JP7587111B2 - Model generation device and model generation method - Google Patents

Description

One aspect of the present invention relates to a model generation device.

For example, in the field of plant engineering, it is necessary to handle a large number of drawings (e.g., past drawings). For this reason, there is a demand for technology that allows efficient search for desired drawings from a large number of drawings.

As an example, Patent Document 1 discloses technology related to the user interface of a drawing search system (drawing management system). The drawing search system of Patent Document 1 accepts selection operations by the user of the equipment name and drawing type name, and searches for a group of drawings corresponding to the selection operations.

JP 2003-99484 A

According to the drawing search system of Patent Document 1, search results (group of drawings) corresponding to the above selection operation can be presented without the user having to input a search word. However, as will be described later, in the drawing search system of Patent Document 1, it is not necessarily easy for a user to quickly find a desired drawing corresponding to a target drawing (e.g., a specific drawing that is considered to be highly related to the target drawing) from the search results.

As such, conventional drawing search technology still has room for improvement in terms of ways to enhance user convenience. One aspect of the present invention aims to improve user convenience in drawing searches more than ever before.

In order to solve the above problem, a model generation device according to one aspect of the present invention is a model generation device that generates a learning model for searching at least one drawing corresponding to a target drawing from a plurality of search target drawings, and includes: a search target drawing content parameter acquisition unit that acquires search target drawing content parameters related to the description of each of the plurality of search target drawings by analyzing the plurality of search target drawings; a dissimilarity calculation unit that calculates the dissimilarity of the description between any two different search target drawings included in the plurality of search target drawings based on the search target drawing content parameters of each of the plurality of search target drawings; a determination unit that determines drawings that are not to be studied from among the plurality of search target drawings according to the dissimilarity; and a learning unit that generates the learning model based on the search target drawing content parameters of each of the plurality of search target drawings excluding the non-study drawings.

In order to solve the above problem, a model generation method according to one aspect of the present invention is a model generation method for generating a learning model for searching at least one drawing corresponding to a target drawing from a plurality of search target drawings, and includes a search target drawing content parameter acquisition step of acquiring search target drawing content parameters related to the description of each of the plurality of search target drawings by analyzing the plurality of search target drawings, a difference calculation step of calculating the difference in the description between any two different search target drawings included in the plurality of search target drawings based on the search target drawing content parameters of each of the plurality of search target drawings, a determination step of determining non-learning drawings from among the plurality of search target drawings according to the difference, and a learning step of generating the learning model based on the search target drawing content parameters of each of the plurality of search target drawings excluding the non-learning drawings.

According to one aspect of the present invention, it is possible to improve user convenience in drawing searches compared to the conventional method.

FIG. 1 is a block diagram showing a configuration of a main part of an information processing system according to a reference embodiment. FIG. 11 is a diagram showing an example of a first table in the reference embodiment. 11A and 11B are diagrams illustrating an example of a specific character string extraction process for a past drawing. FIG. 13 is a diagram showing an example of a second table in the reference embodiment. FIG. 13 is a diagram showing an example of a third table in the reference embodiment. 11A and 11B are diagrams illustrating an example of a specific character string extraction process for a new drawing. 13A and 13B are diagrams illustrating an example of deriving a relevance score in a reference embodiment. FIG. 13 is a diagram showing an example of a drawing search result in the reference embodiment. 1 is a block diagram showing a configuration of a main part of an information processing system according to a first embodiment. FIG. 4 is a diagram illustrating a degree of difference in the first embodiment. FIG. 4 is a diagram illustrating an example of a dissimilarity table according to the first embodiment. 6A to 6C are diagrams illustrating an example of a process of a similar set determination unit in the first embodiment. FIG. 2 is a diagram showing an example of a dendrogram representing a hierarchical structure of a set obtained in the first embodiment. 11A to 11C are diagrams illustrating an example of processing by a non-learning drawing determination unit in the first embodiment. FIG. 11 is a diagram showing an example of an edited second table in the first embodiment.

[Reference form]
Prior to the description of the information processing system 100 of the first embodiment, an information processing system 100s will be described as a reference embodiment. For convenience of description, the same reference numerals will be used in the following embodiments for components having the same functions as those described in the reference embodiment, and the description will not be repeated. For simplicity, the description of matters similar to those in the known technology will be omitted as appropriate.

Please note that each numerical value described below in this specification is merely an example. In this specification, the description "A to B" for two numbers A and B means "greater than or equal to A and less than or equal to B" unless otherwise specified.

(Overview of information processing system 100s)
1 is a block diagram showing the configuration of a main part of an information processing system 100s. The information processing system 100s includes an information processing device 1s, a past property plan DB (Database) 91, and a new property plan DB 92.

The information processing device 1s includes a control device 10s, an input unit 71, a display unit 72, and a memory unit 80. The control device 10s includes a learning device 11s and a drawing search device 12. In this specification, "property" refers to, for example, a "site" in plant engineering.

The information processing device 1s only needs to be communicably connected to the past property plan DB 91 and the new property plan DB 92. Therefore, unlike the example of FIG. 1, at least one of the past property plan DB 91 and the new property plan DB 92 may be provided inside the information processing device 1s.

Drawings to be searched for by the information processing device 1s (more specifically, the control device 10s) (search target drawings) include, for example, specification drawings, design drawings, and production drawings. Search target drawings may also include specifications, design documents, and estimates. In this way, search target drawings are not limited to documents that include the name "drawing." As an example, search target drawings include any type of document related to project plans in the plant engineering field.

The control device 10s comprehensively controls each part of the information processing device 1s. The memory unit 80 stores various data and programs used in the processing of the control device 10s. As described below, the control device 10s uses machine learning to search for at least one drawing that corresponds to the target drawing (e.g., drawing ND) from multiple search target drawings (e.g., past drawings a1 to MN).

