JP5817600B2 - Factor analysis apparatus and factor analysis method - Google Patents

Description

  The present invention relates to a factor analysis apparatus and a factor analysis method for analyzing and estimating an element that is a factor of an event that has occurred in a system composed of a large number of elements.

  In general, in a system consisting of many elements such as vehicles and production equipment, the output results of the system, for example, the running state of the vehicle and the product produced, have an influence on the running state and the quality of the produced product. It is not easy to identify, analyze, and estimate the affecting elements. Further, if the influence on the result of the output is due to a plurality of elements, it becomes more difficult to analyze and estimate the combination of these elements. For example, when trying to estimate an element that affects a specific operation of the vehicle, it is analyzed which combination of elements such as an engine, a driver operation, a vehicle environment, etc. as an element of the vehicle has influenced the specific operation. It is not easy. Moreover, such analysis is often performed when the output result of the system is inconvenient, but if the influential elements are not sufficiently narrowed down, it is also possible to try many combinations of multiple elements. The time, labor, and cost required for analysis cannot be ignored. Moreover, there is a possibility that even a satisfactory analysis result for the inconvenience cannot be obtained. Therefore, conventionally, for example, the device described in Patent Document 1 is used to analyze and estimate factors that have affected the output result of the system.

  The apparatus described in Patent Document 1 is an apparatus that estimates an abnormality in a production facility. The apparatus receives data from a history DB (database) that stores product history information and a quality DB that stores quality information. An analysis data creation unit that creates and analyzes data by searching and processing, and a data analysis unit that analyzes and estimates abnormal equipment using the analysis data are provided. The data analysis unit includes an abnormal event classification unit that classifies analysis data for each manufacturing facility using decision tree analysis, and an abnormality that estimates the manufacturing facility that causes the product quality degradation based on the classification result. An equipment evaluation section is provided. As described above, the apparatus described in Patent Document 1 estimates a manufacturing facility that causes a reduction in product quality by analysis using decision tree analysis.

JP 2006-319220 A

  By the way, since the apparatus described in Patent Document 1 is intended to analyze production facilities, the quality of processing by each production facility is often adopted as a classification condition when performing factor analysis using a decision tree. In other words, in the analysis by the same device, the combination of equipment that causes the deterioration of product quality is estimated based on the quality of processing by the production equipment in which interference between equipment is suppressed. There is a tendency to be relatively narrowed down.

  On the other hand, in a factor analysis using a decision tree for a system where it is difficult to suppress mutual interference between elements to be analyzed (for example, speed, accelerator operation, etc.) such as a vehicle, interference between elements makes classification conditions ambiguous. There are many cases. In addition, in many cases, each element can take a plurality of states with respect to one result, and this also contributes to obscuring the classification conditions in the decision tree. Moreover, such ambiguity of classification conditions also tends to increase the factors that separate the results in the factor analysis using a decision tree. Further, when trying to increase the analysis accuracy of the factor analysis using the decision tree, the factors are naturally subdivided, so that the factors for separating the results increase even for the above-described production facilities. Such an increase in the factor of separation also increases the factor itself included in the analysis result, so that the time and labor required for system verification performed based on the analysis result cannot be ignored.

  Such a problem is a common problem in factor analysis for so-called multivariable systems such as not only a factor analysis device mounted on a vehicle but also a manufacturing facility composed of a large number of components.

  The present invention has been made in view of the above situation, and its purpose is to make it easier for the elements that have an influence on the output result to a system in which a large number of elements have an influence on the output result. Moreover, it is an object of the present invention to provide a factor analysis apparatus and a factor analysis method that can analyze and estimate with high accuracy.

Hereinafter, means for solving the above-described problems and the effects thereof will be described.
In order to solve the above-mentioned problem, the invention according to claim 1 is a factor analysis device, in which a plurality of samples including a plurality of variables are stored with one of a plurality of classification classes assigned to the samples. A plurality of samples for each variable by analysis using a decision tree having the storage unit and the classification class as an objective variable for classifying a plurality of samples, and the sample as an explanatory variable classified by the classification class . When separately classifying, the plurality of samples are sequentially classified into samples for each variable while sequentially selecting the variables so that the variation of the selected specific classification class among the plurality of classification classes is reduced. An analysis unit that obtains a variable group consisting of a plurality of variables that affect the classification, and a plurality of variables included in the variable group Includes a variable narrowing portion to narrow based on the ratio of the specific classification classes configured in a plurality of samples sorted by each variable, the analysis unit, the decision tree, one that includes all of the samples A root node and a plurality of leaf nodes composed of one or a plurality of samples that cannot be classified as a result of sequential classification are obtained, and a plurality of variables included in the variable group are obtained from the one root node. It is composed of variables included in a path up to one of the plurality of leaf nodes, and the variable narrowing-down unit narrows down a plurality of variables included in the variable group. The ratio of a specific classification class associated with a plurality of samples classified by a variable corresponds to one or a plurality of samples in the path and classified on the leaf node side When greater than the proportion of the particular classification classes kicked, and summarized in that done by recursively executed to exclude the variable from the variable group.

In order to solve the above-mentioned problem, the invention according to claim 5 is a factor analysis method, wherein the storage unit converts a plurality of samples including a plurality of variables into one classification class of a plurality of classification classes separately from the samples. And a storage step of storing in the storage unit so that is assigned, and the analysis unit determines that the classification class is an objective variable for classifying a plurality of samples, and the sample is an explanatory variable classified by the classification class When the plurality of samples are classified according to each variable by analysis using a tree, the plurality of samples are selected in order so that the variation of the selected specific classification class among the plurality of classification classes is reduced. samples and analyzing step of obtaining a variable group composed of a plurality of variables affecting the classification by sequentially classifying the variable another sample of the variable narrowing portion, front A plurality of variables in the variable group, with each of the set in the plurality of samples classified by the variable Filter based on the percentage of specific classification class variable narrowing process, in the analyzing step, the A decision tree is obtained as having a plurality of leaf nodes including one root node including all samples and one or a plurality of samples that cannot be classified as a result of sequential classification, and is included in the variable group plural
A variable is composed of a variable included in a path from the one root node to one of the plurality of leaf nodes. In the variable narrowing step, a plurality of variables included in the variable group are narrowed down. The ratio of a specific classification class associated with a plurality of samples classified by a plurality of variables included in the variable group is associated with one or a plurality of samples classified on the leaf node side in the path. When the ratio is larger than the ratio of the specific classification class, the gist is to recursively execute the removal of the variable from the variable group .

  According to such a configuration or method, a variable group obtained as a plurality of variables that reduces variation of a specific classification class, that is, favorably divides a specific classification class, is included in the variable group. For each of a plurality of variables, the filtering is performed based on the ratio of a specific classification class assigned to one or more samples classified by the variable. In other words, the variables included in the variable group are further narrowed down, and the number of combinations of variables based on the one or more variables thus narrowed is reduced, and the time and effort required for verification regarding a specific classification class performed for each of those combinations. Etc. will be reduced. Further, if the variables included in the variable group are narrowed down so that appropriate analysis and estimation can be performed even with the variables further narrowed down from the variable group, the estimation accuracy can be improved as compared with the conventional case. In other words, for a system in which a large number of variables (elements) affect a specific classification class (output result), this factor analysis device makes it easier and more accurate for variables that affect a specific classification class. It can be analyzed and estimated.

Further , according to such a configuration or method, since the variable group is obtained based on the decision tree algorithm, the variables included in the variable group obtained in the factor analysis apparatus to which the decision tree algorithm is applied are appropriately narrowed down. Will be able to.

