributes. However, when the set of attributes is arranged in a node, it is arbitrarily selected. The difference is that it is not done.

Up to 3) of the first embodiment, the description is omitted because it is the same.

4) Attributes to each node of the identification tree (corresponding to 7 in the flowchart of FIG. 6) The arrangement of the identification tree to each node is as follows. First, an arbitrary T _Va of A _eff is set for the root node. Depending on the state of attribute overlap, areas where attribute distribution does not overlap,
It is divided into areas that overlap the distribution of attributes.

When the attribute becomes the value of these non-overlapping areas, the classification is completed at the root node. Overlapping regions cannot be classified between categories, and are re-classified with other attributes. That is, the former is the leaf node N _e,
The latter is a re-classification node _Nc . Set in N _c N _c '
For example, regarding the area 1 shown in FIG. 10, (C _ca , C
_ab , C _a ).

Next, the attribute to be arranged in the re-classification node is selected as follows. Taking region 1 as an example, it is possible to classify two categories of elements of S _c ′. The attributes are as follows: Here, f (C _i , C _j ) = f (C _j ,
C _i ).

F (C _ca , C _ab ) = T _Ib + T _Ic + T _Vb + T _Vc (21) f (C _ab , C _a ) = T _V0 + T _I0 + T _Ia + T _Ib + T _Vb + T _Vc (22) f (C _a , C _ca ) = T _V0 + T _I0 + T _Ia + T _Ic + T _Vc (23)

An attribute that allows all elements of S _c ′ to be classified is
The logical product of these results in the following equation.

F (C _ca , C _ab ) f (C _ab , C _a ) f (C _a , C _ca ) = T _V0 T _Ib + T _V0 T _Ic + T _I0 T _Ib + T _I0 T _Ic + T _Ia T _Ib + T _Ia T _Ic + T _Ib T _Ic + T _Vc + T _V0 T _Vb + T _I0 T _Vb + T _Ia T _Vb + T _Ic T _Vb (24)

As a result, the attributes that are a subset of A _eff are T _Vc , T _Ib T _Ic , and T _Ic T _Vb . For this set of three attributes, an arbitrary _TVc is placed here.

The above operation is performed using the attributes of the subset of A _eff . FIG. 11 shows a part of the result. However, nodes marked with * cannot be separated by the conventional procedures.

Hereinafter, since it is the same as 5) and subsequent steps of the first embodiment, the description is omitted.

( E ) Fifth Embodiment In this embodiment, a set of attributes in which at least one attribute is a combination of the category of the state (i) is obtained, and the attribute of each of all the obtained sets is determined. When an arbitrary set is selected from the set of mutually identifiable attributes evaluated based on the evaluation function and determined above, and the selected set of attributes is arranged in the node, the identification becomes efficient. So that priority is given to efficient attributes. Therefore, in the present embodiment, an efficient attribute set is not selected, but when arranging the selected attribute set in the node, the efficient attribute is preferentially arranged.

In the section (3) of the first embodiment 3), the description is the same up to the point of proceeding with the definition 2. A in the set of (14)
Select ₃ .

4) Attributes to each node of the identification tree (corresponding to 7 in the flowchart of FIG. 6) The arrangement of the identification tree to each node is as follows. First, regarding the root node, the one having the largest evaluation value F (T _k ) of A ₃ is considered as the root node. Here, F (T _Va ) = F
(T _Vc ), so T _Va . Depending on the state of the attribute overlap, the area is divided into an area where the attribute distribution does not overlap and an area where the attribute distribution overlaps.

When the attribute becomes the value of these non-overlapping areas, the classification is completed at the root node. Overlapping regions cannot be classified between categories, and are re-classified with other attributes. That is, the former is the leaf node N _e,
The latter is a re-classification node _Nc . Set in N _c N _c '
For example, regarding the area 1 shown in FIG. 10, (C _ca , C
_ab , C _a ).

Next, the attribute to be arranged in the re-classification node is selected as follows. Taking region 1 as an example, it is possible to classify two categories of elements of S _c ′. The attributes are as follows: Here, f (C _i , C _j ) = f (C _j ,
C _i ).