The input unit 71 accepts user operations (user operations). The display unit 72 displays various data. As an example, the display unit 72 may display data indicating the results of a search by the control device 10s. The input unit 71 and the display unit 72 may be provided as one unit. For example, the input unit 71 and the display unit 72 can be integrated by using a touch panel.

(Past property drawings DB91)
The past property drawing DB 91 stores drawings (strictly speaking, drawing data) for each past property (each existing property). In the following description, "drawing data of drawing A (a certain drawing)" will be abbreviated as simply "drawing A" as appropriate. Also, "drawing number of drawing A" will be abbreviated as simply "drawing A" as appropriate.

In the reference embodiment, the past property drawing DB91 stores drawings for each of a number M of different properties. M is an integer equal to or greater than 1. Hereinafter, the jth property is also referred to as "property j." j is an integer equal to or greater than 1 and equal to or less than M.

In addition, the past property drawing DB91 stores N different drawings (N types) for each of properties 1 to M. N is an integer greater than or equal to 1. Hereinafter, the i-th drawing for property j will also be referred to as "drawing (i, j)." The i-th drawing (i-th type) for each property will also be referred to collectively as drawing i. i is an integer greater than or equal to 1 and less than or equal to N.

As described above, a total of T drawings are stored in the past property drawing DB91. In the reference example, T = M x N. T is an integer equal to or greater than 2. In other words, at least one of M and N is equal to or greater than 1.

Specifically, in the past property drawing DB91, each type of drawing number is listed by property number in the form of the first table TB1 shown in Figure 2. Hereinafter, the first table TB1 is also abbreviated to "TB1". Other elements will be similarly abbreviated as appropriate. TB1 is also referred to as the "past property/drawing correspondence table sorted by property". The cell in the i-th row and j-th column of TB1 indicates the drawing number of drawing (i, j).

In the example of Figure 2, for convenience, properties 1 to 3 are also written as properties A to C, respectively. In the example of Figure 2, the first type of drawing (Drawing 1) is an outline drawing, the second type of drawing (Drawing 2) is an assembly drawing, and the third type of drawing (Drawing 3) is a foundation drawing. In addition, the Nth type of drawing (Drawing N) is a configuration drawing.

For simplicity, in the following, drawings (1,1) to (N,1) in the example of Figure 2 (i.e., each of the N types of drawings for property A) will also be referred to as drawings a1 to aN. For example, drawings a1 to aN in the example of Figure 2 refer to the exterior drawing to configuration drawing of property A, respectively. The drawings for the other properties will be described in the same way.

As described above, the past property drawing DB91 stores a total of T drawings, from drawing a1 to drawing MN. Hereinafter, drawings a1 to MN will be collectively referred to as past drawings. Past drawings are an example of drawings to be searched. For this reason, drawings a1 to MN will also be referred to as a group of drawings to be searched.

In addition, in this specification, any one of multiple past drawings (search target drawings) is also referred to as a candidate drawing. As an example, the processing steps for drawing a1 (exterior drawing of property A) which is the candidate drawing will be mainly illustrated. The processing steps for other drawings will be omitted as appropriate, but are the same as those for drawing a1.

(Learning device 11s)
The learning device 11s includes a past drawing data acquisition unit 111, a past drawing data preprocessing unit 112 (candidate drawing content parameter acquisition unit, search target drawing content parameter acquisition unit), and a learning model generation unit 113s. The learning device 11s generates a learning model for drawing search by the drawing search device 12 based on the drawings a1 to MN. For this reason, the learning device 11s may be referred to as a model generation device. An example of the processing flow of the learning device 11s will be described below.

(Acquisition of candidate drawings)
In the past property drawing DB91, drawings a1 to MN are pre-sorted by drawing type according to TB1s. Therefore, for example, the past drawing data acquisition unit 111 acquires each exterior drawing from the past property drawing DB91 in the order of "drawing a1 → b1 → ... → M1". Next, the past drawing data acquisition unit 111 acquires each assembly drawing from the past property drawing DB91 in the order of "drawing a2 → b2 → ... → M2". And finally, the past drawing data acquisition unit 111 acquires each configuration drawing from the past property drawing DB91 in the order of "drawing aN → bN → ... → MN".

In the above example, the past drawing data acquisition unit 111 first references the cell in the first row and first column of TB1s. Then, the past drawing data acquisition unit 111 acquires the drawing (1,1) corresponding to the above cell, i.e. drawing a1, from the past property drawing DB91. The past drawing data acquisition unit 111 supplies the acquired drawing a1 to the past drawing data pre-processing unit 112.

(Generation of pre-processing data corresponding to candidate drawings)
The previous drawing data pre-processing unit 112 analyzes the drawing a1 to generate pre-processed data corresponding to the drawing a1. As an example, the previous drawing data pre-processing unit 112 performs OCR processing on the drawing a1 to generate post-OCR processing data. In the post-OCR processing data, each character represented as image data in the drawing a1 is converted into a character code.

Then, the past drawing data preprocessing unit 112 extracts (detects) a specific character string that is preset for each drawing from the post-OCR processing data. Hereinafter, the specific character string is referred to as a specific character string. In the first embodiment, it is assumed that L (L is an integer equal to or greater than 2) different kth specific character strings are preset.

In the following, the kth specific string is referred to as the kth specific string. k is an integer greater than or equal to 1 and less than or equal to L. In the following explanation, an example is given in which the first specific string is set to "voltage value", the second specific string is set to "current value", the third specific string is set to "OR", and the Lth specific string is set to "open".

However, the specific character string is not limited to the above example. The specific character string may be, for example, a word that is commonly used to explain product specification information (technical information) in drawings in the engineering field.

Figure 3 is a diagram showing an example of a specific character string extraction process for drawing a1. In the example of Figure 3, the character string "Voltage value 90V" is written in area L1 of drawing a1. In this case, the past drawing data pre-processing unit 112 extracts the series of characters "Voltage value 90V" in L1 from the data after OCR processing. The past drawing data pre-processing unit 112 further extracts the first specific character string "Voltage value" from the series of characters extracted.