Furthermore, according to such a configuration or method, when the ratio of the specific classification class included in the sample on the root node side is larger than the ratio of the specific classification class included in the sample on the leaf node side, the root node The side variable is excluded from the variable group. As a result, variables with a small contribution to the classification of multiple samples are excluded from the variables included in the variable group. Therefore, it is preferable to classify a plurality of samples according to a specific classification class even as a variable group with a narrowed variable. Can be maintained.

The invention according to claim 2 is the factor analysis device according to claim 1 , wherein when the plurality of leaf nodes are generated by one or a plurality of samples classified by the variable, the variable narrowing-down unit The gist is to use a ratio of a specific classification class in a leaf node having a larger ratio of a specific classification class associated with a sample of each leaf node of the leaf node for narrowing down the variable group.

The invention according to claim 6 is the factor analysis method according to claim 5 , wherein, in the variable narrowing-down step, when a plurality of leaf nodes are generated by one or a plurality of samples classified by the variable, The gist is to use a ratio of a specific classification class in a leaf node having a larger ratio of a specific classification class associated with a sample of each leaf node of the leaf node for narrowing down the variable group.

  According to such a configuration or method, a leaf node suitable for narrowing down a variable is selected, and a ratio of a specific classification class included in the sample of the selected leaf node is a plurality of variables included in the variable group. It will be used for narrowing down. That is, it becomes possible to narrow down the variables included in the variable group.

The invention according to claim 3 is the factor analysis device according to claim 1 or 2 , wherein the plurality of classification classes includes two classes, a normal class indicating a normal case and an exception class indicating an exception case, The gist is that the specific classification class is the exception class.

The invention according to claim 7 is the factor analysis method according to claim 5 or 6 , wherein the plurality of classification classes includes two classes, a normal class indicating a normal case and an exception class indicating an exception case, The gist is that the specific classification class is the exception class.

  According to such a configuration or method, it is possible to obtain a variable that seems to have an influence on the occurrence of an exception class from a plurality of samples. If an exception case is an unfavorable case for the system, if the verification of the case that is unfavorable for the system can be performed appropriately by acquiring variables that affect the exception case, the exception case that occurs in the system It is expected to be able to respond.

According to a fourth aspect of the present invention, in the factor analysis device according to the third aspect , the variable is information indicating a vehicle state or a vehicle environment, and the vehicle is a plurality of vehicle states obtained from one vehicle at the same time. Or the vehicle environment and the plurality of samples are samples in which at least one of time and vehicle is different from each other, a specific vehicle state is set as the exception class, and the variable narrowing unit includes the specific vehicle state The gist is to narrow down a plurality of vehicle states or vehicle environments included in the variable group based on the set ratio.

The invention according to claim 8 is the factor analysis method according to claim 7 , wherein information indicating a vehicle state is used as the variable, and a plurality of vehicle states or vehicles obtained from one vehicle at the same time as the sample. In addition to the environment, the plurality of samples are samples in which at least one of time and vehicle is different from each other, a specific vehicle state is set as the exception class, and the specific vehicle state is set in the variable narrowing step. The gist is to narrow down a plurality of vehicle states or vehicle environments included in the variable group based on the ratio.

  In vehicles, there is a correlation between vehicle conditions such as speed and acceleration, there is a correlation between vehicle conditions such as steering operation and road conditions and the vehicle environment, and between each element such as each vehicle condition and vehicle environment. In many cases, there is a correlation, that is, the independence between elements is low and there is mutual interference. It is not easy to estimate a combination of elements causing a specific target vehicle state, a so-called exceptional case, from a plurality of such interference-prone elements. In addition, it is not easy to select an element that suitably separates a target specific vehicle state from elements having a lot of interference, and the number of elements that are selected as having an influence on a specific vehicle state by analysis is large. Tend to be. Therefore, when verifying a specific vehicle state based on the selected combination of elements, many combinations of these elements are verified, and the time and effort required for verification cannot be ignored.

  On the other hand, according to the above configuration or method, the vehicle state or the vehicle environment obtained as having an influence on the specific vehicle state from the vehicle state or the vehicle environment acquired from the vehicle is the vehicle state or the vehicle environment. Based on the percentage of exception classes assigned to one or more samples classified by. That is, even in a system such as a vehicle that tends to have a large number of elements selected to affect a specific vehicle state, the number of vehicle states or vehicle environments that are selected can be narrowed down. As a result, it becomes possible to verify the vehicle state in which the exception class is generated based on the narrowed combination of the vehicle state or the vehicle environment, so the number of verifications corresponding to the number of combinations of the vehicle state or the vehicle environment is reduced. As a result, the time and labor required to verify the exception class are reduced.

The block diagram which shows the system configuration | structure about one Embodiment which actualized the system provided with the factor analysis apparatus concerning this invention. The block diagram which shows the structure about the vehicle which acquires the sample made into the analysis object by the factor analysis apparatus of the embodiment. The figure which shows an example of the content of the database which stores the sample made into the analysis object by the factor analysis apparatus of the embodiment. The schematic diagram which shows typically the decision tree which the factor analysis apparatus of the embodiment created by factor analysis. The figure which shows the exception rate computed from the decision tree by the factor analysis apparatus of the embodiment in a table format. The schematic diagram which shows the example which the factor analysis apparatus of the embodiment selects a node from a decision tree. The schematic diagram which shows the example which the factor analysis apparatus of the embodiment narrows down a node based on an exception rate.

DESCRIPTION OF EMBODIMENTS Hereinafter, an embodiment embodying a system including a factor analysis device according to the present invention will be described with reference to the drawings.
The system according to the present embodiment is based on samples collected from a plurality of vehicles 20 in the information processing center 10, and vehicle statuses that are relevant to specific exceptional events that occur in the vehicle 20 in the information processing center 10. This is a system for estimating the vehicle environment.

First, the system of this embodiment is demonstrated with reference to FIG.1 and FIG.2.
As illustrated in FIG. 1, a plurality of vehicles 20, specifically, four vehicles 20 a to 20 d are wirelessly connected to the information processing center 10. Each vehicle 20a-20d consists of the same vehicle type, and can collect the data of the vehicle state and vehicle environment which each consist of the same type. The vehicle state relates to the state of the vehicle itself, and examples of the vehicle state include speed, acceleration, engine speed, steering operation, lighting / extinguishing of lights, and opening / closing of a door. The vehicle environment is related to the surrounding environment of the vehicle, and examples of the vehicle environment include information on the road such as time, position, weather, brightness (illuminance), road shape, and information on the user.

  As shown in FIG. 2, the vehicle 20 includes a vehicle control device 21 that performs various controls on the vehicle, a storage device 22 that is connected to the vehicle control device 21 so as to be able to input and output data, and a data input and output that is possible. A communication unit 23 connected to the vehicle control device 21 is provided.

  The vehicle control device 21 is connected to a plurality of information output devices that output detected information as data. As an information output device, a clock 30, a GPS 31 as a position sensor, an illuminometer 32 as an illuminance sensor, a speedometer 33 as a speed sensor, a first accelerometer 34 and a second accelerometer as acceleration sensors 35 is connected to the vehicle control device 21.

  The timepiece 30 outputs the current time Ta to the vehicle control device 21. The GPS 31 is a so-called global positioning system (GPS), and outputs a current position Pa detected based on a received GPS satellite signal to the vehicle control device 21. The illuminometer 32 outputs the illuminance value La detected in the traveling environment of the vehicle 20 to the vehicle control device 21. The speedometer 33 outputs the speed Va detected from the vehicle 20 to the vehicle control device 21. The first accelerometer 34 outputs the first acceleration Aa detected with respect to the traveling direction of the vehicle 20 to the vehicle control device 21. The second accelerometer 35 outputs the acceleration Ab detected in the direction intersecting the road surface to the vehicle control device 21.