F (C _ca , C _ab ) = T _Ib + T _Ic + T _Vb + T _Vc (21) f (C _ab , C _a ) = T _V0 + T _I0 + T _Ia + T _Ib + T _Vb + T _Vc (22) f (C _a , C _ca ) = T _V0 + T _I0 + T _Ia + T _Ic + T _Vc (23)

An attribute that allows all elements of S _c ′ to be classified is
The logical product of these results in the following equation.

F (C _ca , C _ab ) f (C _ab , C _a ) f (C _a , C _ca ) = T _V0 T _Ib + T _V0 T _Ic + T _I0 T _Ib + T _I0 T _Ic + T _Ia T _Ib + T _Ia T _Ic + T _Ib T _Ic + T _Vc + T _V0 T _Vb + T _I0 T _Vb + T _Ia T _Vb + T _Ic T _Vb (24)

[0148] As a result, the attribute that is a subset of A ₃ is a _{T Vc, T Ib T Ic,} T Ic T Vb. Regarding a set of these three attributes, the one having the largest evaluation value F (T _k ) of each attribute is considered as a root node. Wherein the set of attributes of the T _Vc is to place the T _Vc since the maximum.

[0149] or more of such an operation performed using the attributes of a subset of A _3. FIG. 11 shows a part of the result. However,
Nodes marked with * cannot be separated by the previous procedure.

The following is the same as 5) et seq. Of the first embodiment, and a description thereof will be omitted.

( F ) Sixth Embodiment In this embodiment, when dividing a child node, an arbitrary attribute is selected without selecting the most efficient attribute. Therefore, in this embodiment, when dividing a child node,
The feature is that efficient attributes are not selected.

Up to 5) of the first embodiment is the same. First
6-1), 6-2) and 6-3) of the embodiment are omitted, and 6-
4) is changed as follows.

6-4) Identification of Child Nodes Categories C _{Da *} , C _{Db *} , C _{DbDc *} , and C _{NDaDbDc *} can be identified as child nodes to be identified from an arbitrary attribute T _Va . The result is shown in FIG.

FIGS. 14 to 19 show flowcharts for classifying data for concretely diagnosing an accident using the numerical values used in the above-described embodiment.

[0155]

As described above, the present invention has the following effects.

(1) In any two categories,
Even if the distribution of the attribute values is not completely separated, a portion where the attribute values do not overlap can be classified.

(2) Even for an overlapping portion, that is, a portion in which a category cannot be identified, the probability of the overlapping portion can be obtained by obtaining the probability distribution of the attribute, and the category can be estimated.

(3) When an arbitrary attribute value is obtained, an appearance probability can be obtained. When categories cannot be separated by the attribute, it is possible to know the certainty to which category the attribute belongs.

(4) From the identification tree and the flowchart,
You can see what categories cannot be classified, and you can know the range of attribute values at that time.

(5) When the attribute values of data have a distribution,
It can be applied to various classifications such as diagnosis, pattern recognition, and image processing.

(6) When the distribution of attribute values is obtained by a simulator or the like, even if the parameters of the simulator are changed, the data can be quickly created by learning the classification of the data by machine learning in accordance with the change. .

(7) It is possible to automatically create an algorithm that does not allow human subjectivity.

(8) An efficient algorithm can be created.

(9) Attributes unnecessary for data can be known.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram showing a result of discrimination tree learning based on data in Table 1.

FIG. 2 is an explanatory diagram showing a relationship between an attribute value distribution and two arbitrary categories.

FIG. 3 is an explanatory diagram showing classification of a category in a case where there is a partial overlap.

FIG. 4 is an explanatory diagram showing an appearance probability of a part having an attribute overlap and an appearance probability at an arbitrary value.

FIG. 5 is a conceptual diagram showing a procedure of creating an algorithm according to the present invention.

FIG. 6 is an overall flowchart of the present invention.

FIG. 7 is a system diagram of a distribution line model in an embodiment of the present invention.

FIG. 8 is an explanatory diagram showing a classifiable / impossible relationship between categories.

FIG. 9 is an explanatory diagram of a distribution of categories in which attributes overlap.