Then, the past drawing data pre-processing unit 112 applies a known syntax analysis technique to the series of character strings, and determines that the character string "90V" following the first specific character string "Voltage value" indicates the attribute of "Voltage value". In this case, the past drawing data pre-processing unit 112 acquires the number "90" contained in the character string "90V" as a value indicating the magnitude of the "voltage value". In this way, the past drawing data pre-processing unit 112 determines through syntax analysis that "the series of character strings written in L1 indicate the specification information that 'the voltage value is 90V.'"

In addition, a series of characters "current value 1.0A" is written in area L2 of drawing a1. The past drawing data pre-processing unit 112 extracts the second specific character string "current value" from the series of characters in the same manner as in the example of L1. The past drawing data pre-processing unit 112 then determines that the character string "1.0A" following the second specific character string "current value" indicates the attribute of "current value". Therefore, the past drawing data pre-processing unit 112 acquires the number "1.0" contained in the character string "1.0A" as a value indicating the magnitude of the "current value". In this way, the past drawing data pre-processing unit 112 determines that "the series of characters written in L2 indicates the specification information that 'the current value is 2.0A.'"

However, the third specific character string "OR" is written in drawing a1, but the third specific character string "OR" is not followed by any other character string. In this case, the past drawing data pre-processing unit 112 detects the number of occurrences of the third specific character string "OR" contained in the post-OCR processing data. As a result of this detection process, the past drawing data pre-processing unit 112 determines that the number of occurrences of the third specific character string "OR" contained in drawing a1 is three (see area L3).

Similarly, drawing a1 contains the Lth specific character string "open", but no characters follow the Lth specific character string "open". Therefore, the past drawing data pre-processing unit 112 determines that the number of Lth specific character strings "open" contained in drawing a1 is one, as in the example of L3 (see area L4).

Then, the past drawing data preprocessing unit 112 generates preprocessed data corresponding to drawing a1 based on the above-mentioned analysis results of the OCR processed data. The preprocessed data is a data set indicating the kth content parameter (hereinafter, Ak) associated with the kth specific character string. The first to Lth content parameters are also collectively referred to as content parameters.

The content parameter can be said to be one of the data that quantifies (quantifies) the contents of the drawing (specifically, the contents related to a specific character string). For this reason, the content parameter is used as an index that indicates the contents of the drawing. Note that, to distinguish from the target drawing content parameters described later, the content parameters of the drawing to be searched (past drawing) are also referred to as the search target drawing content parameters. In addition, the kth content parameter of the drawing to be searched is also referred to as the kth content parameter of the drawing to be searched.

In the following, Ak of drawing (i, j) is also written as Ak(i, j). As described above, the past drawing data preprocessing unit 112 sets Ak(i, j) based on the analysis results for drawing (i, j) (more specifically, the detection results of a specific character string in drawing (i, j)).

As an example, the past drawing data pre-processing unit 112 sets A1=90 for drawing a1 (see also the hatched area in FIG. 4 described below). That is, the past drawing data pre-processing unit 112 sets the value "90" indicating the magnitude of the first specific character string "voltage value" written on drawing a1 as the first content parameter.

Similarly, the past drawing data pre-processing unit 112 sets A2 = 1.0. In other words, the past drawing data pre-processing unit 112 sets the value "1.0" indicating the magnitude of the second specific character string "current value" written on drawing a1 as the second content parameter.

In response to this, the past drawing data pre-processing unit 112 sets A3 = 3. In other words, the past drawing data pre-processing unit 112 sets the value "3" indicating the number of times the third specific character string "OR" is written (appears) in drawing a1 as the third content parameter.

Similarly, the past drawing data preprocessing unit 112 sets AL = 1. That is, the past drawing data preprocessing unit 112 sets the value "1" indicating the number of times the Lth specific character string "open" is written in drawing a1 as the Lth content parameter.

As described above, the past drawing data pre-processing unit 112 sets A1 to AL for drawing a1. As described above, the past drawing data pre-processing unit 112 acquires the content parameters of the candidate drawing by analyzing the candidate drawing. For this reason, the past drawing data pre-processing unit 112 is also called the candidate drawing content parameter acquisition unit.

The past drawing data pre-processing unit 112 also performs similar processing on other past drawings. That is, the past drawing data pre-processing unit 112 sets A1 to AL for each of the drawings a1 to MN.

Then, the past drawing data preprocessing unit 112 generates a second table TB2 that shows A1 to AL of each of the drawings a1 to MN, i.e., A1(1,1) to AL(M,N). As described below, TB2 is an example of teacher data in the reference form. In other words, the past drawing data preprocessing unit 112 also operates as a functional unit that generates teacher data (teacher data generation unit).

TB2 contains the ith sub-table TB2-i within the second table. TB2-i is a table showing A1 to AL for each of the drawings i. Figure 4 shows TB2-1 as an example. TB2-1 shows A1 to AL for each of the drawings a1 to M1 (exterior drawings of properties A to M). TB2s is made up of a set of N sub-tables, TB2-1 to TB2-N. In this way, in the reference form, a sub-table is created for each drawing type.

(Generation of learning models in reference form)
The learning model generation unit 113s generates a learning model corresponding to the drawings a1 to M1 based on the pre-processed data corresponding to the drawings a1 to M1 (i.e., A1 to AL in the drawings a1 to M1). A known machine learning algorithm may be used to generate the learning model. As an example, in the first embodiment, the learning model generation unit 113s generates the learning model by multinomial logistic regression analysis.

As described above, TB2 indicates the number (an example of an identifier) of each of multiple past drawings (e.g., drawing a1) and the correspondence between A1 to AL of each of the multiple past drawings. Therefore, the learning model generation unit 113s acquires TB2 as training data. The learning model generation unit 113s then generates a learning model in the reference form by performing a multinomial logistic regression analysis using the training data.