  The communication unit 23 is a device that enables wireless communication from the vehicle control device 21 to the information processing center 10, and converts the transmission data input from the vehicle control device 21 into a signal for wireless communication, thereby converting the information processing center. 10 is transmitted. Note that the communication unit 23 may receive a signal for wireless communication from the information processing center 10 and output the signal to the vehicle control device 21.

  The storage device 22 is composed of a nonvolatile storage device such as a flash memory or an HDD (Hard Disk Drive), and can be read and written from the vehicle control device 21.

  The vehicle control device 21 includes a microcomputer including a CPU that executes various arithmetic processes, a ROM that stores various control programs, a RAM that is used as a work area for data storage and program execution, an input / output interface, a memory, and the like. It is structured in the center. The vehicle control device 21 stores various programs and various parameters required for various controls in advance. In the present embodiment, the vehicle control device 21 is provided with a program for collecting and recording information output from various information output devices. This program acquires a plurality of pieces of information output from a plurality of information output devices at the same time using sampling conditions such as time and position, and samples the obtained pieces of information as one objective variable, for example, sample D1. Is stored in the storage device 22. The acquisition of the sample and the storage in the storage device 22 are executed each time a sampling condition such as time and position is satisfied, so that at least samples with different times are stored in the storage device 22. Furthermore, in the present embodiment, a program for transmitting the sample D1 stored in the storage device 22 to the information processing center 10 based on a predetermined condition is also provided. The predetermined condition is determined based on the transmission cycle, the communication status, the number of stored samples, and the like.

  Thus, when the sampling condition is satisfied, the vehicle control device 21 acquires the current time Ta, the current position Pa, the illuminance value La, the speed Va, the first acceleration Aa, the second acceleration Ab, etc. The sample is stored in the storage device 22 as the sample D1. Similarly, the vehicle control device 21 causes the storage device 22 to store the samples D2, D3,... Having values acquired with the establishment of the sampling condition. In addition, the vehicle control device 21 transmits one or more samples (D1, D2, D3,...) Stored in the storage device 22 to the information processing center 10 via wireless communication.

  As shown in FIG. 1, an information processing center 10 includes an information processing unit 11 that processes various types of information, a storage device 12 as a storage unit that is connected to the information processing unit 11 so that data can be input and output, and data And a communication unit 13 connected to the information processing unit 11 so as to be able to input and output these.

  The communication unit 13 is a device for receiving a radio signal transmitted from the vehicle control device 21, converts the received radio signal into data, and outputs the data to the information processing unit 11. That is, the information processing unit 11 can acquire the samples D1, D2, D3,... Transmitted from the vehicle 20 via the communication unit 13. Note that the communication unit 13 may convert the data input from the information processing unit 11 into a signal for wireless communication and transmit it to the vehicle control device 21.

  The storage device 12 is composed of a nonvolatile storage device such as an HDD (Hard Disk Drive), and can be read and written from the information processing unit 11. For example, the information processing section 11 can write each sample D1, D2, D3,... Acquired from the vehicle 20 to the storage device 12 or read it from the storage device 12, respectively.

  The storage device 12 is provided with a database 121 in which the samples D1, D2, D3,... Acquired from the vehicle 20 can be registered. As shown in FIG. 3, the database 121 includes 210 samples D1 to D210 composed of a plurality of samples D1, D2, D3,... Acquired from the vehicle 20 (20a to 20d), and each sample D1 to D210. Each is assigned a classification class X1 as an explanatory variable. In each sample D1 to D210, values dn, m of 30 variables S1 to S30 corresponding to each information output device of the vehicle 20 (where n is an integer from 1 to 210 and m is an integer from 1 to 30). include. Each variable S1 to S30 corresponds to any one of a plurality of information output devices provided in the vehicle 20. For example, in the present embodiment, the variable S1 corresponds to the clock 30 of the vehicle 20, the variable S2 corresponds to the GPS 31 of the vehicle 20, and the variable S3 corresponds to the illuminometer 32 of the vehicle 20. The variable S4 corresponds to the speedometer 33 of the vehicle 20, and the variable S5 corresponds to the first accelerometer 34 of the vehicle 20. That is, in the sample D1 of the database 121, the current time Ta is stored as the value d1,1 of the variable S1, the current position Pa is stored as the value d2,1 corresponding to the variable S2, and the value d3 corresponding to the variable S3. , 1 is stored as the illuminance value La. In addition, in the sample D1 of the database 121, the speed Va is stored as the value d4, 1 corresponding to the variable S4, and the first acceleration Aa is stored as the value d5, 1 corresponding to the variable S5.

  The classification class X1 is an example that occurs in the vehicle 20, and sets the presence or absence of correspondence to the case that is the object of the factor analysis in each sample D1 to D210. In the classification class X1, “normal class A” and “exception class B” are provided as two classification classes corresponding to whether or not a target case has occurred in the vehicle 20, respectively. The “exception class B” indicates a specific vehicle state as an exceptional case that occurs in the vehicle 20, and examples of the specific vehicle state include sudden acceleration, sudden deceleration, and vertical vibration. In the present embodiment, an exceptional case that occurs in the vehicle 20 is used for the purpose of factor analysis. Therefore, an exception class B as a specific classification class is set in the sample when the exceptional case occurs in the vehicle 20. On the other hand, “normal class A” is set for the sample when no exception has occurred in the vehicle 20. That is, the sample for which “normal class A” is set is a sample that does not correspond to an exception case, while the sample for which “exception class B” is set is a sample that corresponds to an exception case. For example, each sample D1, D2, Dn, etc. registered in the database 121 has “exception class B” in the classification class X1, while each sample D3, D210 etc. has “normal class A” in the classification class X1. Is set. The classification class X1 may be automatically given based on a predetermined condition, or may be given by an analyst. Further, the classification class X1 may be given by the vehicle 20 at the time of sample acquisition, or may be given when the sample is registered in the database 121 or after it is registered.

  That is, in the storage device 12, a plurality of samples D1 to D210 including a plurality of variables S1 to S30 are one of the plurality of classification classes other than the samples D1 to D210, “normal class A” or “exception. “Class B” is assigned and stored (storage step).

  The information processing unit 11 includes a microcomputer that includes a CPU that executes various arithmetic processes, a ROM that stores various control programs, a RAM that is used as a work area for data storage and program execution, an input / output interface, a memory, and the like. It is structured in the center. The information processing unit 11 stores various programs and various parameters required for information processing related to various vehicle diagnoses in advance. In the present embodiment, the information processing unit 11 is provided with a decision tree analysis program and parameters for performing factor analysis by decision tree analysis, and a narrowing program and parameters for further narrowing down the variables analyzed by the decision tree analysis. . The information processing unit 11 is provided with a decision tree creating unit 111 based on execution processing of the decision tree analysis program, and a variable narrowing unit 113 based on execution processing of the narrowing program. Note that the decision tree analysis program and the narrow-down program may be always executed, or may be executed as necessary.

  The decision tree creation unit 111 generates a decision tree 112 from the samples stored in the database 121 based on decision tree analysis. The decision tree 112 is a tree diagram based on a tree shape model generated by decision tree analysis which is a kind of information processing method.