FIG. 10 is a diagram showing an example of a distribution of attribute values of each category.

FIG. 11 is an explanatory diagram of an identification tree in the embodiment of the present invention.

FIG. 12 is an explanatory diagram showing an example of category division in the present invention.

FIG. 13 is a graph showing an example of a probability distribution of a current in the present invention.

FIG. 14 is a flowchart (1) showing an example in which the present invention is applied to an accident diagnosis.

FIG. 15 is a flowchart (2) showing an example in which the present invention is applied to an accident diagnosis.

FIG. 16 is a flowchart (3) showing an example in which the present invention is applied to an accident diagnosis.

FIG. 17 is a flowchart (4) showing an example in which the present invention is applied to an accident diagnosis.

FIG. 18 is a flowchart (5) showing an example in which the present invention is applied to an accident diagnosis.

FIG. 19 is a flowchart (6) showing an example in which the present invention is applied to an accident diagnosis.

Continued on the front page (56) References Patent No. 3019227 (JP, B2) Information Processing Society of Japan 43rd (late 1991) National Convention 7D-6 "Learning Study of Decision Tree in Distribution Line Accident Diagnosis System" Masato Togami et al.

Claims (6)

(57) [Claims] [Claim 1] (a) category C ₁ ~ to classify the data
C _{i to} C _m are set, and the attribute T of each category is set.
₁ through T _j through T _n aggregates the measurement data for each or calculated
And simulate the result with the corresponding category
And the upper and lower limits for each attribute
Stored in the storage device as an attribute value distribution table representing cloth
That the reference and the step, the attribute value table stored in (b) the storage device
Then, for each attribute, a certain category C _i and another category C _i
-Analysis of the state of overlap of the attribute value distribution with C _j
The state of attribute value distribution for at least one attribute is
Attributes of the category C _j from the distribution of the attribute values of category C _i
Whether the distribution of values can be completely identified (i) or category
The distribution of the attribute values of the distribution of category C _j of attribute values of C _i and one
Partial overlap (ii) or category C
state distribution of the attribute values of _j are included in the distribution of the attribute values of C _i
The process of determining which state (iii) it belongs to
And performing, 1 if (c) each attribute is in the state (i), other
In the case of, define a coefficient of 0 and logically define each attribute.
As variables, and the category C _i and other categories C _j
The set of attributes that can classify
The logical expression of the logical variable with the same attribute
Step of selecting a formula as a set of attributes
If, on the combination of; (d) Category C _i and C _j is the state (i)
, The category C _i and all other categories
A set of attribute sets that can be classified is obtained in step (c).
Performing the process of calculating the logical product of the set of logical expressions
If, on the combination of (e) the category C _i and C _j is the state (i)
To make all categories classifiable with each other.
A set of attribute sets for the logical expression obtained in step (d)
(F) performing the process of obtaining the identification tree from the set of the attribute sets;
In order to select a set of rate attributes, heavy ne of the distribution of the attribute values
Evaluation function based on the state of appearance and the appearance frequency of category C _i
Process to evaluate more and select the most efficient attribute set
(G) setting the attribute set selected in the step (f).
Of the attributes that have the highest rating
-And placed as parent node, and (g-1) parent node
The distribution of attribute values of attributes in a category included
Overlap with the distribution of attribute values of the attribute in other categories
If not, assign the category to the parent node.
Classification is completed by placing it as a child node to
(G-2) When overlapping, use another category and classification
The set of unsuccessful categories is assigned to the child node for the parent node.
(G-3) The category of the child node
The steps (b) to (e) are performed between
The evaluation in the set of attributes selected in step (f)
Excludes attributes used for classification at parent node for child nodes.
The attribute that is the largest of the attributes
Placed as attributes used to classify child nodes as nodes
(G-4) The processing of the above (g-1) to (g-3)
Between the category C _j in Tegori C _i and state (i)
Process until there are no more child nodes to be classified in
Cormorants and steps not completed classification in (h) wherein step (g)
For each child node, the category is divided into an attribute T
The distribution of attribute values for _k is s categories
C _1, ..., if C _i, are overlapping in ... C _s,
The distribution of attribute values for a certain attribute T _k is
Part where Lee C _i does not overlap with all other categories
Min, any category C _i and any other one category
-Overlap, any category C _i and any other 2
