JP2010129003A - Disaster prevention comprehensive plan support system and program of the same - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a disaster prevention comprehensive plan support system and program for creating a rule for efficiently and accurately evaluating the occurrence/non-occurrence of disaster-accident-repair for accident and repair of the sediment disasters caused on a slope or mountain stream by rainfall, natural disasters such as a disaster or river disaster caused by an earthquake or volcanic activity, facilities or equipment such as a road or a water and sewage, a dam, a harbor, and a sand guard, and building structures such as a bridge and steel tower, a plant, and a building. <P>SOLUTION: This disaster prevention comprehensive plan support system includes a risk arithmetic section 12 for calculating the disaster occurrence risk of an observation-inspection object part using separation surface data 17 and observation-inspection data 8 for forming the boundary between the occurrence and non-occurrence of the disaster analyzed by an SVM, a representative data extracting section 13 for extracting representative data 19 using this risk, a rough assembly creating section 22 for forming a rough assembly using the representative data 19, and a rule creating section 23 for extracting a rule of disaster-accident-repair from the rough assembly. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to earth and sand disasters that occur on slopes and mountain streams due to rainfall, natural disasters such as disasters caused by earthquakes and volcanic activities, river disasters, and facilities and equipment such as roads, water and sewage systems, dams, ports, and sabos. Efficient and highly accurate assessment of the occurrence / non-occurrence of disasters / accidents / repairs for aging of buildings, bridges, steel towers, power plants, buildings, etc. and accidents and repairs associated with the above natural disasters Disaster prevention comprehensive plan support system and program for supporting

Natural disasters such as heavy rains, earthquakes, and volcanic activities have caused enormous damage to Japan every year. For example, about 70% of Japan's national land is mountainous, and there are many geologically vulnerable areas, and there are many steep landforms. It is one of the fatal natural disasters in Japan due to social conditions such as population growth in Japan and weather conditions such as typhoons and rainy seasons that are easily affected by heavy rains. However, precious lives have been lost and valuable assets have been destroyed, but there are many places where there are about 520,000 dangerous places for debris disasters (debris flow dangerous mountain streams, steep slope collapse dangerous places, landslide dangerous places) nationwide, Currently, the maintenance rate due to hardware measures is as low as 25%.
The same applies to facilities and equipment such as roads, water and sewage systems, dams, and harbors, or aging of buildings such as bridges, steel towers, power plants, and buildings, and accidents and repairs associated with the natural disasters described above.
In response to such a situation, there are budgetary and time constraints to implement hardware countermeasures and repair / renewal work for all of these dangerous locations and facilities / equipment / buildings. The need for measures to cover delays in hard measures has been recognized, and there are points and places where natural disasters such as dangerous mountain streams that are widely scattered in the region are likely to occur, aging and natural disasters. The importance of strengthening prevention against occurrence of disasters, accidents and repairs by grasping the distribution of particularly dangerous parts and relatively safe parts among structures that may be damaged by accidents involved Yes.
So, for example, to support the planning of disaster prevention / repair business plans, we have been conducting research to obtain highly accurate information by computer processing data related to the occurrence or non-occurrence of actual disasters / accidents / repairs. The present inventors have already disclosed a method for setting a landslide occurrence limit line, an evacuation reference line, and a warning reference line, which are discrimination boundaries between generated rainfall and non-occurrence rainfall used for predicting occurrence of landslide. Announced as shown in
In Non-Patent Document 1, there is a case where a radial basis function network excellent in non-linear discrimination (hereinafter abbreviated as RBFN) is used for the purpose of setting a high-precision generation limit line or the like without performing a linear approximation of complex natural phenomena. ), And proposes a method for setting a non-linear collapse critical rainfall line for each region. In the technology disclosed in Non-Patent Document 1, a non-linear collapse occurrence limit rainfall line is set by using RBFN and determining the optimum weight of the intermediate layer and the output layer using the learning function.
Patent Document 1 discloses a system provided with a means for automatically selecting and displaying a disaster countermeasure to be executed when a disaster occurs, and indicating the progress status as a “disaster countermeasure support system”. Yes.
The disaster countermeasure support system disclosed in this Patent Document 1 basically reads and copes with a list of countermeasures stored in association with events that occur in advance and corresponding countermeasures in an if-then format. is there. It was designed to enable accurate judgments in situations where there is no mental, temporal, or human capacity at the time of a disaster. In addition, by displaying the standard work time, the work time required during actual work, and the remaining time that can be taken, it is possible to grasp the progress of the measure in real time, and at the same time, measures with high importance and low measures It can also be used to select.
Further, Patent Document 2 discloses an invention in which the technology disclosed in Non-Patent Document 1 is applied to a warning and evacuation system. In the invention disclosed in this patent document 2, a short-term rainfall index, for example, the maximum time within 3 hours from the time of occurrence is taken into account, taking into account the topographic factors, geological / soil factors, environmental factors and earthquake factors that affect disasters. CL is calculated by using the rainfall and the effective rainfall for a long-term rainfall index, for example, the half-life at that time is 72 hours.
By using CL thus obtained, it is possible to provide a highly reliable warning and evacuation support system.
Furthermore, Non-Patent Document 2 created a rule for the occurrence / non-occurrence of debris flow by rough set using terrain / geological factors (hereinafter referred to as terrain factors) and rainfall factors. Research that pursues the factors is disclosed. This study explores objective disaster conditions by introducing mathematical methods.
Kazumasa Kuramoto et al. 5: Research on the setting of the non-linear collapse occurrence limit rainfall line using the RBF network, No. 672 / VI-50, pp. 117-132, 2001.3 Masao Okamoto and 4 others: Research on important factors and rule extraction of sediment movement phenomenon by data mining using rough set, Journal of Sabo Society Vol. 54, no. 6, p. 4-15, 2002 JP 2002-230235 A JP 2003-184098 A

However, in the inventions disclosed in Non-Patent Document 1 and Patent Document 2, the focus is on setting a disaster occurrence limit line, an evacuation reference line, and a warning reference line, and a specific area or a certain condition In the regional groups compiled for each, the objective was to objectively evaluate the extent to which the short-term rainfall index and the long-term rainfall index reached the risk of disaster occurrence. Extremely speaking, the data of short-term and long-term rainfall index accumulated at the same point is input, and how much rain will cause a disaster based on the rain data accumulated at that point. Judgment was made.
Even though this is an objective and quantitative evaluation, it is possible to carry out individual specific evaluations for each region or group, but it is not a specific region but a general and universal common to all regions. There was a problem that it was difficult to evaluate.

  In addition, in the invention disclosed in Patent Document 1, a specific implementation of countermeasures using a list-structured data structure with predetermined or known conditions and countermeasures although they are basically complicated. The procedure is shown. Although the countermeasure list can certainly be corrected and updated, the judgment is basically made based on the input data, and the computer only performs the process of combining the event and the countermeasure, and the data itself There is a problem that it is difficult to carry out the evaluation and improve the accuracy.

Further, in Non-Patent Document 2, there is a problem that a very diverse rule is created because it is intended to explain the occurrence / non-occurrence of as many disasters as possible only by the analysis result of the rough set. This is presumed that the influence of noisy data included in the population data (data in which a disaster has occurred / not occurred due to local and special conditions) is large.
If the rules are too diverse, it will be difficult to perform the evaluation efficiently even if the accuracy is high. In other words, the number of points and places that can be covered by the rules are reduced, and it becomes difficult to evaluate the occurrence / non-occurrence of disasters. That is, there has been a problem that it is impossible to carry out highly accurate and efficient evaluation on occurrence / non-occurrence of disasters. This also applies to Non-Patent Document 1, Patent Document 1, and Patent Document 2.
In this application, “rule” means a combination of category sections for each factor in the rough set, and “rough set” means a category section for attribute values related to this factor for disaster / accident / repair occurrence factors. This is a set obtained by combining the category intervals for each factor to form a model and substituting observation and inspection data into this model, and is not completely separated by the subset formed by the combination of category intervals Means a set. In addition, in this application, “category section” refers to attribute values of factors causing disasters, accidents, repairs, such as observed values or inspection values such as rainfall, basin average slope, basin length, age, degree of damage, etc. And that section divided as a section. For example, when the factor is a basin average gradient, three ranges of 10 ° to 20 °, 20 ° to 30 °, and 30 ° to 40 ° are considered as examples of category sections. This also applies to east, west, south, north, etc. that do not have a physical order such as mountain stream direction.