As an example, the learning model generation unit 113s calculates the kth weighting coefficient (hereinafter, Bk) associated with the kth specific character string in drawing a1 based on each of A1 to AL in TB2-1. The first to Lth weighting coefficients are also collectively referred to as weighting coefficients. Furthermore, Bk in drawing (i, j) is also written as Bk(i, j). Bk(i, j) is the weighting coefficient corresponding to Ak(i, j). The weighting coefficient is used to calculate the relevance score, which will be described later.

For example, the learning model generation unit 113s calculates B1 to BL for drawing a1 using the above-mentioned multinomial logistic regression analysis. In the reference embodiment, the weighting coefficients for drawing a1 are calculated as B1 = -0.5, B2 = -1.2, B3 = 0.6, ..., BL = 0.7 (see the hatched areas in Figure 5 described below).

The learning model generation unit 113s performs similar processing for other past drawings. That is, the learning model generation unit 113s sets B1 to BL for each of drawings a1 to MN. In this way, the learning model generation unit 113s generates a learning model that includes B1 to BL for each of drawings a1 to MN, i.e., B1(1,1) to BL(M,N).

Then, the learning model generation unit 113s generates a third table TB3 showing B1(1,1) to BL(M,N). TB3 includes an i-th sub-table TB3-i within the third table. TB3-i is a table showing B1 to BL of each of the drawings i. FIG. 5 shows TB3-1 as an example. TB3-1 shows B1 to BL of each of the drawings a1 to M1. In this way, TB3-1 shows weighting coefficients corresponding to the content parameters of TB2-1. TB3 is composed of a set of N sub-tables, TB3-1 to TB3-N.

(New property drawing DB92)
In the new property drawing DB 92, drawings (hereinafter collectively referred to as new drawings) relating to new properties (e.g., at least one property that is scheduled to be constructed) are stored as a new property data set. In the following, one new property (property T) included in the new property data set will be described.

As an example, new property drawing DB92 stores N different drawings (outline drawing to configuration drawing) for property T that are of the same type as previous properties. Below, the outline drawing for property T is given as an example of a new drawing. The outline drawing for property T is also referred to as drawing ND (see also Figure 6 below).

(Drawing search device 12)
The drawing search device 12 includes a new drawing data acquisition unit 121, a new drawing data preprocessing unit 122 (target drawing content parameter acquisition unit), a score calculation unit 123, and a search result data generation unit 124. The score calculation unit 123 and the search result data generation unit 124 may be collectively referred to as a search unit.

In the reference embodiment, the drawing search device 12 searches for at least one drawing by matching the target drawing with each of the multiple search target drawings using the learning model generated by the learning device 11s. Drawing ND in the reference embodiment is an example of a target drawing. As described below, the drawing search device 12 uses the learning model to match drawing ND with each of drawings a1 to MN.

(Acquisition of new drawings)
The new drawing data acquisition unit 121 is a functional unit paired with the past drawing data acquisition unit 111. As an example, the new drawing data acquisition unit 121 acquires a predetermined new drawing (e.g., drawing ND) included in the new property data set of the new property drawing DB 92 when the input unit 71 accepts a predetermined user operation. The new drawing data acquisition unit 121 supplies the acquired drawing ND to the new drawing data pre-processing unit 122.

(Generating pre-processing data for new drawings)
The new drawing data pre-processing unit 122 is a functional unit paired with the previous drawing data pre-processing unit 112. The new drawing data pre-processing unit 122 generates pre-processed data corresponding to the drawing ND by the same process as the previous drawing data pre-processing unit 112. That is, the new drawing data pre-processing unit 122 performs OCR processing on the drawing ND and extracts a specific character string from the drawing ND. Then, the new drawing data pre-processing unit 122 sets content parameters of the drawing ND based on the extraction result of the specific character string. Hereinafter, the k-th content parameter of the drawing ND is also referred to as Ck.

That is, the new drawing data pre-processing unit 122 obtains the content parameters of the drawing ND by analyzing the drawing ND using the same analysis method as the past drawing data pre-processing unit 112. Note that, to distinguish from the above-mentioned search target drawing content parameters, the content parameters of the target drawing (drawing ND) are also referred to as target drawing content parameters. Also, the kth content parameter of the target drawing is also referred to as the kth content parameter of the target drawing.

Figure 6 is a diagram showing an example of a specific character string extraction process for a drawing ND. Figure 6 is a diagram paired with Figure 3. In the example of Figure 6, the character string "Voltage value 50V" is written in region R1 of drawing ND. Therefore, the new drawing data pre-processing unit 122 sets C1 = 50 based on the result of extraction of the first specific character string (see also Figure 7 described below).

The position of region R1 in drawing ND is different from the position of region L1 in drawing a1 (see Figures 3 and 6). However, although the description position is different in drawing ND from that in drawing a1, the same type of specification information (voltage value) is described as in drawing a1. According to the method of setting the first content parameter by the new drawing data pre-processing unit 122, the consistency of the description content (specification information related to the first specific string) between drawing ND and drawing a1 can be quantitatively evaluated regardless of the apparent difference between drawing ND and drawing a1 (e.g., the difference in the description position of the first specific string). In this regard, the same applies to C2 to CL (in other words, regions R2 to R4) described below.

In the example of FIG. 6, the character string "current value 2.5A" is written in region R2 of drawing ND. Therefore, the new drawing data preprocessing unit 122 sets C2=2.5 based on the extraction result of the second specific character string (see FIG. 6).

Furthermore, the new drawing data pre-processing unit 122 determines that the number of occurrences of the third specific character string "OR" contained in the drawing ND is three (see region R3 in FIG. 6). Therefore, the new drawing data pre-processing unit 122 sets C3=3.