  As shown in FIG. 3, the decision tree creating unit 111 uses a plurality of samples of cases targeted for factor analysis, for example, “exception class B” which is one specific classification class selected from two classification classes. The decision tree 112 is generated by repeating suitable classification from D1 to D210 (analysis step). The suitable classification means that the decision tree creating unit 111 classifies the plurality of samples D1 to D210 with a variable (S1 or the like) selected so that the variation of the “exception class B” becomes small. As shown in FIG. 4, in the decision tree 112, the above “classification” is illustrated as “branch”.

  As shown in FIG. 4, the decision tree 112 includes one “root node N1” where classification (branch) is started, and four “intermediate nodes N2 to N5” that are branch points other than “root node N1”. And 6 “leaf nodes LP1 to LP3, LN1 to LN3” which are end points that are not branched any more. That is, a plurality of branches occur from one “root node N1”, and one “intermediate node N2 to N5” is provided at each of the branches. Further, “leaf nodes LP1 to LP3, LN1 to LN3” are provided as end points at which no branching occurs at such branch destinations. Accordingly, a plurality of “intermediate nodes N2 to N5” and a plurality of “leaf nodes LP1 to LP3, LN1 to LN3” are provided for one “root node N1”. In addition, when the “leaf node” is provided in one branch from the “root node”, the “intermediate node” is not provided. From such a decision tree 112, a “route” as a route from each “leaf node LP1 to LP3, LN1 to LN3” to “root node N1” is obtained for each “leaf node LP1 to LP3, LN1 to LN3”. It is done. The “root” thus obtained includes the “leaf node” that is the end point of the “route”, the “intermediate node” included in the route from the “leaf node” to the “root node”, and the “root node”. Each piece of information is included.

  In this embodiment, the “root node N1” includes all samples D1 to D210 as information. All the samples D1 to D210 are divided into two “intermediate node N2” and “intermediate node N3” by classification. That is, each of “intermediate node N2” and “intermediate node N3” includes some of the samples D1 to D210. In each of “intermediate node N2” and “intermediate node N3”, samples included in each are classified into two “intermediate nodes” or “leaf nodes”. Then, when such classification is repeated and the number of samples included in the classified node becomes one, or when the sample included in the classified node becomes unclassifiable, the nodes are referred to as “leaf nodes”. At the same time, the classification ends.

  Here, with reference to FIG. 4, the decision tree 112 created by the decision tree creation unit 111 based on the samples D1 to D210 set in the database 121 will be described. Note that the decision tree creation unit 111 sets the case targeted for the factor analysis as “exception class B” which is one of a plurality of classification classes. The 210 samples D1 to D210 included in the “root node N1” include 60 samples in which “normal class A” is set and 150 samples in which “exception class B” is set. And Further, each sample D1 to D210 includes values dn, 1 to dn, 30 corresponding to 30 variables S1 to S30. For convenience of explanation, the samples D1 to D210 are represented by five variables S1 to S5. A case will be described in which only the values dn, 1 to dn, 5 are used for classification, that is, used for classification.

  The decision tree creation unit 111 starts classification (separation) from “root node N1” having 210 samples D1 to D210. Note that the decision tree creation unit 111 applies a decision tree analysis algorithm to “root node N1”, so that “exception class B” can be appropriately classified (isolated) from “root node N1”. It is determined that it is preferable to select S1 as a variable and to set the threshold value C1 as the threshold value for evaluating the values d1 and m of the variable S1. Then, the decision tree creation unit 111 sets “intermediate node N2” including samples whose values d1 and m are larger than the threshold C1 (d1> C1) from the “root node N1”, and values d1 and m are equal to or less than the threshold C1 (d1 ≦ C1) “intermediate node N3” including the sample is classified (separated). Note that d1 is abbreviated as the subscript “m” of the values d1 and m. As a result, the “intermediate node N2” includes 100 samples. Among those samples, there are 50 samples in which “normal class A” is set and samples in which “exception class B” is set. 50 pieces shall be included. On the other hand, the “intermediate node N3” includes 110 samples, including 10 samples for which “normal class A” is set and 100 samples for which “exception class B” is set. It shall be included.

  Subsequently, the decision tree creation unit 111 applies the decision tree analysis algorithm to the “intermediate node N2”, thereby appropriately classifying the “exception class B” from the “intermediate node N2”. It is determined that it is preferable to select the threshold value C2 as a threshold value for selecting the variable and evaluating the values d2 and m2 of the variable S2. Then, the decision tree creating unit 111 determines from “intermediate node N2” “leaf node LP1” that includes samples whose values d2 and m are greater than the threshold C2 (d2> C2), and values d2 and m that are equal to or less than the threshold C2 (d2 ≦ C2) is classified as “leaf node LN1” including the sample. As a result, the “leaf node LP1” includes 60 samples, and in those samples, 50 samples for which “normal class A” is set and samples for which “exception class B” is set are included. 10 shall be included. On the other hand, the “leaf node LN1” includes 40 samples. Among these samples, there are 0 samples for which “normal class A” is set and 40 samples for which “exception class B” is set. It shall be included.

  In addition, the decision tree creation unit 111 applies the decision tree analysis algorithm to the “intermediate node N3”, thereby appropriately classifying the variable S3 from the “intermediate node N3” to the “exception class B”. It is determined that it is preferable to set the threshold value C3 as the threshold value for evaluating the values d3 and m of the variable S3. Then, the decision tree creating unit 111 determines, from “intermediate node N3”, “intermediate node N4” including samples whose values d3 and m are greater than the threshold C3 (d3> C3), and values d3 and m are equal to or less than the threshold C3 (d3 ≦ C3) is classified as “intermediate node N5” including the sample. As a result, 68 samples are included in the “intermediate node N4”, and 8 samples in which “normal class A” is set and 8 samples in which “exception class B” is set are included in these samples. 60 pieces shall be included. On the other hand, the “intermediate node N5” includes 42 samples, including two samples for which “normal class A” is set and 40 samples for which “exception class B” is set. It shall be included.

  Furthermore, the decision tree creation unit 111 applies the decision tree analysis algorithm to the “intermediate node N4”, thereby appropriately classifying the variable S4 from the “intermediate node N4” to the “exception class B”. And the threshold value for evaluating the value d4, m of the variable S4 is preferably set to the threshold value C4. Then, the decision tree creation unit 111 sets “leaf node LP2” including samples from the “intermediate node N4” where the values d4 and m are larger than the threshold C4 (d4> C4), and the values d4 and m are equal to or less than the threshold C4 (d4 ≦ C4) is classified as “leaf node LN2” including the sample. As a result, the “leaf node LP2” includes 8 samples. Among these samples, there are 7 samples for which “normal class A” is set and samples for which “exception class B” is set. One shall be included. On the other hand, the “leaf node LN2” includes 60 samples, of which one sample is set with “normal class A” and 59 samples are set with “exception class B”. It shall be included.

  Subsequently, the decision tree creation unit 111 applies the decision tree analysis algorithm to the “intermediate node N5”, thereby appropriately classifying the variable S5 from the “intermediate node N5” to the “exception class B”. It is determined that it is preferable to select the threshold value C5 as the threshold value for selecting the variable and evaluating the value d5, m of the variable S5. Then, the decision tree creation unit 111 has “leaf node LP3” including samples from “intermediate node N5” whose value d5 is larger than the threshold C5 (d5> C5), and the value d5 is equal to or less than the threshold C5 (d5 ≦ C5). The “leaf node LN3” including the sample is classified. Thereby, the “leaf node LP3” includes two samples, and these samples include two samples for which “normal class A” is set, and “exception class B” is set. Assume that 0 samples are included. On the other hand, the “leaf node LN3” includes 40 samples. Among these samples, there are 0 samples for which “normal class A” is set and 40 samples for which “exception class B” is set. It shall be included.