  The present invention has been made in response to such a conventional situation, such as a landslide disaster that occurs on a slope or a mountain stream caused by rainfall, a natural disaster such as a disaster or a river disaster caused by an earthquake or a volcanic activity, and a road. Efficient and accurate disasters for facilities and equipment such as water and sewage systems, dams, ports, sabos, bridges, steel towers, power plants, buildings, etc., and accidents and repairs associated with the above natural disasters -It aims at providing the disaster prevention comprehensive plan support system which can produce | generate the rule (Rule) for performing occurrence / non-occurrence evaluation of an accident and repair, and its program.

In order to achieve the above object, the comprehensive disaster prevention plan support system according to the first aspect of the present invention provides an occurrence factor of disaster / accident / repair (where the number of factors is n. N ≧ 2) Corresponding to the coordinate axes that make up the n-dimensional coordinate space, the n-dimensional observation / inspection data observed and inspected for each cause of disaster / accident / repair is used as learning data, and the disaster / accident at the location subject to observation / inspection・ The occurrence / non-occurrence of repairs is used as a teacher value, and the occurrence / non-occurrence of disasters / accidents / repairs is analyzed using a support vector machine (hereinafter referred to as SVM) to generate disasters / accidents / repairs. The n-dimensional observation / inspection data indicates the separation plane that forms a non-occurrence boundary and the n-dimensional observation / inspection data for each factor observed / inspected at the observation / inspection target location. Said The distance from the coordinate point in the dimensional coordinate space to the separation surface (however, this distance may be positive when the coordinate point is on the origin side of the n-dimensional coordinate space with respect to the separation surface, Alternatively, the risk may be negative if it is on the origin side.) A risk calculation unit that calculates the risk of occurrence of disaster, accident, or repair at the location subject to observation / inspection, and a desired risk for occurrence of disaster, accident, or repair The corresponding disaster / accident / repair occurrence risk is extracted on the basis of the threshold value or condition, and n-dimensional observation is performed for each factor corresponding to the extracted disaster / accident / repair occurrence risk・ A representative data extraction unit that extracts inspection data as representative data, and a category section for attribute values related to the factor of occurrence of n-dimensional disaster / accident / repair, and a category for each factor. -A section is combined to form a model, the representative data is substituted into this model, a rough set generation unit that forms a rough set, and a combination of category sections for each factor in the rough set is generated as a rule, Calculate the certainty of disaster, accident and repair for each rule and support, and set the rule based on the desired threshold or condition for the certainty and / or the desired threshold or condition for the support. And a rule generation unit for extracting disasters / accidents / repairs.
In this application, in order to calculate the n-dimensional observation / inspection data observed and inspected for each cause of disaster, accident, and repair as learning data for analyzing the separation surface, and the distance to the separation surface The n-dimensional observation / inspection data for each factor may be the same or different. If they are different, they may be observed / inspected at the same point / location at different times, or may be observed / checked at different locations / locations. Further, in this application, “threshold” and “condition” are separately described, but this means that “threshold” is a concept that is limited only by numerical values, and “condition” is, for example, This is because the concept of including a logical expression including a threshold is also considered.

  The disaster prevention comprehensive plan support system according to claim 2 constitutes an n-dimensional coordinate space for disaster / accident / repair occurrence factors (the number of factors is n. N ≧ 2) in the observation / inspection target locations. The n-dimensional observation / inspection data observed and inspected for each cause of disaster, accident, and repair corresponding to the coordinate axis is used as learning data, and the occurrence, non-occurrence of disaster, accident, and repair at the location subject to observation and inspection As a teacher value, and the possibility of occurrence / non-occurrence of disasters / accidents / repairs using a support vector machine (hereinafter referred to as SVM) to form the boundary of occurrence / non-occurrence of disasters / accidents / repairs Using the separation surface calculation unit that analyzes the separation surface, the separation surface analyzed by the separation surface calculation unit, and the n-dimensional observation / inspection data for each of the factors observed and inspected at the observation / inspection target location, N dimension The distance from the coordinate point in the n-dimensional coordinate space indicated by the observation / inspection data to the separation plane (however, this distance is the origin point of the n-dimensional coordinate space with respect to the separation plane) A risk calculation unit that calculates the risk of occurrence of a disaster, an accident, or a repair at the observation / inspection location; Based on the desired threshold or condition for the accident / repair occurrence risk, the corresponding disaster / accident / repair occurrence risk is extracted, and the extracted disaster / accident / repair occurrence risk is dealt with. A representative data extraction unit that extracts n-dimensional observation / inspection data for each factor as representative data, and category categories for attribute values related to the n-dimensional disaster / accident / repair occurrence factor A model is formed by combining the category intervals for each factor, the representative data is substituted into the model, and a rough set generation unit for forming a rough set, and a category interval for each factor in the rough set A combination is generated as a rule, and the certainty of disaster, accident and repair and the support for each rule are calculated, and a desired threshold or condition for the certainty and / or a desired threshold or condition for the support And a rule generation unit for disaster / accident / repair that extracts the rule with reference to the above.

  The comprehensive disaster prevention plan support program according to claim 3 is a program executed by a computer to generate rules for disasters, accidents, and repairs, and the computer is provided with an observation / inspection location. Causes of occurrence of disasters / accidents / repairs (the number of factors is n. N ≧ 2) corresponded to the coordinate axes constituting the n-dimensional coordinate space, and were observed and inspected for each cause of occurrence of disasters / accidents / repairs. Use n-dimensional observation / inspection data as learning data and use the occurrence / non-occurrence of disasters / accidents / repairs at the locations subject to observation / inspection as teacher values to support the possibility of occurrence / non-occurrence of disasters / accidents / repairs The separation plane that is analyzed by using a vector machine (hereinafter referred to as SVM) to form the boundary of occurrence / non-occurrence of disaster / accident / repair, and observation / point The distance from the coordinate point in the n-dimensional coordinate space indicated by the n-dimensional observation / inspection data to the separation plane (however, this distance) May be positive when the coordinate point is on the origin side of the n-dimensional coordinate space with respect to the separation plane, or may be negative when the coordinate point is on the origin side). Based on the risk calculation process for calculating the risk of occurrence of disaster, accident, or repair at a location, and the desired threshold or condition for the risk of occurrence of disaster, accident, or repair, the applicable disaster, accident, or repair risk A representative data extraction step of extracting n-dimensional observation / inspection data for each factor corresponding to the extracted disaster / accident / repair occurrence risk as representative data, and the n-dimensional disaster For accident / repair factors, a category section for the attribute value related to this factor is provided, a model is formed by combining the category sections for each factor, and the representative data is substituted into this model to form a rough set. A rough set generation step and a combination of category sections for each factor in the rough set are generated as rules, and the certainty of disaster / accident / repair and support for each rule are calculated, and a desired threshold for the certainty is calculated. A disaster / accident / repair rule generation step for extracting the rule based on a value or condition and / or a desired threshold value or condition for the support is performed.