Similarly, the new drawing data pre-processing unit 122 determines that the number of occurrences of the Lth specific character string "kai" contained in the drawing ND is two (see region R4 in FIG. 6). Therefore, the new drawing data pre-processing unit 122 sets CL=2.

As described above, the new drawing data pre-processing unit 122 sets C1 to CL by performing the same processing on the drawing ND as the past drawing data pre-processing unit 112. After that, the new drawing data pre-processing unit 122 generates a new drawing content parameter table TB-ND indicating C1 to CL (see FIG. 7).

(Calculation of relevance score)
The score calculation unit 123 calculates a relevance score based on the learning model (more specifically, B1(1,1) to BL(M,N)) generated by the learning model generation unit 113s. The relevance score is a score (index) indicating the degree of relevance of each past drawing (drawings a1 to MN) to drawing ND. Hereinafter, the relevance score of drawing (i,j) is also written as SCORE(i,j). SCORE(i,j) is also abbreviated as SCORE.

Below, an example of calculating the relevance score is shown with reference to FIG. 7 (particularly the hatched areas in FIG. 7). For ease of explanation, the following example shows a case where the relevance score is calculated for each of the previous drawings of the same type as drawing ND (i.e., drawings a1 to M1, which are outline drawings). However, please note that the relevance score is also derived for previous drawings of a different type from drawing ND, in the same way as in the example of FIG. 7.

Figure 7 is a diagram explaining an example of deriving a relevance score. Figure 7 mainly describes an example of calculating the SCORE of drawing a1. The score calculation unit 123 calculates the first raw score to the Lth raw score of drawing a1 using (i) C1 to CL derived by the new drawing data preprocessing unit 122 and (ii) B1 to BL of drawing a1 derived by the learning model generation unit 113s.

Hereinafter, the kth raw score will also be written as SSk. The kth raw score of drawing (i, j) will also be written as SSk(i, j). The first raw score to the Lth raw score will also be collectively referred to as raw scores. SSk corresponds to Bk (the kth weighting coefficient).

Specifically, the score calculation unit 123 calculates the product of the k-th content parameter (Ck) of the drawing ND and the k-th weighting coefficient (Bk) of the drawing a1 as SSk.
SSk=Ck×Bk…(1)
Calculate.

In the example of FIG. 7, C1=50, B1=-0.5. Therefore, the score calculation unit 123 derives SS1 as SS1=50×-0.5=-25. In a similar manner, the score calculation unit 123 calculates each subsequent raw score. For example, the score calculation unit 123 calculates each raw score as SS2=-3, SS3=1.8, ..., SSL=1.4.

Next, the score calculation unit 123 calculates a relevance score (SCORE) of the drawing a1 based on the first raw score to the Lth raw score of the drawing a1. Specifically, the score calculation unit 123 calculates the sum of each raw score from the first raw score to the Lth raw score as the SCORE. That is, the score calculation unit 123 calculates the relevance score (SCORE) of the drawing a1 based on the first raw score to the Lth raw score of the drawing a1.
SCORE=ΣSSk…(2)
The range of k on the right side of the formula (2) is from 1 to L. This also applies to the formula (2A) etc. described later.

As described above, the score calculation unit 123 calculates SCORE using each Ck and each Bk. Hereinafter, the relevance score of drawing a1 will also be referred to as SCORE(a1). The relevance scores of other past drawings will be similarly represented. In the example of Figure 7, the result of calculation by the score calculation unit 123 was SCORE(a1) = -30.4.

Then, the score calculation unit 123 uses a similar method to calculate the SCORE of each of the remaining past drawings (drawings b1 to M1 in the example of FIG. 7). In the example of FIG. 7, the results of calculations by the score calculation unit 123 are SCORE(b1) = 41.1, SCORE(c1) = 36.5, ..., SCORE(M1) = 4.2.

In the example of FIG. 7, it was confirmed that the maximum relevance score of the M SCOREs, SCORE(a1) to SCORE(M1), was SCORE(b1). It was also confirmed that SCORE(c1) was the second largest of the M SCOREs. It was also confirmed that the minimum relevance score of the M SCOREs was SCORE(a1).

The score calculation unit 123 generates a fourth table TB4 that indicates its own calculation results. TB4 includes an i-th sub-table TB4-i within the fourth table. TB4-i is a table that indicates SS1 to SSL and SCORE for each of the drawings i. FIG. 7 shows TB4-1 as an example. TB4-1 shows SS1 to SSL and SCORE for each of the drawings a1 to M1. In this way, TB4-1 is generated based on TB-ND and TB3-1. TB4 is composed of a set of N sub-tables, TB4-1 to TB4-N.

In addition, the above formula (1) is
SSk(i,j)=Ck×Bk(i,j)...(1A)
In this way, the score calculation unit 123 calculates SSk(i,j) based on Bk(i,j).

Moreover, the above formula (2) is
SCORE (i, j) = ΣSSk (i, j)
=Σ{Ck×Bk(i,j)}...(2A)
In this way, the score calculation unit 123 calculates SCORE(i,j) based on each SSk(i,j) (i.e., each Ck and each Bk(i,j)).

Moreover, the above formula (2A) is
SCORE(i,j)=Σ{Ak(i,j)×Bk(i,j)}...(3)
In this way, the score calculation unit 123 can calculate the SCORE, which is a dependent variable (objective variable), by applying C1 to CL as independent variables (explanatory variables) in the learning model.
The right side of equation (3) is an example of a linear model in multinomial logistic regression analysis (multinomial logistic regression analysis model).

In addition, SSk can also be expressed as the output of the learning model corresponding to Ck. From this, it can also be expressed that the score calculation unit 123 calculates SCORE based on the outputs of the learning model (SS1 to SSk).

Alternatively, the learning model may be configured to output a SCORE. In this case, the score calculation unit 123 obtains the SCORE as an output of the learning model and supplies the SCORE to the search result data generation unit 124.