Where the categories overlap, ..., any category
-Part where C _i and any other s-1 categories overlap
To divide, for these to split the category attribute T _k
Other than the case of empty set
Performing a process to make Lee, using the probability distribution in the range of attribute values for each (i) attribute, stearyl
New category C _i and category created in top (h)
-Calculate the attribute value probabilities of areas where C _j does not overlap at all
Step a, the attribute stored in the storage device (j) said step (a) to
From the attribute values in the value table, find the frequency of the category,
The frequency of occurrence and the distribution of attribute values obtained in step (i)
Evaluation is performed using an evaluation function that uses the probability of overlap, and
Steps to select the attributes that are most effective for classifying the code
When could not be classified in (k) the step (c) ~ (g)
For the child node, the attribute that is most effective in classifying the child node
Of the category obtained in step (h) using the property
Perform processing to classify in a new category created by splitting
And (l) the knowledge created by the steps (b) to (k).
Create and store data classification flowcharts from different trees
Storing the data in a device, and classifying the data according to the flowchart. 2. (a) Categories C ₁ to C for Classifying Data
C _{i to} C _m are set, and the attribute T of each category is set.
₁ through T _j through T _n aggregates the measurement data for each or calculated
And simulate the result with the corresponding category
And the upper and lower limits for each attribute
Stored in the storage device as an attribute value distribution table representing cloth
That the reference and the step, the attribute value table stored in (b) the storage device
Then, for each attribute, a certain category C _i and another category C _i
-Analysis of the state of overlap of the attribute value distribution with C _j
The state of attribute value distribution for at least one attribute is
Attributes of the category C _j from the distribution of the attribute values of category C _i
Whether the distribution of values can be completely identified (i) or category
The distribution of the attribute values of the distribution of category C _j of attribute values of C _i and one
Partial overlap (ii) or category C
state distribution of the attribute values of _j are included in the distribution of the attribute values of C _i
The process of determining which state (iii) it belongs to
And performing, 1 if (c) each attribute is in the state (i), other
In the case of, define a coefficient of 0 and logically define each attribute.
As variables, and the category C _i and other categories C _j
The set of attributes that can classify
The logical expression of the logical variable with the same attribute
Step of selecting a formula as a set of attributes
And (d) the categories C _i and C _j are combinations of the state (i)
In all categories, all categories
In step (c), a set of attribute sets that can be classified
Step for performing the process of obtaining the logical product of the set of logical expressions obtained in
And (e) the most effective way to create an identification tree from the set of attribute sets.
In order to select a rational set of attributes,
Evaluation function based on the state of appearance and the appearance frequency of category C _i
Process to evaluate more and select the most efficient attribute set
And (f) the set of attributes selected in step (e).
Of the attributes that have the highest rating
-And placed as parent node together with (f-1) parent node
The distribution of attribute values of attributes in a category included
Overlap with the distribution of attribute values of the attribute in other categories
If not, assign the category to the parent node.
Classification is completed by placing it as a child node to
(F-2) If they overlap, use another category and classification
The set of unsuccessful categories is assigned to the child node for the parent node.
(F-3) Category of its child node
The steps (b) to (d) are performed between pairs of
Of the attribute set selected in step (e)
Excludes attributes used for classification at parent node for child nodes.
The attribute that is the largest of the attributes
Placed as attributes used to classify child nodes as nodes
(F-4) The processing of the above (f-1) to (f-3) is performed
Between the category C _j in Tegori C _i and state (i)
Process until there are no more child nodes to be classified in
Cormorants and steps not completed classification (g) In step (f)
For each child node, the category is divided into an attribute T
The distribution of attribute values for _k is s categories
C _1, ..., if C _i, are overlapping in ... C _s,
The distribution of attribute values for a certain attribute T _k is
Part where Lee C _i does not overlap with all other categories
Min, any category C _i and any other one category
-Overlap, any category C _i and any other 2
Where the categories overlap, ..., any category
-Part where C _i and any other s-1 categories overlap
To divide, for these to split the category attribute T _k
Other than the case of empty set