  The disaster prevention comprehensive plan support program according to claim 4 is a program that is executed by a computer to generate rules for disaster, accident, repair occurrence, and non-occurrence.・ The occurrence factor of accident / repair (the number of factors is n. N ≧ 2) corresponds to the coordinate axes that constitute the n-dimensional coordinate space, and n is observed and inspected for each occurrence factor of the disaster / accident / repair. Dimensional observation / inspection data is used as learning data, and the occurrence / non-occurrence of disasters / accidents / repairs at the observation / inspection locations is used as a teacher value to support the possibility of occurrence / non-occurrence of disasters / accidents / repairs. Using the machine (hereinafter referred to as SVM), the separation plane calculation process for analyzing the separation plane that forms the boundary of occurrence / non-occurrence of disaster / accident / repair, and analysis by this separation plane calculation section In the n-dimensional coordinate space indicated by the n-dimensional observation / inspection data, using the separated plane and the n-dimensional observation / inspection data for each of the factors observed / inspected at the observation / inspection target location. The distance from the coordinate point to the separation plane (however, this distance may be positive when the coordinate point is on the origin side of the n-dimensional coordinate space with respect to the separation plane, or the origin side The risk level calculation step for calculating the disaster / accident / repair occurrence risk at the location subject to observation / inspection, and a desired threshold for the disaster / accident / repair occurrence risk Or, based on the conditions, the corresponding disaster / accident / repair occurrence risk is extracted, and n-dimensional observation / inspection data for each factor corresponding to the extracted disaster / accident / repair occurrence risk is obtained. Representative data For the n-dimensional disaster / accident / repair occurrence factor, a category section is provided for the attribute value related to this factor, and a model is formed by combining the category sections for each factor. By substituting data into this model, a rough set generation process that forms a rough set and a combination of category sections for each factor in the rough set are generated as rules, and the certainty of disaster, accident, and repair for each rule And a rule generation step for disaster, accident, and repair that calculates a support and extracts the rule based on a desired threshold value or condition for the certainty factor and / or a desired threshold value or condition for the support. , Are executed.

The present invention does not use the observation / inspection data obtained at disaster risk points or facilities / equipment / buildings as a rough set as it is to evaluate the occurrence / non-occurrence of disasters / accidents / repairs as a preliminary stage.・ A part of the observation / inspection data is objectively extracted as representative data using the separation surface of inspection data and SVM analysis, and a rough set is generated using the representative data. Since a rule for evaluating occurrence / non-occurrence of repair is generated, noise data included in observation / inspection data can be deleted, and an objective and highly accurate rule can be generated.
In addition, although the generated rule is composed of a small number of factors, it has an advantage that it is highly versatile and can explain a lot of observation / inspection data.

Hereinafter, in the disaster prevention comprehensive plan support system according to the embodiment of the present invention, SVM is used to extract representative data for generating disaster / accident / repair rules, and disaster / accident / repair rules A rough set was used for generation. Therefore, before explaining the disaster prevention comprehensive plan support system according to the present embodiment, first, SVM and rough set will be explained.
SVM is a technique proposed by Vapnik et al. In 1995 and is currently attracting attention as the most powerful pattern classification technique. In SVM, a data group that cannot be linearly separated is mapped to a high-dimensional feature space using a non-linear function, and a function of clearly separating the data group by making it linearly separable. At this time, a surface that separates data in the high-dimensional feature space is called a separation hyperplane. The concept of this separation hyperplane is shown in FIG. In the present application, this separation hyperplane is simply referred to as a separation surface.
Using this function, the degree of risk can be evaluated by calculating the distance between each data and the separated hyperplane that discriminates the occurrence / non-occurrence of disasters / accidents / repairs. FIG. 2 shows the concept of calculating the distance between the separated hyperplane and each data to evaluate the risk level. In this application, this risk assessment method is applied to represent the original observation / inspection data in determining the disaster / accident / repair rules (representing data with higher risk as disaster / accident / repair occurrence data) As a non-occurrence data, the work of extracting the safer).
The distance between the data point and the separation hyperplane does not need to be described again, but a perpendicular is drawn from the coordinates of the data point to the separation hyperplane to distance the length. However, in FIG. 2, since the dangerous side is the positive direction as the coordinate and the distance (f (x)) is calculated from the separation hyperplane toward the data point, that is, it is considered as a vector, the data point is separated. The distance increases from the hyperplane in the positive direction, and the distance increases as a negative value. Conversely, as the data point moves away from the separation hyperplane in the negative direction (closer to the coordinate origin), the distance increases. In this embodiment, although such a coordinate is adopted, on the contrary, when the safe side is a positive direction as a coordinate, the negative value is safer and the positive value is more positive. Needless to say, it can be dangerous.

Next, the rough set will be described. FIG. 3 shows a conceptual diagram of the rough set. Three categories are provided for the entire set using two factors (factor 1 and factor 2), and a model is formed in which each combination is considered to be divided into nine regions. As shown in FIG. 3, data relating to the occurrence / non-occurrence of disasters / accidents / repairs is plotted for this model.
In the rough set shown in FIG. 3, the subsets are not completely separated even if divided into nine regions as described above. Such a set that is not completely separated is called a rough set. When a set is divided by these factors, data in which all data in the same area is the same type is called consistent data (data in the hatched portion in FIG. 3), and mixed data is called contradictory data. As an index for evaluating the combination of factors, the degree of consistency defined by the following equation (1) is used.

The degree of consistency is determined by replacing the category section name described above (for example, indicating the serial number assigned to each category section) as an attribute value, and determining the occurrence / non-occurrence of disaster / accident / repair attribute value (for example, (Indicates that the value is 1 when disaster / accident / repair occurs, and 0 when non-occurrence occurs)), individual data is classified, areas with inconsistent data with different decision attribute values are inconsistent areas, others are consistent areas As a result, the degree of matching is used as an index for evaluating a combination of factors in the region.
In addition, as the scale representing the accuracy and versatility of the created rule (Rule), the certainty factor and support shown in equations (2) and (3) are used.

  For example, in FIG. 3, when both the category divisions of factor 1 and factor 2 are 2, that is, the region of the central division is an occurrence rule, the certainty in this region is 100% (3/3), and support is 14 % (3/22). Therefore, all the locations that conform to this rule are composed only of generated data, which means that the accuracy is 100%, and the ratio of the total data is 14%.

Next, explanation will be given on the expansion of the rule area.
When a rough set is used for an analysis of sediment disasters, each of the created rules has a small range that can be explained, and since only a part of the database can be explained, it is often difficult to say that it is a useful rule. In such a case, it is effective to expand the rule area.
For example, when the hatching area shown in FIG. 4A is an occurrence rule, the concept of extending the rule area is that an area including a subset that is considered to be more dangerous as shown in FIG. By doing so (the concept of “above” is given to the category division), the number of occurrence points included in the rule increases, and the rule becomes versatile. The same applies to the non-occurrence rule, but in this case, the concept of “below” is given to the category division. Also in this application, it is possible to create a more versatile rule by using the extension of the rule area for the representative data extracted by the SVM.
In the above, SVM and rough set which are important concepts in explaining the disaster prevention comprehensive plan support system according to the embodiment of the present invention have been described.