(Decision of the ranking of candidates)
The search result data generating unit 124 determines a candidate order (candidate ranking) corresponding to the drawing ND based on the SCORE(a1) to SCORE(M1) calculated by the score calculating unit 123. Specifically, the search result data generating unit 124 determines a candidate order (first candidate to Mth candidate) for each of the drawings a1 to M1 (outline drawing) based on the SCORE(a1) to SCORE(M1).

The search result data generation unit 124 sorts SCORE(a1) to SCORE(M1) according to the value of each SCORE. As an example, the search result data generation unit 124 sorts SCORE(a1) to SCORE(M1) in descending order of value. Then, the search result data generation unit 124 determines the past drawing corresponding to the gth largest SCORE as the gth candidate (hereinafter also referred to as candidate g). g is an integer greater than or equal to 1 and less than or equal to M.

By setting candidate numbers according to SCORE in this way, it is expected that the smaller the candidate number (the higher the candidate), the higher the relevance between the past drawing and drawing ND. For example, candidate 1 refers to the past drawing among drawings a1 to M1 that is shown by SCORE to have the highest relevance with drawing ND.

As described above, SCORE(b1) is the maximum relevance score (the largest relevance score) among SCORE(a1) to SCORE(M1). Therefore, the search result data generation unit 124 determines drawing b1 as candidate 1 (see the hatched area in Figure 8). In this way, the search result data generation unit 124 selects drawing b1 as the past drawing that is most highly related to drawing ND.

Furthermore, SCORE(c2) is the second largest relevance score among SCORE(a1) to SCORE(M1). Therefore, the search result data generation unit 124 determines drawing c2 as candidate 2. In this way, the search result data generation unit 124 selects drawing c2 as the past drawing with the second highest relevance to drawing ND.

Furthermore, SCORE(a1) is the smallest relevance score (the Mth largest relevance score) among SCORE(a1) to SCORE(M1). Therefore, the search result data generation unit 124 determines drawing a1 as candidate M. In this way, the search result data generation unit 124 selects drawing a1 as the past drawing with the lowest relevance (the Mth highest) to drawing ND.

The search result data generation unit 124 generates data indicating the candidate numbers of drawings a1 to M1 for drawing ND. In this data, the drawing numbers (identifiers) of drawings a1 to M1 are sorted according to SCORE. More specifically, in this data, the drawing numbers of drawings a1 to M1 are sorted in descending order of SCORE.

The search result data generation unit 124 performs the same process on past drawings of the same type as the new drawing even if the type of the new drawing is different from an outline drawing. For example, consider a case where the type of the new drawing is an assembly drawing. In this case, the search result data generation unit 124 determines the candidate numbers for each of drawings a2 to M2 (assembly drawings) for the new drawing based on SCORE(a2) to SCORE(M2). The same applies to other types of drawings.

As described above, the search result data generation unit 124 sets candidate numbers for drawings (i,1) to (i,M) for drawings ND belonging to type i based on SCORE(i,1) to SCORE(i,N). Specifically, the search result data generation unit 124 identifies SCORE(i,g), which is the gth largest SCORE among SCORE(i,1) to SCORE(i,M). Then, the search result data generation unit 124 determines the past drawing corresponding to SCORE(i,g), i.e., drawing (i,g), as candidate g.

As an example, the search result data generation unit 124 generates a fifth table TB5 as the search result data. In TB5, past drawings are sorted for each type of drawing ND in descending order of candidate number (in descending order of SCORE). An example of TB5 is shown in FIG. 8. The search result data generation unit 124 supplies TB5 to the display unit 72 and causes the display unit 72 to display the TB5.

By displaying TB5 on the display unit 72, the search results of the drawing search device 12 (more specifically, the search result data generation unit 124) can be presented to the user in a highly viewable manner.

However, the search result data generation unit 124 does not necessarily have to present the entire TB5 to the user as search result data. This is because if there are too many past drawings, presenting the entire TB5 to the user may actually reduce user convenience.

Therefore, the search result data may be at least a part of TB5. The search result data may be data indicating at least one drawing (at least one past drawing) corresponding to the target drawing (drawing ND).

As mentioned above, the drawing in candidate 1 is the drawing with the highest SCORE among multiple past drawings. Therefore, from the perspective of reducing the amount of data in the search result data, it is also possible to generate the search result data as data showing only the drawing in candidate 1.

However, when designing equipment in the field of plant engineering, designers need to make a comprehensive judgment based on various technical information. For this reason, when designing new equipment described in drawing ND using previous drawings as reference, the previous drawing with the highest SCORE (candidate 1 drawing) is not necessarily the most suitable reference for the design. For example, candidate 2 drawing (drawing with a lower SCORE) may actually be more suitable as a reference for the design than candidate 1 drawing.

Therefore, it is preferable that the search result data is data indicating the top multiple candidates (e.g., the first γ candidates). In other words, the search result data may be data indicating drawings of candidates 1 to γ, where γ is an integer greater than or equal to 1 and less than or equal to g.

(Effects of the reference form)
In plant engineering, as a project for object T progresses, it may be necessary to search for past drawings that are highly related to drawing ND (target drawing) (see also Patent Document 1). This is because in the above project, various plans (e.g., forecast of manufacturing period, trial calculation of manufacturing cost) for a specific type of equipment in object T are often made with reference to the design and manufacturing track record of the same type of equipment in past objects.

Here, if the past drawing contains similar specification information as drawing ND, it is expected that it will be highly relevant to drawing ND. However, in plant engineering, it is common for there to be a huge number of drawings for past projects. For this reason, even if the drawing search system of Patent Document 1 is used, it is not necessarily easy for a user to quickly find the desired drawing (a past drawing that is highly relevant to drawing ND) from the search results. This is because when a keyword-less search is performed with the drawing search system of Patent Document 1, it is likely that a large number of past drawings will be hit as search results.