And the step of performing a process to make Lee, using the probability distribution of the range of attribute values for each (h) attribute, stearyl
New category C _i and category created in top (g)
-Calculate the attribute value probabilities of areas where C _j does not overlap at all
Step a, the attribute stored in the storage device by (i) step (a) to
From the attribute values in the value table, find the frequency of the category,
The frequency of occurrence and the distribution of attribute values obtained in step (h)
Evaluation is performed using an evaluation function that uses the probability of overlap, and
Steps to select the attributes that are most effective for classifying the code
When could not be classified in (j) said step (c) ~ (f)
For the child node, the attribute that is most effective in classifying the child node
Of the category obtained in step (g) using the property
Perform processing to classify in a new category created by splitting
Cormorants and step, identification created by (k) the step (b) ~ (j)
Create and store data classification flowcharts from different trees
Storing the data in a device, and classifying the data according to the flowchart. 3. (a) Categories C ₁ to C for Classifying Data
C _{i to} C _m are set, and the attribute T of each category is set.
₁ through T _j through T _n aggregates the measurement data for each or calculated
And simulate the result with the corresponding category
And the upper and lower limits for each attribute
Stored in the storage device as an attribute value distribution table representing cloth
That the reference and the step, the attribute value table stored in (b) the storage device
Then, for each attribute, a certain category C _i and another category C _i
-Analysis of the state of overlap of the attribute value distribution with C _j
The state of attribute value distribution for at least one attribute is
Attributes of the category C _j from the distribution of the attribute values of category C _i
Whether the distribution of values can be completely identified (i) or category
The distribution of the attribute values of the distribution of category C _j of attribute values of C _i and one
Partial overlap (ii) or category C
state distribution of the attribute values of _j are included in the distribution of the attribute values of C _i
The process of determining which state (iii) it belongs to
And performing, 1 if (c) each attribute is in the state (i), other
In the case of, define a coefficient of 0 and logically define each attribute.
As variables, and the category C _i and other categories C _j
The set of attributes that can classify
The logical expression of the logical variable with the same attribute
Step of selecting a formula as a set of attributes
If, on the combination of; (d) Category C _i and C _j is the state (i)
, The category C _i and all other categories
A set of attribute sets that can be classified is obtained in step (c).
Performing the process of calculating the logical product of the set of logical expressions
If, on the combination of (e) the category C _i and C _j is the state (i)
To make all categories classifiable with each other.
A set of attribute sets for the logical expression obtained in step (d)
(F) performing a process of obtaining a logical product of the logical product of the sets; and (f) selecting a set of the attribute sets obtained in the step (e).
Performing a process of selecting one set of arbitrary attributes from the set
When, the set of attributes selected in (g) said step (f)
Inside the parent attribute with any attributes
Placed as a node, and (g-1)
Distribution of attribute values of attributes in other categories
Does not overlap with the distribution of attribute values for that attribute in the
In some cases, the certain category is
(G-2)
When overlapped, it could not be classified into another category
A set of categories as child nodes for the parent node
(G-3) step between the set of its child nodes.
(B) to (e) and perform processing on the child node.
Any of the attributes except those used for classification at the parent node
Attribute is a child node having the child node as a parent node.
(G-4) of the above (g-1) to (g-3)
The process is defined as category C _i and category C _{j in} state (i).
Until there are no more child nodes to classify between
And performing a process, not completed classification in (h) wherein step (g)
For each child node, the category is divided into an attribute T
The distribution of attribute values for k is s categories
C _1, ..., if C _i, are overlapping in ... C _s,
The distribution of attribute values for a certain attribute T _k is
Part where Lee C _i does not overlap with all other categories
Min, any category C _i and any other one category
-Overlap, any category C _i and any other 2
Where the categories overlap, ..., any category
-Part where C _i and any other s-1 categories overlap
To divide, for these to split the category attribute T _k
Other than the case of empty set
Performing a process to make Lee, using the probability distribution in the range of attribute values for each (i) attribute, stearyl
New category C _i and category created in top (h)
-Calculate the attribute value probabilities of areas where C _j does not overlap at all
Step a, the attribute stored in the storage device (j) said step (a) to