Next, the disaster prevention comprehensive plan support system according to the embodiment of the present invention will be described with reference to FIGS. 5 and 6.
FIG. 5 is a block diagram of the disaster prevention comprehensive plan support system, and FIG. 6 is a flowchart showing calculation steps executed in the disaster prevention comprehensive plan support system.
In FIG. 5, the disaster prevention comprehensive plan support system 1 generally includes an input unit 2, an SVM analysis unit 3, a rough set analysis unit 6, an analysis database 4, a database 5, and an output unit 7.
The input unit 2 stores the data shown in the analysis database 4 and the database 5 in advance, including the observation / inspection data 8 and the analysis condition data 9 stored in the analysis database 4 and the database 5, and stores them in a readable manner. Used to keep. Further, it is also used to directly input observation / inspection data 8 and analysis condition data 9 to the SVM analysis unit 3. Specific examples of the input unit 1 include a plurality of types of devices such as a keyboard, a mouse, a pen tablet, an optical reading device, or a receiving device that receives data from an analysis device such as a computer or a measurement device via a communication line. An apparatus that can be used properly according to the purpose is conceivable.
Further, as the output unit 7, specifically, a display device such as a CRT, liquid crystal, plasma or organic EL, a display device such as a printer device, and a transmitter such as a transmitter for transmitting to an external device, etc. Can be considered.
The SVM analysis unit 3 includes an analysis condition setting unit 10, a separation plane calculation unit 11, a risk level calculation unit 12, and a representative data extraction unit 13.
The analysis condition setting unit 10 sets analysis conditions for executing the analysis of the SVM. The analysis condition setting unit 10 reads the analysis condition data 9 directly input from the input unit 2 and sets the analysis condition data 9 as an analysis condition. It is also possible to read and set the analysis condition data 9 stored in the analysis database 4 from the input unit 2. As an analysis condition in the SVM analysis, for example, in the case of the soft margin method, it is necessary to set two parameters, that is, C, which indicates the degree to which misclassification is allowed, and r, which is the radius of Gaussian kernel (data influence degree). This is included as an analysis condition. Alternatively, a threshold value or a condition for extracting representative data is included as an analysis condition in the risk level calculation described later. That is, the analysis condition is necessary in addition to the observation / inspection data 8, the separation surface function data 14 constituting the separation surface, or the risk function data 15 for executing the risk calculation when executing the SVM analysis. The analysis condition setting unit 10 is an element for setting these parameters in the SVM analysis.

As shown in step S1 of FIG. 6, the separation plane calculation unit 11 of the SVM analysis unit 3 generates / does not generate disasters / accidents / repairs in disaster / risk points and observation / inspection objects such as facilities / equipment / buildings. Observed and inspected data as learning data 16 and the occurrence / non-occurrence of disasters / accidents / repairs at observation / inspection locations as teacher values are input to the separation plane function data 14, respectively. A separation plane that separates the possibility of occurrence of repair is calculated, and the separation plane data 17 obtained by the calculation is stored in the database 5 so as to be readable. Of course, when the separation plane data 17 is stored in the database 5 via the input unit 2 in a readable manner, the separation plane calculation unit 11 may be deleted. The learning data 16 is data observed and inspected in the observation / inspection object, but the contents of the data are numerical values observed and inspected for factors related to occurrence / non-occurrence of disasters / accidents / repairs. In the learning data 16 and the observation / inspection data 8, this coordinate axis is conceptualized as a dimension corresponding to the number of factors (n, n ≧ 2) and a multi-dimensional coordinate axis corresponding to the dimension. The learning data 16 and the observation / inspection data 8 are plotted in the coordinate space formed by, and the distance to the separation surface, which is also conceptualized in the n-dimensional space, is calculated by the risk calculating unit 12 described later. .
As described above, the learning data 16 and the observation / inspection data 8 may be the same data, or different data, that is, data obtained at different times with the same observation / inspection points / locations or different points / Any of the data obtained by observation / inspection at the location may be used.
In the present embodiment, the separation surface function data 14 is stored in advance in the analysis database 4 and the learning data 16 is stored in the database 5 in advance, and the separation surface calculation unit 11 reads out and uses them. Even if it is not stored in the database 4 or the database 5, it may be input via the input unit 2 when the separation plane calculation unit 11 executes the analysis. Hereinafter, also in the case of using data stored in the analysis database 4 or the database 5 in the present embodiment, the data may be read out from these databases and used, or the input unit 2 may be externally provided as necessary. You may make it input via.
As shown in step S <b> 2 in FIG. 6, the risk level calculation unit 12 is based on a separation plane that is obtained as a result of calculation by the separation plane calculation unit 11 or that is configured by the separation plane data 17 stored in the database 5 in advance. Then, the distance from the observation / inspection data 8 is calculated. More specifically, the separation plane data 17 and the observation / inspection data 8 are input to the risk function data 15 which is a function for calculating the distance, and the separation plane and the observation / inspection data 8 in the n-dimensional coordinate space. The distance from the plot point is calculated. This distance is stored in the database 5 so as to be readable as the risk data 18.
The concept of this distance is as described above. Then, using this distance as a risk level, the risk level for the observation / inspection data 8 is determined. In the present application, for the sake of convenience, a risk index is used as a risk index, but it is needless to say that a safety index may be used as a safety index.

Next, as shown in step S3 in FIG. 6, the representative data extraction unit 13 sets the risk level set as desired with respect to the risk level (risk level data 18) calculated by the risk level calculation unit 12. A risk level is selected based on a threshold value or a condition, and observation / inspection data 8 corresponding to the selected risk level is extracted as representative data 19. The threshold value or condition for the degree of risk is stored in the analysis database 4 in advance as the analysis condition data 9 and read and used by the representative data extraction unit 13, or the representative data extraction unit 13 via the input unit 2 at the time of analysis It is good to read it.
The threshold value or condition for the risk level for extracting the representative data 19 will be described in more detail when the embodiment is described.
The extracted representative data 19 is stored in the database 5 so as to be readable.

The rough set generation unit 22 of the rough set analysis unit 6 reads the representative data 19 obtained by the SVM analysis unit 3 from the database 5 and generates a rough set, as shown as step S4 in FIG. Specifically, first, for an n-dimensional disaster / accident / repair occurrence factor corresponding to the representative data 19, a category section for an attribute value related to this factor is provided. This category section is as described when the rough set was explained earlier. This category section is set in advance for each disaster / accident / repair occurrence factor and stored in the analysis database 4 as the analysis condition data 9 or is output at the time of analysis by the rough set generation unit 22. It is preferable to ask the user to ask about the setting of the category section via the unit 7 and to enable the input of the category section via the input unit 2. A model is formed by combining the category intervals for each factor. This is also as described for the rough set earlier. A combination of factors having a plurality of category sections is referred to as a model in the present application. Further, the contents of the representative data 19 are substituted into this model to form a rough set, and the rough set data 20 is readably stored in the database 5.
At this point, a rough set as shown in FIG. 3 has been generated.

The disaster / accident / repair rule generation unit 23 of the rough set analysis unit 6 uses the rough set data 20 generated by the rough set generation unit 22 as described in step S5 shown in FIG. A rule relating to the occurrence / non-occurrence of an accident / repair is generated, and the data is stored in the database 5 so as to be read out as the rule data 21.
More specifically, the disaster / accident / repair rule generation unit 23 calculates the certainty factor and support for each rule, which is a combination of each factor included in the model for each category section, with respect to the rough set data 20. To do. And a factor satisfying either or both of the threshold value or condition set as desired for the certainty factor and / or the threshold value or condition set as desired for the support. Search and extract category segment combinations (rules). Accordingly, whether each rule is a single rule or a plurality of rules depends on the certainty factor and / or the support threshold value or condition.
At that time, the above-mentioned rule expansion may be further executed. Also in this case, a certainty factor and / or support threshold value or condition for permitting the extension of the rule is determined in advance, stored as analysis condition data 9, and read during analysis. Good.
Alternatively, the certainty factor, the support, and the threshold value or conditions for this expansion may be input via the input unit 2 during the analysis.
When a combination (rule) of a factor that satisfies a threshold or condition for confidence and / or support and a category section is found, the rule generation unit 23 for disaster / accident / repair reads this into the database 5 as rule data 21 Store as possible.

Needless to say, each component included in the SVM analysis unit 3 and the rough set analysis unit 6 can output data to be output to the output unit 7.
Further, in the present embodiment, the system invention has been described. However, considering the system shown in FIG. 5 as a general-purpose computer and executing the flowchart shown in FIG. As an explanation of the embodiment of the program for extracting representative data 19 from the observation / inspection data 8 while the computer executes each process and generating a disaster / accident / repair rule based on the representative data 19 The actions and effects according to the embodiment of this program are the same as the actions and effects according to the embodiment of the disaster prevention comprehensive plan support system described above.