  In this way, the decision tree creating unit 111 generates the decision tree 112. From this decision tree 112, as a route that can suitably classify “exception class B”, a route composed of “intermediate node N2” and “root node N1”, “intermediate node N4” and “intermediate node” A route consisting of “N3” and “root node N1” and a route consisting of “intermediate node N5”, “intermediate node N3” and “root node N1” are obtained. That is, the first variable group as the variable group including the variable S2 and the variable S1 is obtained from the route including the “intermediate node N2” and the “root node N1”. Further, a second variable group as a variable group including the variable S4, the variable S3, and the variable S1 is obtained from the route including the “intermediate node N4”, “intermediate node N3”, and “root node N1”. Further, a third variable group including a variable S5, a variable S3, and a variable S1 is obtained from the route including the “intermediate node N5”, “intermediate node N3”, and “root node N1”.

  Conventionally, factor analysis and estimation are performed for each variable group based on a combination obtained from the variable group. That is, three combinations of “variable S2 and variable S1,” “variable S2,” and “variable S1” are obtained from the first variable group. Further, from the second variable group, “variable S4 and variable S3 and variable S1”, “variable S4 and variable S3”, “variable S4 and variable S1”, “variable S3 and variable S1”, “variable S4”, Seven combinations of “variable S3” and “variable S1” are obtained. Further, from the third variable group, “variable S5 and variable S3 and variable S1”, “variable S5 and variable S3”, “variable S5 and variable S1”, “variable S3 and variable S1”, “variable S5”, Seven combinations of “variable S3” and “variable S1” are obtained. That is, 17 types of variable combinations can be obtained from the decision tree 112, and 13 types can be obtained by excluding duplication.

  The variable refining unit 113 affects the “exception class B” from the decision tree 112 generated by the decision tree creation unit 111 based on the ratio of the “exception class B” set in the sample included in each node. Narrow down the variables (factors) to be affected.

  First, as illustrated in FIG. 5, the variable narrowing unit 113 calculates an exception rate for each node provided in the decision tree 112. The exception rate is a ratio of “exception class B” set in the classification class X1 of a plurality of samples included in the node, and is expressed as “number of exception classes B / (number of normal classes A + number of exception classes B”). ) ”. That is, the exception rate of “root node N1” is calculated as “0.71” (= 150 / (60 + 150)) because “normal class A” is 60 and “exception class B” is 150. The Also, the exception rate of “intermediate node N2” is calculated as “0.50” (= 50 / (50 + 50)) because “normal class A” is 50 and “exception class B” is 50. The Further, the exception rate of “leaf node LP1” is calculated as “0.17” (= 10 / (50 + 10)) because “normal class A” is 50 and “exception class B” is 10. On the other hand, the exception rate of “leaf node LN1” is calculated as “1.00” (= 40 / (0 + 40)) because “normal class A” is 0 and “exception class B” is 40. Is done. Also, the exception rate of “intermediate node N3” is calculated as “0.91” (= 100 / (10 + 100)) because “normal class A” is 10 and “exception class B” is 100. In addition, the exception rate of “intermediate node N4” is calculated as “0.88” (= 60 / (8 + 60)) because “normal class A” is 8 and “exception class B” is 60. Is done. Further, the exception rate of “leaf node LP2” is calculated as “0.13” (= 1 / (7 + 1)) because “normal class A” is seven and “exception class B” is one. On the other hand, the exception rate of “leaf node LN2” is calculated as “0.98” (= 59 / (1 + 59)) because “normal class A” is 1 and “exception class B” is 59. Is done. Also, the exception rate of “intermediate node N5” is calculated as “0.95” (= 40 / (2 + 40)) because “normal class A” is 2 and “exception class B” is 40. The Further, the exception rate of “leaf node LP3” is calculated as “0.00” (= 0 / (2 + 0)) because “normal class A” is two and “exception class B” is zero. On the other hand, the exception rate of “leaf node LN3” is calculated as “1.00” (= 40 / (0 + 40)) because “normal class A” is 0 and “exception class B” is 40. Is done.

  The variable narrowing unit 113 narrows down elements included in the route selected from the decision tree 112 that “exception class B” is preferably classified. At this time, the route includes “leaf node LN1 (with a higher ratio of“ exception class B ”out of two“ leaf nodes LP1 and LN1 (LP2, LN2 or LP3, LN3) ”classified from the intermediate node”. LN2, LN3) ”are included. As shown in FIG. 6, “Route R1”, “Route R2”, and “Route R3” are selected from the decision tree 112. That is, the “route R1” is a route including a route in the order of the “intermediate node N2” and the “root node N1”, and has a combination of the variable S2 and the variable S1. The “route R2” is a route composed of a route in the order of “intermediate node N4”, “intermediate node N3”, and “root node N1”, and a combination of a variable S4, a variable S3, and a variable S1. Have. Further, the “route R3” is a route composed of a route in the order of “intermediate node N5”, “intermediate node N3”, and “root node N1”, and includes a combination of a variable S5, a variable S3, and a variable S1. Have.

  That is, the variable narrowing unit 113 narrows down the nodes (variables) included in the selected routes R1, R2, and R3. The narrowing is performed for each route, and the ratio of “exception class B” in the leaf node having the larger “exception class B” ratio among the two leaf nodes classified from the intermediate node is used for narrowing down. Specifically, in the route to be narrowed down, the exception rate of one node (for example, intermediate node N3) is higher than the exception rate of other nodes (for example, intermediate node N4) classified from the one node. When it is larger, excluding the one node (intermediate node N3) from the route is performed recursively. It is assumed that one node (intermediate node N3) excluded by the above-described narrowing is a node classified from the original node (for example, root node N1). In this case, in narrowing down the original node (root node N1), the exception rate of the original node (root node N1) and other nodes further classified from the excluded one node (intermediate node N3) ( The exception rate of the intermediate node N4) is compared. That is, since the other node (intermediate node N4) further classified from the excluded one node (intermediate node N3) is a node classified from the original node (root node N1), the other node ( The exception rate of the intermediate node N4) is compared with the exception rate of the original node (root node N1).

Next, the operation of the variable narrowing unit 113 will be described with reference to FIG.
The variable narrowing-down unit 113 determines that the exception rate “0.50” of “intermediate node N2” is higher than the exception rate “1.00” of “leaf node LN1” classified from “intermediate node N2” in “route R1”. When it is larger, “intermediate node N2” is excluded from “route R1”. However, since the exception rate “0.50” of “intermediate node N2” is smaller than “Exception” “1.00” of “leaf node LN1” (YES), “intermediate node N2” is not excluded from “route R1”. . In addition, the variable narrowing-down unit 113 sets the exception rate “0.71” of “root node N1” in “route R1” to the exception rate “0.50” of “intermediate node N2” classified from “root node N1”. "Root node N1" is excluded from "root R1". As it is, the exception rate “0.71” of “root node N1” is larger than the exception rate “0.50” of “intermediate node N2” (NO), so “root node N1” is excluded from “root R1”. Is done.

  As a result, the two elements included in the “route R1”, “intermediate node N2” and “root node N1” are narrowed down, and a combination of elements including the “intermediate node N2” is selected with respect to the “route R1”. .