One idea is to use a known drawing search system with a keyword search function to search for the desired drawing that is highly related to the drawing ND. For example, it is possible to search for past drawings using a specific character string written in the drawing ND as a keyword (search word). However, even when performing such a keyword search, it is not necessarily easy for the user to quickly find the desired drawing from the search results. This is because even when performing a keyword search, it is likely that a considerable number of past drawings will be hit as search results. Furthermore, if the user cannot determine an appropriate search word, it is difficult to use a keyword search in the first place.

As described above, it is difficult for users who do not have sufficient knowledge of plant engineering (e.g., new designers) to quickly search for desired drawings using conventional drawing search systems. Unless the user is well versed in plant engineering (e.g., experienced designers), it is not necessarily easy to efficiently search for desired drawings using conventional drawing search systems.

In consideration of these problems with the conventional technology, the inventors of the present application (hereinafter, the inventors) have newly created an information processing system 100s (more specifically, an information processing device 1s). The information processing device 1s can search for a desired drawing corresponding to the drawing ND using a learning model obtained by machine learning for each past drawing.

First, in the learning device 11s of the information processing device 1s, content parameters (search target drawing content parameters) related to a specific character string are set for each past drawing. Then, a weighting coefficient for each past drawing is set based on the search target drawing content parameters.

Then, in the drawing search device 12 of the information processing device 1s, a relevance score, which is an index showing the degree of relevance between the drawing ND and each of the past drawings, is calculated for each past drawing using the content parameters of the drawing ND (target drawing content parameters) and the weighting coefficient of each past drawing. Then, each past drawing is sorted in descending order of relevance score. In other words, at least one past drawing is presented to the user as a candidate for the desired drawing based on the relevance score.

In this way, the drawing search device 12 allows the user to efficiently search for the desired drawing by keyword-less search, unlike known drawing search systems that have a keyword search function. In particular, the drawing search device 12 allows the user to perform an efficient drawing search even when the user cannot determine an appropriate search word. For this reason, the drawing search device 12 can be said to be particularly suitable for improving the work efficiency of, for example, the above-mentioned new designers.

The drawing search system of Patent Document 1 is common to the information processing device 1s (more specifically, the drawing search device 12) in that it is based on keyword-less search. However, unlike the information processing device 1s, the drawing search system of Patent Document 1 does not order the search results. For this reason, in the drawing search system of Patent Document 1, past drawings that are ranked high in the search results do not necessarily have a high relevance to the drawing ND.

In response to this, in the information processing device 1s, data in which each past drawing is sorted in descending order of relevance score is presented to the user as a search result. This allows the user to review each past drawing presented as a search result in descending order of relevance score. In other words, the user can be given priority in reviewing past drawings that are ranked as higher-ranking candidates.

In this way, unlike the drawing search system of Patent Document 1, the information processing device 1s does not require the user to thoroughly check each past drawing presented as a search result (using a brute force method). Therefore, even compared to the drawing search system of Patent Document 1, the information processing device 1s makes it possible for the user to find the desired drawing corresponding to drawing ND more easily than before. In other words, the information processing device 1s can further improve user convenience compared to the drawing search system of Patent Document 1.

Incidentally, in recent years, image matching technology (e.g., facial recognition technology) that utilizes machine learning has come into widespread use. It is conceivable to use this image matching technology to search for a desired drawing. However, with the above-mentioned image matching technology, the relevance between the drawing ND and each past drawing is determined solely based on the "appearance" (e.g., layout) of the image. For this reason, a drawing search method based on image matching technology (hereinafter, image matching-based drawing search) may not be as suitable for drawing search as the information processing device 1s. This point will be described below.

For example, if the format for describing drawings is not standardized (unified), it is possible that the way in which drawings are described varies depending on the creator of each drawing. Therefore, even if similar content (e.g., specification information) is described in two drawings, the content may be written in different locations in each drawing. For example, see "area L1 in Figure 3" (hereinafter also referred to as area A) and "area R1 in Figure 6" (hereinafter also referred to as area B).

In such a case, in image matching-based drawing search, drawing a1 may be evaluated as having low relevance to drawing ND because the positions of areas A and B are different. However, as described above, the same type of specification information (specification information related to voltage values) is described in areas A and B. Therefore, in reality, drawing a1 may be considered to have a certain degree of relevance to drawing ND.

As described above, image matching-based drawing search cannot evaluate the relevance between drawing ND and drawing a1 by taking into account the content of the specification information written on each drawing. Therefore, it cannot properly search for a previous drawing (drawing b1 in the example of Figure 8) that looks different from drawing ND but contains similar content.

In contrast to this, the information processing device 1s, unlike the image matching-based drawing search, focuses on specific character strings written in each drawing and sets content parameters (search target drawing content parameters and target drawing content parameters). In other words, the information processing device 1s can take into account the content of the specification information written in each drawing. Then, based on the content parameters, the relevance between the drawing ND and each past drawing is evaluated (e.g., scored).

Therefore, unlike image matching-based drawing search, the information processing device 1s can appropriately search for past drawings (e.g., drawing b1) that look different from drawing ND but contain similar content. In this way, the information processing device 1s is more suitable for drawing search than image matching-based drawing search.

In addition, the greater the number of past drawings, the more difficult it becomes for a conventional drawing search system (or image matching-based drawing search) to properly search for past drawings that are highly relevant to a given new drawing. Therefore, the greater the number of past drawings, the more useful the information processing device 1s is for improving user convenience.

[Modifications]
(1) The information processing device according to one aspect of the present invention can also be applied to searching for drawings in fields other than plant engineering. The drawings according to one aspect of the present invention may be drawings from which content parameters can be obtained by the drawing search device.

(2) In the reference embodiment, the learning model generation unit 113s uses multinomial logistic regression analysis to generate a learning model. However, the method for generating a learning model is not limited to this.