Hereinafter, examples using specific data will be described.
In this example, the disaster occurrence occurred in Hiroshima Prefecture on June 29, 1999, when debris flow disasters were concentrated in a concentrated manner, using the disaster prevention comprehensive plan support system 1 in the Hiroshima area (Asa We tried to generate a highly versatile rule from data obtained at 1,235 locations in Kita Ward, Asanan Ward, Saeki Ward). The study area is shown in FIG.
The breakdown of data acquired at 1,235 locations is 213 locations for debris flow generation basins and 1,022 locations for non-generation basins. In addition, the debris flow in this example is used as a general term for sediment movement phenomena in mountain streams including hillside collapse.
In addition, the topographic data used in the description of this example is that the debris flow dangerous mountain stream is divided into primary valleys and the factors measured from the 1 / 10,000 topographic map and the debris flow dangerous mountain stream survey are used to explain the basin characteristics of the mountain stream. A total of 17 factors shown in Table 1 were used as topographic factors, using the factors measured in the field based on the results (here, the two factors of the sediment thickness and the stream width). Here, since the rough set cannot handle continuous value data, the continuous value data is discretized into categories.
Rainfall data used hourly rainfall at 29 stations (Amedas, Hiroshima Prefecture and the Ministry of Land, Infrastructure, Transport and Tourism) in the Hiroshima area.

  As a method for obtaining the rainfall distribution, since the rain at the time of the disaster that occurred on June 29, 1999 was a localized torrential rain, a third mesh (about 1 km) for obtaining a more detailed rainfall distribution in this embodiment. The hourly rainfall (hereinafter referred to as mesh rainfall) was calculated at the position of 1 km). In addition, it has been confirmed that the peak of rainfall and the disaster occurrence time are almost the same. The maximum time rainfall recorded on June 29 is estimated as the disaster occurrence time, and the rainfall at that time is calculated in mesh units. Calculated. Therefore, as the rainfall factors used in this study, the maximum hourly rainfall and the cumulative rainfall during the maximum hourly rainfall (hereinafter referred to as hourly rainfall and cumulative rainfall, respectively) are used as the rainfall factors. The discretization process shown in Table 2 was performed with reference to the research. The category classification for the occurrence / non-occurrence of disasters as shown in Table 2 is stored in advance in the analysis database 4 as the analysis condition data 9 or is input via the input unit 2.

First, representative data was extracted from the original data (observation / inspection data 8) using the SVM analysis unit 3.
When performing an SVM analysis, a parameter study for performing an optimal analysis is required. In the SVM of the soft margin method, two parameters of C, which indicates the degree to which misclassification is allowed, and r, which is the radius of Gaussian kernel (data influence degree), must be set. In this embodiment, C = 5, 10, 50, 100, 200, 300, 400, 500, 8 cases, r = 0.1, 0.5, 1.0, 2.0, 3.0, 4.. The parameter study was conducted in 56 cases, 7 cases of 0, 5.0. In the parameter study, we verified the accuracy of the separation plane constructed in each case (how accurately data classification is performed). Here, the accuracy of the separation surface was evaluated by the hit rate defined in the following formula (4).

Table 3 shows a list of hit rates in each case as a parameter study result.
In the past research, parameters that did not reach 100% accuracy (complete separation) were used in order to ensure the versatility of the separation surface. However, in this example, it was clearly separated between disaster occurrence and non-occurrence. Therefore, a parameter set with a hit rate of 100% was adopted.
However, considering that the final goal of this embodiment is the creation of general-purpose rules, it is desirable to adopt classification results on the general-purpose separation surface as much as possible while ensuring complete separation. In SVM, data satisfying | f (x) | ≦ 1 is called a support vector, and can be considered as a data group close to the separation plane. Since a general-purpose separation plane is considered to be constructed by an analysis in which the number of support vectors is reduced as much as possible, C = 300, r = 3 corresponding to this in the parameter study results of this embodiment. Was used as an analysis parameter. This combination of C and r is also stored in the analysis database 4 as analysis condition data 9 and is read from the analysis database 4 at the time of analysis. Or you may input via the input part 2 at the time of an analysis.

Based on the f (x) value calculated by the SVM analysis executed by the SVM analysis unit 3 using the above parameters as the analysis condition data 9, it can be representative of the original data (observation / inspection data 8) in determining the disaster occurrence rule. Representative data 19 was extracted. The representative data 19 for creating the rule is considered to have higher risk as disaster occurrence data and higher safety as non-occurrence data. Thought.
Case 1 is a case where all of the data of | f (x) |> 1 is used as representative data. In this case, generated data: 30 locations and non-generated data: 710 locations are extracted as representative data. In Case 1, the extraction condition is also a threshold value of | f (x) |> 1 as it is.
Case 2 is a case in which the same number is arranged from the higher ranks of the generated data and the non-occurrence data with high risk and high safety. In this case, the data of 30 locations where the f (x) value of the generated original data is smaller than −1 and the non-occurrence data 30 locations where the f (x) value is the largest are the representative data. And extracted from each.
In case 3, as in case 2, data is extracted from the higher ranks of data with high risk and data with high safety, but the number of data to be extracted is This is a case that follows the condition of matching the ratio of the number of non-occurrence data. Here, since the number of generated data whose f (x) value is smaller than −1 is 30, non-generated data amount ratio (213 locations: 1022 locations) in the original data is maintained so that the relationship can be maintained. From the generated data, 144 points were extracted from the largest of the data in which f (x) is greater than 1, and representative data was constructed.
Hereinafter, the rough set analysis unit 6 is used for these three cases to generate rules based on the rough set.
In this embodiment, three cases of cases 1 to 3 are considered, and conditions or threshold values are set when extracting each representative data 19. It is not limited to such conditions, and depending on the contents of the observation / inspection data 8, conditions and threshold values may be set while considering appropriateness.

As described above, a rough set is used to create a rule from representative data. What is important here is the relationship between the minimum number of factors required for rough set analysis and the degree of consistency. Increasing the degree of matching increases the number of factors and complicates the rules. Also, if the degree of matching is lowered, the number of factors is reduced and the rules can be simplified, while the data quality tends to be lowered. In this embodiment, since the rule is created using the representative data 19 regarding occurrence and non-occurrence, the consistency level is set to 100% as a required level (threshold) in order to prevent the data quality from being reduced, and as few as possible among them. A rule that satisfies the condition of being composed of the number of factors was adopted as a rule.
When there are a large number of rules configured with the same number of factors under a degree of consistency of 100%, it is difficult to select an optimal combination. Therefore, the calculation is executed by adding a condition that makes a combination that includes a large number of occurrences of individual factors among the combinations having a matching degree of 100% an optimal combination. In addition, when there are many combinations with the same number of occurrences and it is not possible to narrow down the important factors, previous research has shown that it contributes greatly to the separation of occurrence / non-occurrence of disasters. A condition that prioritizes a combination that includes more “time rainfall” was set.
In the present embodiment, a combination of category intervals for each factor in the rough set is used as a rule, the certainty factor and the support are calculated, and this threshold value or condition is used as the analysis condition data 9. However, the degree of matching in the above description can also be used as a threshold value or condition for narrowing down the number of constituent elements of the rule. In this case, the threshold value or condition regarding the degree of consistency is stored in advance in the analysis database 4 as the analysis condition data 9, and is read out and used by the disaster / accident / repair rule generation unit 23, or the disaster / accident The repair rule generation unit 23 may read and use a threshold value or a condition input via the input unit 2.

  In the rule generation, the rule area is also expanded in this embodiment in order to improve versatility. Among the expanded rules, the threshold value for the certainty level is set to 90% or more (set as a ratio that allows a certain degree of misjudgment), and only those that exceed this value are created as disaster rules. In this example, only the confidence threshold is used to expand the rules. However, in consideration of the purpose and application of the analysis, the support threshold and conditions are used instead of the confidence. Alternatively, an expansion that satisfies both may be performed.