  Further, the variable narrowing-down unit 113 sets the exception rate “0.88” of “intermediate node N4” in “route R2” to the exception rate “0.98” of “leaf node LN2” classified from “intermediate node N4”. "Is greater than", exclude "intermediate node N4" from "root R2". However, the exception rate “0.88” of “intermediate node N4” is smaller than “0.98” of “leaf node LN2” (YES), so “intermediate node N4” is not excluded from “route R2”. . Next, in the “route R2”, the variable narrowing unit 113 sets the exception rate “0.91” of “intermediate node N3” to the exception rate “0...” Of “intermediate node N4” classified from “intermediate node N3”. When it is larger than 88, “intermediate node N3” is excluded from “route R2”. As shown, the exception rate “0.91” of “intermediate node N3” is larger than the exception rate “0.88” of “intermediate node N4” (NO), so “intermediate node N3” is excluded from “route R2”. Is done. Subsequently, the variable narrowing-down unit 113 tries to compare the exception rate of “root node N1” with the exception rate of “intermediate node N3” classified from “root node N1” in “route R2”. Since “intermediate node N3” is an excluded node, “intermediate node N4” classified from “intermediate node N3” is further compared. That is, the variable narrowing-down unit 113 sets “exception ratio“ 0.71 ”of“ root node N1 ”in“ root R2 ”to“ intermediate node N4 ”classified from“ root node N1 ”to“ intermediate node N3 ”. When the exception rate is higher than “0.88”, “root node N1” is excluded by “root R2”. However, the exception rate “0.71” of the “root node N1” is smaller than the exception rate “0.88” of the “intermediate node N4” (YES), so the “root node N1” is not excluded from the “route R2”. .

  In this way, the variable narrowing unit 113 recursively compares the exception rate of the node on the root node side and the exception rate of the node on the leaf node side, so that the exception rate of the node selected in each route Even in the comparison, “exception rate of nodes on the leaf node side”> “exception rate of nodes on the root node side” is set.

  As a result, the three elements included in the “route R2”, “intermediate node N4”, “intermediate node N3”, and “root node N1” are narrowed down, and “intermediate node N4” and “root” A combination of elements consisting of “node N1” is selected.

  Further, the variable refining unit 113 sets the exception rate “0.95” of “intermediate node N5” in “route R3” to the exception rate “1.00” of “leaf node LN3” classified from “intermediate node N5”. When the value is larger than “”, “intermediate node N5” is excluded from “route R3”. However, since the exception rate “0.95” of “intermediate node N5” is smaller than “Exception” “1.00” of “leaf node LN3” (YES), “intermediate node N5” is not excluded from “route R3”. . Next, in the “route R3”, the variable narrowing-down unit 113 sets the exception rate “0.91” of “intermediate node N3” to the exception rate “0.0” of “intermediate node N5” classified from “intermediate node N3”. When greater than 95, “intermediate node N3” is excluded from “route R3”. However, since the exception rate “0.91” of “intermediate node N3” is smaller (YES) than the exception rate “0.95” of “intermediate node N5”, “intermediate node N3” is not excluded from “route R3”. . Subsequently, in the “root R3”, the variable narrowing unit 113 sets the exception rate “0.71” of “root node N1” to the exception rate “0...” Of “intermediate node N3” classified from “root node N1”. When it is greater than 91, “root node N1” is excluded from “root R3”. However, the exception rate “0.71” of “root node N1” is smaller than the exception rate “0.91” of “intermediate node N3” (YES), so “root node N1” is not excluded from “root R3”. .

  As a result, although the three elements included in “Route R3”, “Intermediate Node N5”, “Intermediate Node N3”, and “Root Node N1” are narrowed down, none of the nodes are removed from “Route R3”. A combination of elements including “intermediate node N5”, “intermediate node N3”, and “root node N1” is selected for “root R3”.

  In this way, “intermediate node N2” is selected as an element from “route R1”, “intermediate node N4” and “root node N1” are selected as elements from “route R2”, and “root R3” is selected. Are selected as elements “intermediate node N5”, “intermediate node N3”, and “root node N1” (variable narrowing step).

  By the way, each node corresponds to each information output device of the vehicle 20 through a variable used for classification determination. Specifically, “intermediate node N2” is associated with “GPS31” of vehicle 20 via “variable S2”. Further, “intermediate node N4” is associated with “speedometer 33” of vehicle 20 via “variable S4”, and “root node N1” is associated with “clock 30” of vehicle 20 via “variable S1”. ". Further, “intermediate node N5” is associated with “first accelerometer 34” of vehicle 20 via “variable S5”. That is, as the information output device of the vehicle 20 suitable for the classification of “exception class B”, “GPS31” of the vehicle 20 is selected based on “route R1”, and “speedometer” of the vehicle 20 is selected based on “route R2”. 33 ”and“ clock 30 ”are selected, and“ first accelerometer 34 ”,“ illuminance meter 32 ”, and“ clock 30 ”of the vehicle 20 are selected based on“ route R3 ”.

  Thereby, one combination of GPS31 (variable S2) is obtained from "route R1". From the “route R2”, three combinations of the speedometer 33 (variable S4), the clock 30 (variable S1), the speedometer 33, and the clock 30 are obtained. Furthermore, from the “route R3”, the first accelerometer 34 (variable S5), the illuminometer 32 (variable S3), the clock 30 (variable S1), the first accelerometer 34, the illuminometer 32, and the first acceleration. Seven combinations of the meter 34 and the clock 30, the illuminometer 32 and the clock 30, the first accelerometer 34, the illuminometer 32, and the clock 30 are obtained. That is, the variable narrowing-down unit 113 can obtain 11 types of variable combinations from the decision tree 112 and 10 types excluding duplication.

  Thus, the combination of the information output devices (elements) of the vehicle 20 selected from each route is provided from the variable narrowing-down unit 113 to the vehicle diagnosis device or the like as a combination that can suitably classify the “exception class B”. In the vehicle diagnosis apparatus or the like, verification regarding the occurrence of “exception class B” is performed based on the combination of the narrowed elements provided from the variable narrowing unit 113. At this time, since combinations of variables are narrowed down, it is possible to reduce the processing load by the vehicle diagnostic apparatus or the like and improve the processing speed.

  For example, the vehicle diagnosis apparatus can verify the occurrence of a case corresponding to “exception class B” based on the current position Pa obtained from the GPS 31. Further, the vehicle diagnostic apparatus can verify the occurrence of the case corresponding to “exception class B” based on a combination of the speed Va obtained from the speedometer 33 and the time Ta obtained from the timepiece 30. it can. Further, the vehicle diagnostic apparatus can obtain the occurrence of the case corresponding to “exception class B” from the acceleration Aa obtained from the first accelerometer 34, the velocity Va obtained from the speedometer 33, and the timepiece 30. It can be verified based on the combination of the time Ta.

As described above, according to the factor analysis device of the present embodiment, the effects listed below can be obtained.
(1) From the decision tree 112, a variable group is obtained as a plurality of variables (S1 to S30) that reduce the variation of the “exception class B”, that is, favorably divide the “exception class B”. A plurality of variables included in the variable group are “exception class B” assigned to each node (N1 to N5, LN1 to LN3) having one or a plurality of samples classified by the variable. Filter based on percentage. In other words, the variables included in the variable group are further narrowed down, and the number of variable combinations based on the one or more variables thus narrowed is reduced, and the time required for the verification related to “exception class B” performed for each of these combinations, Time and effort will be reduced. In addition, if the variables included in the variable group are narrowed down so that appropriate analysis and estimation can be performed using the variables further narrowed down from the variables, the estimation accuracy can be improved as compared with the prior art. That is, in contrast to a system in which a large number of variables (elements) affect the “exception class B” (output result), this factor analysis device makes it easier and more expensive for the variable that has influenced the “exception class B” It can be analyzed and estimated with accuracy.