From the attribute values in the value table, find the frequency of the category,
The frequency of occurrence and the distribution of attribute values found in step (i)
Evaluation is performed using an evaluation function that uses the probability of overlap, and
Steps to select the attributes that are most effective for classifying the code
When could not be classified in (k) the step (c) ~ (g)
For the child node, the attribute that is most effective in classifying the child node
Of the category obtained in step (h) using the property
Perform processing to classify in a new category created by splitting
And (l) the knowledge created by the steps (b) to (k).
Create and store data classification flowcharts from different trees
Storing the data in a device, and classifying the data according to the flowchart. 4. (a) Categories C ₁ to C for Classifying Data
C _{i to} C _m are set, and the attribute T of each category is set.
₁ through T _j through T _n aggregates the measurement data for each or calculated
And simulate the result with the corresponding category
And the upper and lower limits for each attribute
Stored in the storage device as an attribute value distribution table representing cloth
That the reference and the step, the attribute value table stored in (b) the storage device
Then, for each attribute, a certain category C _i and another category C _i
-Analysis of the state of overlap of the attribute value distribution with C _j
The state of attribute value distribution for at least one attribute is
Attributes of the category C _j from the distribution of the attribute values of category C _i
Whether the distribution of values can be completely identified (i) or category
The distribution of the attribute values of the distribution of category C _j of attribute values of C _i and one
Partial overlap (ii) or category C
state distribution of the attribute values of _j are included in the distribution of the attribute values of C _i
The process of determining which state (iii) it belongs to
And performing, 1 if (c) each attribute is in the state (i), other
In the case of, define a coefficient of 0 and logically define each attribute.
As variables, and the category C _i and other categories C _j
The set of attributes that can classify
The logical expression of the logical variable with the same attribute
Step of selecting a formula as a set of attributes
If, on the combination of; (d) Category C _i and C _j is the state (i)
, The category C _i and all other categories
A set of attribute sets that can be classified is obtained in step (c).
Performing the process of calculating the logical product of the set of logical expressions
If, on the combination of (e) the category C _i and C _j is the state (i)
To make all categories classifiable with each other.
A set of attribute sets for the logical expression obtained in step (d)
(F) performing the process of obtaining the identification tree from the set of the attribute sets;
In order to select a rational set of attributes,
Evaluation function based on the state of appearance and the appearance frequency of category C _i
Process to evaluate more and select the most efficient attribute set
(G) setting the attribute set selected in the step (f).
Inside the parent attribute with any attributes
Placed as a node, and (g-1)
Distribution of attribute values of attributes in other categories
Does not overlap with the distribution of attribute values for that attribute in the
In some cases, the certain category is
(G-2)
When overlapped, it could not be classified into another category
A set of categories as child nodes for the parent node
(G-3) between a set of categories of its child nodes
Performs the processing of steps (b) to (e)
In the attribute set selected in (f), the child node
Arbitrary attributes except those used for classification at the parent node
Is used to classify child nodes whose parent node is the child node.
(G-4) The above (g-1) ~
Ca in the (g-3) processing the category C _i and state (i)
There is no child node to classify with the category C _j
And (h) the classification is not completed in the step (g).
For each child node, the category is divided into an attribute T
The distribution of attribute values for _k is s categories
C _1, ..., if C _i, are overlapping in ... C _s,
The distribution of attribute values for a certain attribute T _k is
Part where Lee C _i does not overlap with all other categories
Min, any category C _i and any other one category
-Overlap, any category C _i and any other 2
Where the categories overlap, ..., any category
-Part where C _i and any other s-1 categories overlap
To divide, for these to split the category attribute T _k
Other than the case of empty set
Performing a process to make Lee, using the probability distribution in the range of attribute values for each (i) attribute, stearyl
New category C _i and category created in top (h)
-Calculate the attribute value probabilities of areas where C _j does not overlap at all
Step a, the attribute stored in the storage device (j) said step (a) to
From the attribute values in the value table, find the frequency of the category,
The frequency of occurrence and the distribution of attribute values obtained in step (i)
Evaluation is performed using an evaluation function that uses the probability of overlap, and