Based on the representative data 19 extracted in Case 1 in Table 4, the rough set analysis unit 6 generates rules. The results are shown on the left side of Table 5. In the factors constituting the rule, the notation of the up arrow and the down arrow in the table means below the category value or more based on the concept of extending the rule area. “*” Means that any category value is acceptable. In Case 1, the minimum number of factors was six factors: basin area, stream width, valley deposit thickness, basin length, valley depth ratio, and hourly rainfall. Only one non-occurrence rule was created. As for the content of the rule, it was a result that can be said to be non-occurrence if the basin layer thickness is 5 or less.
The results when this rule is applied to the original data (observation / inspection data 8, 1,235 locations) are shown on the right side of Table 5. Shown here is the result of adding the number of data calculated for each rule for circles 1 to 3, and circle 4 is the number of data included in each rule from all 1,235 streams (circle 1). Is subtracted. Circles 5 and 6 are the cumulative certainty and support of each rule.
As a result in this case, it was found that only one non-occurrence rule can explain the Hiroshima district 1,235 mountain stream. However, as shown in Table 1, the category value 5 of the gorge deposit thickness is the maximum category value of “4.0 ≦ GG (valve deposit thickness of 4.0 m or more)”. It is agreed that such a value is acceptable, and all the six factors can be any value for the rule as a whole. Further, it is determined that an effective rule could not be created because no generation rule could be specified. As the cause of this, the number of non-occurrence data is 710 and the number of non-occurrence data is too large for 30 occurrences of data, so it is presumed that rule creation depending on non-occurrence was made.

Case 1 has a problem that there is too much non-occurrence data. In case 2, 30 locations of the number of data with f (x) value of -1 or less are extracted as generated data from the SVM results, and the same number of 30 locations with large f (x) values are also generated for non-occurring data The rule was created by using 60 points of data extracted as a representative data. The results are shown on the left side of Table 6.
The minimum number of factors in Table 6 is the steepest slope, hourly rainfall, and cumulative rainfall, and a total of four rules were created with two non-occurrence rules and two occurrence rules.
The result of applying this rule to the original data is shown on the right side of Table 6. Here, four rules are created. Circles 1 to 3 indicate the number of data when each rule is accumulated. For example, the data number 185 of the circle 1 in the rule 4 includes the data 107 of the rule 3. In other words, if only the data of rule 4 is considered, there are 185-107 = 78 locations. The overall confidence and support of the four rules created are shown in Rule 9, which is the last line. In this case, the confidence of the four rules is 31.0%, and support is 97.1%. It becomes.
Unlike Case 1, both non-occurrence rules and occurrence rules were created, but the non-occurrence support when applied to the original data (observation / inspection data 8) is low at 15.0% (see Circle 6). This is because “96.2% of the 185 data corresponding to the non-occurrence rule is considered safe, but only 15.0% (185) of the corresponding data out of the total 1,235. Although there was a rule that could explain a lot of data, the cumulative certainty level was low at 31.9% and it was not reliable, and the goal of “creating a rule with high confidence level and high support” was achieved. The result was not made.
Therefore, it was determined from this case 2 that a typical rule could not be created. The reason for this is considered to be that the non-occurrence representative data became too small because the non-occurrence data was made the same number as the 30 occurrence data with respect to the 30 occurrence data.

In order to extract the representative data 19 from the problem regarding the number of data in Case 1 and Case 2, a reduction method that does not change the quality of the entire database is required. In case 3, representative data is extracted in a form that maintains the balance of the quantity ratio of generated / non-generated data in the original data (observation / inspection data 8). The rule generation result is shown on the left side of Table 7. As shown here, the minimum number of factors in Case 3 is five factors: water system pattern, steepest slope, main stream length, stream width, and hourly rainfall, total of three non-occurrence rules and three occurrence rules. Six rules were created.
However, the water system pattern is “*” in all rules, and the actual minimum number of factors is four.
The results when this rule is applied to the original data (observation / inspection data 8) are shown on the right side of Table 7. As shown here, the non-occurrence support is 71.4 %%, and the non-occurrence data with a large number is actually well captured. In addition, although the total support of non-occurrence and occurrence is 84.8%, which is slightly lower than 97.1% of case 2, the confidence of case 3 is 81.9, and 31.9 of case 2 Rules with much higher quality than% are created.
From the above, it is possible to calculate the percentage of data where the f (x) value is -1 or less in the SVM result, apply this to non-occurrence, extract representative data, and create a rule for more accuracy. It is thought that rule creation with a high balance was possible.

Here, the contents of the rules created in Study Case 3 (see Table 8) will be considered. Of the created rules, the non-occurrence of debris flow estimated from the non-occurrence rule is as follows: `` The steepest slope is 20 degrees or less, the main stream length is 0.20 km or less, the stream width is as small as 2.0 m, and the hourly rainfall The mountain stream that occurs is “the mountain stream with the steepest slope of 30 degrees or more, the main stream length of 0.60 km or more, and the hourly rainfall of 60 mm / hr or more”.
Considering the factors, if the slope of the gorge is 20 degrees or less, it is appropriate as a non-occurrence factor because there are few landslide slopes in the upstream area, and conversely, if it is 30 degrees or more, it is a sediment production source. The occurrence area increases and the risk of debris flow is very high. This is consistent with the fact that the incidence increases at 30 degrees or higher even when the actual steepest slope in the slope shown in FIG. 8 is seen. As for the length of the main stream, the longer the extension, the greater the amount of unstable sediment on the bed and the higher the risk of debris flow. Especially when the main stream length is 0.60km or less, as shown in Fig. 9, the incidence is about 15%, but at 0.60km or more, the incidence is increasing to nearly 30%. . As for the mountain stream width, the occurrence rate is low at 2.0 m or less from Fig. 10, which is appropriate as a non-occurrence rule item. The same applies to rainfall. As shown in FIG. 11, as the amount of rainfall increases, the risk of occurrence increases, which is appropriate as an item for occurrence or non-occurrence rules.
Table 8 shows a comparison of this example with the results of previous studies in which the extraction of important factors and the generation of rules for the occurrence and non-occurrence of debris flows using rough sets in the Hiroshima area. As shown in Table 8, the water system pattern, steepest slope, basin width, basin length, valley depth ratio, number of zero-order valleys and maximum hourly rainfall are listed as important factors. The four important factors extracted in the example are the four factors: water system pattern, steepest slope, basin width, and maximum hourly rainfall. The results of the important factors resulting from the disaster are similar. It was. Also from this, it is considered that the important factors extracted by the system having both the SVM analysis unit 3 and the rough set analysis unit 6 are appropriate.

In addition, according to Table 8, in the past research, the data of all the factors (19 factors) were used to create a large number of rules with 136 occurrence rules and 250 non-occurrence rules. However, it was not practical. However, in the present embodiment, the rule creation was limited to the representative data 19 and the number of rules was successfully reduced to three generated rules and three non-generated rules.
Comparing the maximum value of support for one rule, it was 4.0% non-occurrence and 0.5% occurrence in the past study, but in case 3 of this example, it was 65.3% non-occurrence and 4.8% occurrence. It is recognized that the property is greatly improved. In addition, comparing the non-occurrence data with the non-occurrence rule and the maximum number of occurrence data that could be matched with the occurrence rule, the previous study showed 49 non-occurrence and 6 occurrences. There were 717 non-occurrence and 31 occurrences, confirming that the quality of rule conformance was high.