  (2) Since the variable group is obtained based on the algorithm of the decision tree 112, the variables included in the variable group are appropriately narrowed down from the variable groups obtained in various analysis devices to which the algorithm of the decision tree 112 is applied. become.

  (3) The “exception class” included in the samples (D1 to D210) on the root node (N1) side rather than the ratio of “exception class B” included in the samples (D1 to D210) on the leaf nodes (LN1 to LN3) side When the ratio of “B” is large, the variables (S1 to S30) on the root node (N1) side are excluded from the variable group. As a result, since variables that have a small contribution to the classification of a plurality of samples are excluded from the variables included in the variable group, a plurality of samples are classified according to “exception class B” even in the variable group in which the variables are narrowed down. It can maintain suitably.

  (4) Leaf nodes (LN1 to LN3) suitable for variable narrowing are selected, and the ratio of “exception class B” included in the sample of the selected leaf nodes (LN1 to LN3) is included in the variable group. It is used for narrowing down a plurality of variables (S1 to S30). That is, it becomes possible to narrow down the variables included in the variable group.

  (5) Variables (S1 to S30) that are considered to have an influence on the occurrence of “exception class B” can be acquired from a plurality of samples D1 to D210. The exceptional case is “exception class B”, which is an undesirable case for the vehicle 20. From this, it is possible to appropriately deal with an exceptional case occurring in the vehicle 20 because acquisition of a variable that affects the exceptional case can suitably execute verification regarding a case that is not preferable for the vehicle 20. Be expected.

  (6) In the vehicle 20, there is a correlation between the vehicle state such as the speed Va and the acceleration Aa, there is a correlation between the vehicle state such as the steering operation and the road state, and the vehicle environment. In many cases, there is a correlation between elements (variables) such as, that is, the independence between elements is low and there is mutual interference. It is not easy to estimate a combination of elements that cause a specific target vehicle state, so-called exceptional case (for example, “exception class B”), from a plurality of such interference-prone elements. In addition, it is not easy to select an element that suitably separates a target specific vehicle state from elements having a lot of interference, and the number of elements that are selected as having an influence on a specific vehicle state by analysis is large. Tend to be. Therefore, when verifying a specific vehicle state based on the selected combination of elements, many combinations of these elements are verified, and the time and effort required for verification cannot be ignored.

  On the other hand, according to the present embodiment, the vehicle state or vehicle environment obtained from the vehicle state or vehicle environment acquired from the vehicle 20 as affecting the “exception class B” is the vehicle state or vehicle environment. Narrowing is performed based on the ratio of “exception class B” assigned to one or more samples (D1 to D210) classified by the environment. That is, even in a system such as the vehicle 20 that tends to have a large number of variables selected to affect a specific vehicle state, the number of vehicle states or vehicle environments to be selected can be narrowed down. . As a result, it becomes possible to verify the vehicle state that generates the “exception class B” based on the narrowed combination of the vehicle state or the vehicle environment, and therefore the verification corresponding to the number of combinations of the vehicle state or the vehicle environment. The number is reduced, and the time and labor required for verification of “exception class B” are reduced.

(Other embodiments)
In addition, the said embodiment can also be implemented in the following aspects, for example.
-In above-mentioned embodiment, it illustrated about the case where each vehicles 20a-20d consist of the same vehicle type. However, the present invention is not limited to this, and if the exception example to be analyzed and estimated is not influenced by the vehicle type or is a small exception case, each vehicle may be a different vehicle type instead of the same vehicle type. This improves the applicability of the factor analysis device.

  -In above-mentioned embodiment, it illustrated about the case where each vehicle 20a-20d collects the data which consist of the same kind. However, the present invention is not limited to this, and data that does not affect the exception cases to be analyzed and estimated, such as data that is not used for analysis, may be collected according to the situation of each vehicle. On the contrary, it is not necessary to collect. Thereby, the freedom degree of the structure of the vehicle to which a factor analysis apparatus is applied comes to improve.

  In the above embodiment, the case where the sample D1 stored in the storage device 22 of the vehicle 20 is transmitted to the information processing center 10 is illustrated. However, the present invention is not limited to this, and the vehicle may immediately transmit the acquired sample to the information processing center. As a result, the degree of freedom in designing a system to which factor analysis can be applied is improved.

  In the above embodiment, the case where the information processing center 10 and the plurality of vehicles 20 are wirelessly connected has been illustrated. However, the present invention is not limited to this, and if the vehicle data can be collected in the information center, the stopped vehicle may be wired to the information center and transferred from the vehicle to the information center. After data is stored in a storage medium such as an HDD or a CD-ROM, the data may be transferred from the storage medium to the information center. As a result, the degree of freedom in designing the system configuration can be improved.

  In the above embodiment, the time Ta, the current position Pa, the illuminance value La, the speed Va, the first acceleration Aa, the second acceleration Ab, and the like acquired at the same time are included in one sample D1 (D2 to D210). Exemplified the case. However, the present invention is not limited to this, and values such as time, current position, illuminance value, travel speed, first acceleration, and second acceleration included in the sample can be acquired simultaneously if they can be used for factor analysis. It does not have to be. For example, a value may be added later, or a value having a time difference may be used as one sample. This expands the possibilities of the system to apply this factor analysis device.

  -In above-mentioned embodiment, it illustrated about the case where a some sample consists of 210 samples D1-D210. However, the present invention is not limited to this, and the number of the plurality of samples may be less than 210 or more than 210. An analysis result corresponding to the number of samples can be obtained regardless of the number of samples. Thereby, the design freedom of a factor analysis apparatus can be raised.

  In the above embodiment, the case where each sample D1 to D210 includes 30 values dn, m corresponding to 30 types of variables S1 to S30 is illustrated. However, the present invention is not limited to this, and the number of variables included in each sample, that is, the number of values acquired from the information output device may be less than 30 or more than 30, respectively. An analysis result based on the value from the information output device can be obtained regardless of the number of the information output devices. Thereby, the design freedom of a factor analysis apparatus can be raised.

  In the above embodiment, the case where each sample D1 to D210 includes values dn and m corresponding to 30 types of variables S1 to S30 is illustrated. However, the present invention is not limited to this, and the sample may not include some information output device values. Even if a sample has a missing value, it can be used for factor analysis that does not affect the missing value. On the other hand, the sample can be excluded from the factor analysis that affects the missing value. Thereby, the possibility of the system which can apply this factor analysis apparatus is expanded.

  In the above embodiment, the case where two classes of “normal class A” and “exception class B” are provided as the classification class is illustrated. However, the present invention is not limited to this, and the classification class may be provided with more than two classes. Even in this case, it is possible to appropriately narrow down the variables that affect the target class. As a result, the factor analysis apparatus can be applied to various decision tree analyses.

  In the above embodiment, the case where the decision tree 112 is created by dividing the samples D1 to D210 by the five variables S1 to S5 is illustrated. However, the present invention is not limited to this, and the number of variables required for the factor analysis may be at least more than five than five. Thereby, it becomes possible to perform factor analysis for various states.

  In the above embodiment, the variable narrowing-down unit 113 is exemplified for the case where the exception rate is also calculated for the leaf nodes LP1 to LP3. However, the present invention is not limited to this, and the variable narrowing-down unit is not used for narrowing down the leaf node LP1. It is not necessary to calculate the exception rate of ~ LP3. Thereby, the freedom degree of the design concerning factor analysis comes to improve.