Steps to select the attributes that are most effective for classifying the code
When could not be classified in (k) the step (c) ~ (g)
For the child node, the attribute that is most effective in classifying the child node
Of the category obtained in step (h) using the property
Perform processing to classify in a new category created by splitting
And (l) the knowledge created by the steps (b) to (k).
Create and store data classification flowcharts from different trees
Storing the data in a device, and classifying the data according to the flowchart. 5. (a) Categories C ₁ to C for Classifying Data
C _{i to} C _m are set, and the attribute T of each category is set.
₁ through T _j through T _n aggregates the measurement data for each or calculated
And simulate the result with the corresponding category
And the upper and lower limits for each attribute
Stored in the storage device as an attribute value distribution table representing cloth
That the reference and the step, the attribute value table stored in (b) the storage device
Then, for each attribute, a certain category C _i and another category C _i
-Analysis of the state of overlap of the attribute value distribution with C _j
The state of attribute value distribution for at least one attribute is
Attributes of the category C _j from the distribution of the attribute values of category C _i
Whether the distribution of values can be completely identified (i) or category
The distribution of the attribute values of the distribution of category C _j of attribute values of C _i and one
Partial overlap (ii) or category C
state distribution of the attribute values of _j are included in the distribution of the attribute values of C _i
The process of determining which state (iii) it belongs to
And performing, 1 if (c) each attribute is in the state (i), other
In the case of, define a coefficient of 0 and logically define each attribute.
As variables, and the category C _i and other categories C _j
The set of attributes that can classify
The logical expression of the logical variable with the same attribute
Step of selecting a formula as a set of attributes
If, on the combination of; (d) Category C _i and C _j is the state (i)
, The category C _i and all other categories
A set of attribute sets that can be classified is obtained in step (c).
Performing the process of calculating the logical product of the set of logical expressions
If, on the combination of (e) the category C _i and C _j is the state (i)
To make all categories classifiable with each other.
A set of attribute sets for the logical expression obtained in step (d)
(F) performing a process of obtaining a logical product of the logical product of the sets; and (f) selecting a set of the attribute sets obtained in the step (e).
Performing a process of selecting one set of arbitrary attributes from the set
And (g) from the set of attributes selected in step (f) above.
Place the most efficient attributes on the nodes for creating the identification tree
In order, the distribution of attribute values overlapping state and category C _i
Perform evaluation processing using an evaluation function based on the frequency of appearance
And (h) the set of attributes selected in step (f).
And the group having the highest evaluation in step (g)
Property with its included categories as a parent node.
(H-1) in a certain category included in the parent node
Distribution of attribute values in other categories
If there is no overlap with the attribute value distribution of the attribute,
Place category as child node for the parent node
To complete the classification, and (h-2)
A group of categories that could not be classified with other categories
Is arranged as a child node to the parent node, and (h−
3) step between the set of categories of its child nodes
Perform the processing of (b) to (e) and select in step (f)
The evaluation in step (g) above is
The attribute used for classification at the parent node for the child node
Set the largest attribute among the excluded attributes to the parent node of the child node.
Arranged as attributes used to classify child nodes as nodes
(H-4) The processing of (h-1) to (h-3) is
Between the category C _j in Tegori C _i and state (i)
Process until there are no more child nodes to be classified in
Cormorants and steps not completed classification in (i) wherein step (h)
For each child node, the category is divided into an attribute T
The distribution of attribute values for _k is s categories
C _1, ..., if C _i, are overlapping in ... C _s,
The distribution of attribute values for a certain attribute T _k is
Part where Lee C _i does not overlap with all other categories
Min, any category C _i and any other one category
-Overlap, any category C _i and any other 2
Where the categories overlap, ..., any category
-Part where C _i and any other s-1 categories overlap
To divide, for these to split the category attribute T _k
Other than the case of empty set
Performing a process to make Lee, using the probability distribution in the range of attribute values for each (j) attribute, stearyl
New category C _i and category created in step (i)
-Calculate the attribute value probabilities of areas where C _j does not overlap at all
Step a, the attribute stored in the storage device (k) the step (a) to
From the attribute values in the value table, find the frequency of the category,