In the above, the Example which produced | generated the rule about the occurrence form of a debris flow disaster using specific data for the disaster prevention comprehensive plan support system was demonstrated.
Next, in this second embodiment, road inspection data of the Osaka Loop Line on behalf of buildings and facilities such as roads, water and sewage systems, dams, harbors, sabos, bridges, steel towers, power plants, buildings, etc. Was used in the disaster prevention comprehensive plan support system 1 to try to generate highly versatile rules regarding the necessity of repair (occurrence / non-occurrence).
The acquisition year of observation and inspection data used is 1985, 1987, 1990, 1992, 1995, 1999, 2004. The data from 1985 to 1999 was used as learning data. There are nine factors related to the occurrence of repair in this example (bolt loosening, bolt deficiency, abnormal noise, drainage pipe, that is, drainage pipe damage, water leakage, water stop work, expansion / contraction body damage, rust, Corrosion) is used. In addition, for the teacher value, repair history (occurrence / non-occurrence of repair) is used.
Here, there are many data with different teacher values despite the inspection results of the same factors, and there are obvious contradictions. Therefore, all of these occurrence / non-occurrence data were deleted. There were 1152 learning data (occurrence 292, non-occurrence 860), but by processing obvious contradiction data, the deleted learning data became 327 (occurrence 132, non-occurrence 195).
First, a C and r parameter study on SVM for use as analysis condition data 9 was performed in advance. The results of this parameter study are shown in Table 9.
Among these parameter combinations, C = 400, r = 2 was used as a combination that can extract a large amount of representative data 19 while maintaining 93.0%, which has the highest separability regarding the necessity of repair.

In the separation by the SVM analysis unit 3, a value for which the f (x) value determined to be particularly dangerous is smaller than −1 is adopted as the threshold value for the risk level, and the corresponding data is the risk level calculation unit. According to 12, it was 24 places. Furthermore, the condition that the non-occurrence data is selected so as to maintain the quantitative ratio between the non-occurrence data (195 places) and the occurrence data (132 places) is adopted. When calculated, it was 24 * 195/132 = 35 locations.
Therefore, the representative data 19 extracted by the representative data extracting unit 13 in the present embodiment is 24 occurrences and 35 non-occurrence locations, and rules were created using these data and the rough set analysis unit 6.
The rules generated by the rough set analysis unit 6 were performed to verify whether the 2004 observation / inspection data could be predicted. The results are shown in Table 10.
Here, some of the data had different inspection items depending on the observation / inspection points, so missing values were observed. Therefore, 0 in the table indicates a missing item, and the rule is completed when the data is complemented.

As a result of the verification, it was found that 157 locations of all data in 2004 can be explained with 5 factors and 2 rules. The cumulative certainty in the two rules is 95.5% and the support is 100%, and it is considered that a high-quality and versatile rule could be created.
Here, as a result of comparison of how much difference (number of rules, number of factors, certainty, support, etc.) occurs when using a conventional technique that creates a rule only for a rough set without extracting representative data 19 Is shown in Table 11.
When a rule is created using only a rough set, the confidence factor is 100% and support is 14.6% using all nine factors as shown on the left side of Table 11. It became clear that there was a problem with sex. On the other hand, when using the representative rule by SVM, as shown on the right side of Table 11, the confidence level is 95.5% and the support is 100% while maintaining the same quality as the analysis result of only the rough set. We were able to greatly improve. Also from this, the combination of the extraction of the representative data 19 by the SVM analysis unit 3 and the rule generation in the rough set analysis unit 6 in the disaster prevention comprehensive plan support system 1 according to the present embodiment reduces the factors and the number of rules, while reducing the number of rules. It was confirmed that it is effective for improving quality (accuracy) and versatility.
In addition, this time, the disaster prevention comprehensive plan support system was used to generate rules for the occurrence and non-occurrence of debris flows in Example 1 and the rules for necessity of repair of the Osaka Loop Line in Example 2. Sediment disasters that occur on slopes and mountain streams, natural disasters such as disasters caused by earthquakes and volcanic activities, river disasters, roads and waterworks, dams, ports, sabo facilities and equipment, bridges and steel towers, If there is observation / inspection data for factors related to the occurrence of disasters or accidents or the necessity of repairs for accidents or repairs associated with aging or natural disasters in power plants, buildings, etc., SVM By executing the analysis unit and the rough set analysis unit in combination, it is possible to generate a rule excellent in accuracy and versatility. Accordingly, these may be used alone for planning evaluation and support such as disaster prevention, accident prevention or repair selection, or may be used as a comprehensive system by combining them.

As described above, according to the disaster prevention comprehensive plan support system according to the present embodiment, a combination of the SVM analysis unit 3 and the rough set analysis unit 6 is used to set a certain threshold or condition in the observation / inspection data 8. Extracted below to create representative data 19, which is used to generate disaster / accident / repair rules by the rough set analysis unit 6 to generate highly accurate and versatile disaster / accident / repair rules. It is possible to do.
Occurrence of accidents and repairs due to natural disasters, aging of facilities, structures, equipment, etc. is not always theoretically possible, and regular observation and inspection are essential. There are many items to be inspected, and the accuracy cannot be denied depending on the experience and knowledge of the inspector and the state of the observation / inspection points. Therefore, the credibility of all data is not always at a constant level, and noise is often included.
Therefore, not all observation / inspection data is used to acquire data from observation / inspection of natural disasters and the aforementioned facilities, structures, equipment, etc., and evaluate their occurrence / non-occurrence, It is important to make a selection.
In the disaster prevention comprehensive plan support system and its program according to the embodiment of the present invention, all observation / inspection data are not adopted equally, but are analyzed in two stages, that is, data selection is performed by SVM. By implementing using the analysis unit 3 and then generating rules using the rough set analysis unit 6, the accuracy is improved and high versatility is also exhibited.
The ability to generate highly accurate and versatile rules for disasters, accidents, and repairs can improve the reliability of assessments regarding the necessity of disaster prevention, accident prevention, or repairs. Support for promoting efficiency is possible. Therefore, it is possible to ensure and maintain the safety of local residents against disasters and accidents.

  It has a wide range of uses such as disaster prevention planning in public institutions such as local governments and disaster prevention centers, planning work related to the necessity of accident prevention or repair, and hazard map creation. It is also expected to be used as a teaching material for disaster prevention and accident prevention and evacuation drills at educational institutions, etc., and for private companies operating construction and civil engineering businesses, they will discover the needs of disaster prevention and maintenance management businesses and make proposals It can be used as a tool for business or as a shared tool for cooperation with public institutions, and can also be applied to research and development and design projects related to corporate disaster prevention and maintenance technologies.

It is a conceptual diagram which shows typically the basic concept of SVM. It is a two-dimensional image figure expressing the basic concept about the distance and risk with respect to the observation / inspection data by SVM. It is a conceptual diagram which shows typically the basic concept of a rough set. (A) And (b) is a conceptual diagram for demonstrating expansion of a rough set. It is a structure figure of the disaster prevention comprehensive plan support system concerning this embodiment. It is a flowchart which shows the process of the calculation performed in the disaster prevention comprehensive plan assistance system which concerns on this Embodiment. It is a conceptual diagram which shows the area which acquired the observation and inspection data used in the Example. It is a graph which shows the relationship between the steepest bed slope in the observation and inspection data used in the Example, and a debris flow. It is a graph which shows the relationship between the mountain stream width in the observation and inspection data used in the Example, and a debris flow. It is a graph which shows the relationship between the main stream length and the debris flow in the observation and inspection data used in the Example. It is a graph which shows the relationship between the hourly rainfall in the observation and inspection data used in the Example, and debris flow.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Disaster prevention comprehensive plan support system 2 ... Input part 3 ... SVM analysis part 4 ... Analysis database 5 ... Database 6 ... Rough set analysis part 7 ... Output part 8 ... Observation / inspection data 9 ... Analysis condition data 10 ... Analysis condition setting part DESCRIPTION OF SYMBOLS 11 ... Separation surface calculation part 12 ... Risk level calculation part 13 ... Representative data extraction part 14 ... Separation surface function data 15 ... Risk level function data 16 ... Learning data 17 ... Separation surface data 18 ... Risk level data 19 ... Representative data 20 ... Rough set data 21 ... Rule data 22 ... Rough set generation unit 23 ... Disaster / accident / repair rule generation unit

Claims (4)