  In the above embodiment, the case where the information processing center 10 is outside the vehicle 20 is illustrated. However, the present invention is not limited to this, and all or part of the functions in the information processing center may be provided in the vehicle. For example, a single vehicle may acquire a sample and perform a factor analysis on the vehicle, or one vehicle may acquire information on another vehicle from another vehicle through an information processing center or inter-vehicle communication. Then, factor analysis for the vehicle may be executed. Further, all or part of the functions in the information processing center may be provided in another vehicle.

As a result, the degree of freedom of design and the possibility of adopting such a factor analysis device are improved.
In the above embodiment, the case where a node satisfying “exception rate of nodes on the leaf node side”> “exception rate of nodes on the root node side” is selected is illustrated. However, the present invention is not limited to this, and a node satisfying “exception rate of nodes on the leaf node side” ≧ “exception rate of nodes on the root node side” may be selected, and at least one may be multiplied by a predetermined coefficient for comparison. It may be. As a result, the analysis result of the factor analysis device can be suitably adjusted.

  In the above embodiment, when the ratio of “exception class B” of the node on the root node side is larger than the ratio of “exception class B” of the node on the leaf node side, the variable of the node on the root node side is excluded. The case of repeating recursively was illustrated. However, the present invention is not limited to this, and the node (variable) exclusion may not be repeated recursively but may be interrupted in the middle. Even if it is interrupted, the number of combinations can be reduced according to the excluded variables. Thereby, the design flexibility of the factor analysis device can be improved.

  In the above embodiment, the case of analyzing the cause of “exception class B” is exemplified. However, the present invention is not limited to this, and normal class factors may be analyzed. At this time, for example, if the variable narrowing unit narrows the variables based on the normal rate, it can narrow the variables included in the variable group. As a result, the application range of the factor analysis device can be expanded.

  In the above embodiment, the case where the factor analysis of the vehicle 20 is performed is illustrated. However, the present invention is not limited to this, and if multiple samples containing the same type of variable can be obtained, it can be used for moving objects such as railways, ships, and airplanes, manufacturing equipment such as production lines and plants, and data processing such as road information. Can also be applied. Thereby, the application range of such a factor analysis device can be expanded.

  DESCRIPTION OF SYMBOLS 10 ... Information processing center, 11 ... Information processing part, 12 ... Memory | storage device, 13 ... Communication part, 20 ... Vehicle, 20a-20d ... Vehicle, 21 ... Vehicle control apparatus, 22 ... Memory | storage device, 23 ... Communication part, 30 ... Clock 31, GPS 32 illuminometer 33 33 speedometer 34 first accelerometer 35 second accelerometer 111 decision tree creation unit 112 decision tree 113 variable narrowing unit 121 ... Database, C1-C5 ... Threshold, D1-D210 ... Sample, LN1-LN3 ... Leaf node, LP1-LP3 ... Leaf node, N1 ... Root node, N2-N5 ... Intermediate node, R1-R3 ... Root, X1 ... Classification class.

Claims (8)

A storage unit in which a plurality of samples including a plurality of variables are stored by assigning one of a plurality of classification classes to the samples;
The classification class is an objective variable for classifying a plurality of samples, and the sample is an explanatory variable classified by the classification class. The classification is affected by sequentially classifying the plurality of samples into samples according to the variables while sequentially selecting the variables so that the variation of the selected specific classification class among the plurality of classification classes is reduced. An analysis unit that obtains a variable group consisting of a plurality of variables that affect
A factor analysis device comprising:
A variable refining unit that narrows down a plurality of variables included in the variable group based on a ratio of the specific classification class set in a plurality of samples classified by each variable ;
The analysis unit obtains the decision tree as having one root node including all samples and a plurality of leaf nodes including one or a plurality of samples that cannot be classified as a result of sequential classification. A plurality of variables included in the variable group are configured from variables included in a path from the one root node to one of the plurality of leaf nodes,
The variable narrowing unit is configured to narrow down a plurality of variables included in the variable group, and a ratio of a specific classification class that is associated with a plurality of samples classified by a plurality of variables included in the variable group. By recursively executing the removal of the variable from the variable group when the ratio is greater than the ratio of a specific classification class associated with one or more samples classified on the leaf node side in the path A factor analysis device characterized by that. When a plurality of leaf nodes are generated by one or a plurality of samples classified by the variable, the variable narrowing unit has a ratio of a specific classification class associated with each leaf node sample of the plurality of leaf nodes. The factor analysis apparatus according to claim 1 , wherein a ratio of a specific classification class in a larger leaf node is used for narrowing down the variable group. The plurality of classification classes are composed of two classes, a normal class indicating a normal case and an exception class indicating an exception case.
Factors Analysis device according to claim 1 or 2 wherein the specific classification class is the exception class. The variable is information indicating a vehicle state or a vehicle environment,
The sample comprises a plurality of vehicle states or vehicle environments obtained at the same time from one vehicle, and the plurality of samples comprise samples in which at least one of time and vehicle is different from each other,
As the exception class, a specific vehicle state is set,
The factor analysis device according to claim 3 , wherein the variable narrowing-down unit narrows down a plurality of vehicle states or vehicle environments included in the variable group based on a ratio at which the specific vehicle state is set. A storage step in which the storage unit stores a plurality of samples including a plurality of variables in the storage unit such that one classification class of a plurality of classification classes is assigned to each of the samples;
The analysis unit uses the classification class as an objective variable for classifying a plurality of samples, and uses the sample as an explanatory variable classified by the classification class. When the classification is performed, the plurality of samples are sequentially classified into samples for each variable while sequentially selecting the variables so that the variation of the selected specific classification class among the plurality of classification classes is reduced. An analysis process for obtaining a variable group consisting of a plurality of variables that affect classification in
A variable refining unit comprising a variable refining step of refining a plurality of variables included in the variable group based on a ratio of the specific classification class set in a plurality of samples classified by each variable ;
In the analysis step, the decision tree is obtained as having one root node including all samples and a plurality of leaf nodes including one or a plurality of samples that cannot be classified as a result of sequential classification. A plurality of variables included in the variable group are configured from variables included in a path from the one root node to one of the plurality of leaf nodes,
In the variable narrowing-down step, the ratio of a specific classification class that is associated with a plurality of samples classified by a plurality of variables included in the variable group by narrowing down a plurality of variables included in the variable group, By recursively executing the removal of the variable from the variable group when the ratio is greater than the ratio of a specific classification class associated with one or more samples classified on the leaf node side in the path A factor analysis method characterized by this. In the variable narrowing-down step, when a plurality of leaf nodes are generated by one or a plurality of samples classified by the variable, a ratio of a specific classification class associated with each leaf node sample of the plurality of leaf nodes is The factor analysis method according to claim 5 , wherein a ratio of a specific classification class in a larger leaf node is used for narrowing down the variable group. The plurality of classification classes are composed of two classes, a normal class indicating a normal case and an exception class indicating an exception case.
The factor analysis method according to claim 5, wherein the specific classification class is the exception class. Using information indicating the vehicle state as the variable,
The sample is a plurality of vehicle states or vehicle environments obtained from a single vehicle at the same time, and the plurality of samples are samples whose time and vehicle are different from each other,
As the exception class, set a specific vehicle state,
The factor analysis method according to claim 7 , wherein in the variable narrowing-down step, narrowing down a plurality of vehicle states or vehicle environments included in the variable group is performed based on a set ratio of the specific vehicle state.

Images