Of the frequency of occurrence and the distribution of attribute values found in step (j)
Evaluation is performed using an evaluation function that uses the probability of overlap, and
Steps to select the attributes that are most effective for classifying the code
When could not be classified in (l) said step (c) ~ (h)
For the child node, the attribute that is most effective in classifying the child node
Min category determined in step (i) with sex
Perform processing to classify in a new category created by splitting
Cormorants and step, identification created by (m) the step (b) ~ (l)
Create and store data classification flowcharts from different trees
Storing the data in a device, and classifying the data according to the flowchart. 6. (a) Categories C ₁ to C for Classifying Data
C _{i to} C _m are set, and the attribute T of each category is set.
₁ through T _j through T _n aggregates the measurement data for each or calculated
And simulate the result with the corresponding category
And the upper and lower limits for each attribute
Stored in the storage device as an attribute value distribution table representing cloth
That the reference and the step, the attribute value table stored in (b) the storage device
Then, for each attribute, a certain category C _i and another category C _i
-Analysis of the state of overlap of the attribute value distribution with C _j
The state of attribute value distribution for at least one attribute is
Attributes of the category C _j from the distribution of the attribute values of category C _i
Whether the distribution of values can be completely identified (i) or category
The distribution of the attribute values of the distribution of category C _j of attribute values of C _i and one
Partial overlap (ii) or category C
state distribution of the attribute values of _j are included in the distribution of the attribute values of C _i
The process of determining which state (iii) it belongs to
And performing, 1 if (c) each attribute is in the state (i), other
In the case of, define a coefficient of 0 and logically define each attribute.
As variables, and the category C _i and other categories C _j
The set of attributes that can classify
The logical expression of the logical variable with the same attribute
Step of selecting a formula as a set of attributes
If, on the combination of; (d) Category C _i and C _j is the state (i)
, The category C _i and all other categories
A set of attribute sets that can be classified is obtained in step (c).
Performing the process of calculating the logical product of the set of logical expressions
If, on the combination of (e) the category C _i and C _j is the state (i)
To make all categories classifiable with each other.
A set of set of eyes of the attribute step (d) in the obtained logical formula
(F) performing the process of obtaining the identification tree from the set of the attribute sets;
In order to select a rational set of attributes,
Evaluation function based on the state of appearance and the appearance frequency of category C _i
Process to evaluate more and select the most efficient attribute set
(G) setting the attribute set selected in the step (f).
Of the attributes that have the highest rating
-And placed as parent node, and (g-1) parent node
The distribution of attribute values of attributes in a category included
Overlap with the distribution of attribute values of the attribute in other categories
If not, assign the category to the parent node.
Classification is completed by placing it as a child node to
(G-2) When overlapping, use another category and classification
The set of unsuccessful categories is assigned to the child node for the parent node.
(G-3) The category of the child node
The steps (b) to (e) are performed between
The evaluation in the set of attributes selected in step (f)
Excludes attributes used for classification at parent node for child nodes.
The attribute that is the largest of the attributes
Placed as attributes used to classify child nodes as nodes
(G-4) The processing of the above (g-1) to (g-3)
Between the category C _j in Tegori C _i and state (i)
Process until there are no more child nodes to be classified in
Cormorants and steps not completed classification in (h) wherein step (g)
For each child node, the category is divided into an attribute T
The distribution of attribute values for _k is s categories
C _1, ..., if C _i, are overlapping in ... C _s,
The distribution of attribute values for a certain attribute T _k is
Part where Lee C _i does not overlap with all other categories
Min, any category C _i and any other one category
-Overlap, any category C _i and any other 2
Where the categories overlap, ..., any category
-Part where C _i and any other s-1 categories overlap
To divide, for these to split the category attribute T _k
Other than the case of empty set
Performing a process of making a lie, and (i) Classification failed in the above steps (c) to (g)
For the child node, use any attribute
New made by the division of the obtained category in (h)
Performing a process of classifying by category; and (j) the knowledge created in steps (b) to (i).
Create and store data classification flowcharts from different trees
Storing the data in a device, and classifying the data according to the flowchart.