Causes of occurrence of disasters / accidents / repairs at the locations subject to observation / inspection (the number of factors is n, where n ≧ 2) corresponds to the coordinate axes constituting the n-dimensional coordinate space, and causes each occurrence of disaster, accident, or repair. N-dimensional observation / inspection data observed and inspected as learning data, and using the occurrence / non-occurrence of disasters / accidents / repairs at the observation / inspection locations as teacher values, occurrence / non-occurrence of disasters / accidents / repairs Analyzing the possibility of occurrence using a support vector machine (hereinafter referred to as SVM) and forming the boundary of occurrence / non-occurrence of disasters / accidents / repairs, and observation / inspection at the locations subject to observation / inspection The distance from the coordinate point in the n-dimensional coordinate space indicated by the n-dimensional observation / inspection data to the separation plane (however, this distance) Is said seat The case where the point is on the origin side of the n-dimensional coordinate space with respect to the separation plane may be positive, or the case where the point is on the origin side may be negative). A risk calculation unit that calculates the accident / repair occurrence risk,
Based on a desired threshold value or condition for the disaster / accident / repair occurrence risk, the corresponding disaster / accident / repair occurrence risk is extracted, and the extracted disaster / accident / repair occurrence risk is extracted. A representative data extracting unit for extracting, as representative data, n-dimensional observation / inspection data for each factor corresponding to
For the n-dimensional disaster / accident / repair occurrence factor, a category interval for the attribute value related to this factor is provided, a model is formed by combining the category intervals for each factor, and the representative data is substituted into this model. A rough set generator for forming a rough set;
In the rough set, a combination of category intervals for each factor is generated as a rule, and a certainty level or condition for the certainty factor and / or a certainty factor of disaster, accident, and repair for each rule is calculated. A disaster / accident / repair rule generation unit that extracts the rule based on a desired threshold or condition for the support; and
Disaster prevention comprehensive plan support system characterized by having. Causes of occurrence of disasters / accidents / repairs at the locations subject to observation / inspection (the number of factors is n, where n ≧ 2) corresponds to the coordinate axes constituting the n-dimensional coordinate space, and causes each occurrence of disaster, accident, or repair. N-dimensional observation / inspection data observed and inspected as learning data, and using the occurrence / non-occurrence of disasters / accidents / repairs at the observation / inspection locations as teacher values, occurrence / non-occurrence of disasters / accidents / repairs A separation plane computing unit that analyzes the separation plane that forms the boundary of occurrence / non-occurrence of disaster / accident / repair using a support vector machine (hereinafter referred to as SVM),
Using the separation surface analyzed by the separation surface calculation unit and the n-dimensional observation / inspection data for each factor observed and inspected at the observation / inspection target location, the n-dimensional observation / inspection data is The distance from the coordinate point in the n-dimensional coordinate space to the separation plane (provided that the distance is positive when the coordinate point is on the origin side of the n-dimensional coordinate space with respect to the separation plane) Or may be negative if it is on the origin side). A risk level calculation unit that calculates the disaster / accident / repair occurrence risk level of the observation / inspection target part,
Based on a desired threshold value or condition for the disaster / accident / repair occurrence risk, the corresponding disaster / accident / repair occurrence risk is extracted, and the extracted disaster / accident / repair occurrence risk is extracted. A representative data extracting unit for extracting, as representative data, n-dimensional observation / inspection data for each factor corresponding to
For the n-dimensional disaster / accident / repair occurrence factor, a category interval for the attribute value related to this factor is provided, a model is formed by combining the category intervals for each factor, and the representative data is substituted into this model. A rough set generator for forming a rough set;
In the rough set, a combination of category intervals for each factor is generated as a rule, and a certainty level or condition for the certainty factor and / or a certainty factor of disaster, accident, and repair for each rule is calculated. A disaster / accident / repair rule generation unit that extracts the rule based on a desired threshold or condition for the support; and
Disaster prevention comprehensive plan support system characterized by having. A program executed by a computer to generate rules for occurrence / non-occurrence of disasters / accidents / repairs,
The computer causes the disaster / accident / repair occurrence factor (the number of factors to be n. N ≧ 2) to correspond to the coordinate axes that make up the n-dimensional coordinate space, and the disaster / accident / repair The n-dimensional observation / inspection data observed / inspected for each occurrence factor is used as learning data, and the occurrence / non-occurrence of disasters / accidents / repairs at the locations subject to observation / inspection is used as a teacher value. The separation plane that analyzes the possibility of occurrence / non-occurrence using a support vector machine (hereinafter referred to as SVM) to form the boundary of occurrence / non-occurrence of disaster / accident / repair and N-dimensional observation / inspection data for each factor observed / inspected, and the distance from the coordinate point in the n-dimensional coordinate space indicated by the n-dimensional observation / inspection data to the separation plane (however, This The distance may be positive when the coordinate point is on the origin side of the n-dimensional coordinate space with respect to the separation plane, or may be negative when the coordinate point is on the origin side). A risk calculation process to calculate the risk of occurrence of disasters, accidents and repairs at the target location;
Based on a desired threshold value or condition for the disaster / accident / repair occurrence risk, the corresponding disaster / accident / repair occurrence risk is extracted, and the extracted disaster / accident / repair occurrence risk is extracted. A representative data extraction step of extracting n-dimensional observation / inspection data for each factor corresponding to the above as representative data;
For the n-dimensional disaster / accident / repair occurrence factor, a category interval for the attribute value related to this factor is provided, a model is formed by combining the category intervals for each factor, and the representative data is substituted into this model. A rough set generation process for forming a rough set;
In the rough set, a combination of category intervals for each factor is generated as a rule, and a certainty level or condition for the certainty factor and / or a certainty factor of disaster, accident, and repair for each rule is calculated. A rule generation process for disaster / accident / repair that extracts the rule based on a desired threshold or condition for the support;
Disaster prevention comprehensive plan support program characterized by having A program executed by a computer to generate rules for occurrence / non-occurrence of disasters / accidents / repairs,
The computer causes the disaster / accident / repair occurrence factor (the number of factors to be n. N ≧ 2) to correspond to the coordinate axes that make up the n-dimensional coordinate space, and the disaster / accident / repair The n-dimensional observation / inspection data observed / inspected for each occurrence factor is used as learning data, and the occurrence / non-occurrence of disasters / accidents / repairs at the locations subject to observation / inspection is used as a teacher value. Separation plane calculation process to analyze the separation plane that forms the boundary of occurrence / non-occurrence of disaster / accident / repair using support vector machine (hereinafter referred to as SVM)
Using the separation surface analyzed by the separation surface calculation unit and the n-dimensional observation / inspection data for each factor observed and inspected at the observation / inspection target location, the n-dimensional observation / inspection data is The distance from the coordinate point in the n-dimensional coordinate space to the separation plane (provided that the distance is positive when the coordinate point is on the origin side of the n-dimensional coordinate space with respect to the separation plane) Or may be negative if it is on the origin side), a risk degree calculation step for calculating the disaster / accident / repair occurrence risk of the observation / inspection target part,
Based on a desired threshold value or condition for the disaster / accident / repair occurrence risk, the corresponding disaster / accident / repair occurrence risk is extracted, and the extracted disaster / accident / repair occurrence risk is extracted. A representative data extraction step of extracting n-dimensional observation / inspection data for each factor corresponding to the above as representative data;
For the n-dimensional disaster / accident / repair occurrence factor, a category interval for the attribute value related to this factor is provided, a model is formed by combining the category intervals for each factor, and the representative data is substituted into this model. A rough set generation process for forming a rough set;
In the rough set, a combination of category intervals for each factor is generated as a rule, and a certainty level or condition for the certainty factor and / or a certainty factor of disaster, accident, and repair for each rule is calculated. A rule generation process for disaster / accident / repair that extracts the rule based on a desired threshold or condition for the support;
Disaster prevention comprehensive plan support program characterized by having