The learning model generation unit 113s only needs to be able to generate a learning model by machine learning using the above-mentioned teacher data (it only needs to be able to calculate each weighting coefficient). For example, the learning model generation unit 113s can also generate a learning model using a publicly known neural network technique (e.g., a publicly known deep learning technique).

(3) The search result data generating unit 124 may further acquire specific data corresponding to the drawing selected as candidate g. In the example of FIG. 8, the search result data generating unit 124 may acquire specification data of a specific device corresponding to drawing b1 (candidate 1). This is because in various plant engineering plans, past drawings are often referenced along with the specifications of the specific device described in the past drawings.

In this case, the search result data generating unit 124 can present the specification data to the user along with the search result data. This can further improve user convenience. As an example, a table showing the correspondence between the drawings a1 to MN and the specification data may be prepared in advance in the information processing device 1s. The search result data generating unit 124 can refer to the table and obtain the specification data corresponding to the candidate g.

[Embodiment 1]
9 is a block diagram showing the configuration of the main parts of the information processing system 100 according to the first embodiment. The information processing device of the information processing system 100 is referred to as the information processing device 1. The control device of the information processing device 1 is referred to as the control device 10. The learning device of the control device 10 is referred to as the learning device 11. First, prior to a specific description of the processing of the information processing system 100, possible improvements of the information processing system 100s according to the reference embodiment will be described.

As described above, in the information processing device 1s of the reference embodiment, all past drawings (drawings a1 to MN) are used for learning. For this reason, in the information processing device 1s, it is conceivable that the quality of the learning model will vary depending on the contents of the group of past drawings (drawings a1 to MN) stored in the past property drawing DB91. This is because, as described below, if the contents of the group of past drawings are biased, a learning model based on biased drawing contents (biased learning model) will be generated.

(Pattern classification regarding the quality of learning models in reference form)
The inventors have examined the quality of the learning model in the reference embodiment. Specifically, the inventors have classified the relationship between the past drawings and the quality of the learning model into the following four patterns.

In the following explanation of past drawings (e.g. production drawings), "contents linked to the drawing number" typically means "the contents of the specification drawings (drawings containing specifications) used to create the past drawings." Examples of production drawings include outline drawings and assembly drawings. However, as mentioned above, please note that past drawings are not limited to production drawings.

For simplicity's sake, in the following example, we will assume that the specification drawings for one property contain all the specifications for creating all the production drawings for that one property. We will then assume that all the production drawings for that one property are created based on the specification drawings for that one property.

Also, in the following example, if a production drawing belonging to a certain drawing type (e.g. outline drawing) and another production drawing have the same (common) specifications for the certain drawing type (e.g. specifications for creating the outline drawing), the same drawing number will be assigned to the one production drawing and the other production drawing. However, naturally, if the specifications for the certain drawing type differ between the one production drawing and the other production drawing, different drawing numbers will be assigned to the one production drawing and the other production drawing.

(Pattern 1: When the group of past drawings is composed only of past drawings with different contents linked to the drawing numbers and different drawing numbers.)
Different drawing contents can be learned from multiple past drawings with different drawing numbers. This allows a particularly high-quality learning model (a learning model with sufficient generalization) to be generated. As a result, high search accuracy is achieved.

(Pattern 2: When the past drawings contain drawings with different contents linked to the drawing numbers but the same drawing number)
Different drawing contents are learned from drawings with the same drawing number. That is, learning is performed using ambiguous data. However, in general, machine learning can generate a high-quality learning model even when learning using such ambiguous data. Therefore, in pattern 2, as in pattern 1, high search accuracy is achieved.

(Pattern 3: When the past drawings contain drawings with the same or similar content linked to the drawing numbers but different drawing numbers)
The same or similar drawing contents are learned from multiple past drawings with different drawing numbers. That is, biased learning is performed. For this reason, in pattern 3, a biased learning model is generated. As a result, the search accuracy may decrease. For example, there is a high possibility that multiple different drawings having the same or similar drawing contents will be hit as candidates 1 to g.

(Pattern 4: When the group of past drawings contains past drawings with the same or similar content linked to the drawing number and the same drawing number)
The same or similar drawing contents are learned multiple times using past drawings with the same drawing number. In this case, biased learning is also performed. Therefore, a biased learning model is generated in pattern 4 as well. As a result, the search accuracy may decrease.

As can be seen from patterns 3 and 4 above, a biased learning model can lead to a decrease in search accuracy. For this reason, a biased learning model can also be considered a low-quality learning model. If a drawing search is performed using such a biased learning model, the chances of obtaining appropriate search results may decrease.

The inventors therefore created the information processing system 100 of embodiment 1 in order to further improve the information processing system 100s. As described below, the information processing system 100 (more specifically, the learning device 11) can prevent degradation of the quality of the learning model. Therefore, it is possible to achieve even higher search accuracy than the information processing system 100s.

(Learning device 11)
Unlike the learning device 11s, the learning device 11 further includes a dissimilarity calculation unit 114, a similar set determination unit 115, and a non-learning drawing determination unit 116. The similar set determination unit 115 and the non-learning drawing determination unit 116 may be collectively referred to as a determination unit. Moreover, the learning device 11 includes a learning model generation unit 113 (learning unit) instead of the learning model generation unit 113s of the learning device 11s.

(Difference calculation unit 114)
The dissimilarity calculation unit 114 calculates the dissimilarity between any two past drawings based on the content parameters (first to Lth content parameters) of the two past drawings derived by the past drawing data preprocessing unit 112. do.

In the following explanation, any two past drawings will be referred to as drawing I and drawing J. I and J are integers greater than or equal to 1 and less than or equal to T. For convenience, in the following explanation, the kth content parameter of drawing I will be referred to as Ik. Similarly, the kth content parameter of drawing J will be referred to as Jk. Furthermore, the degree of difference between drawing I and drawing J will be referred to as D(I, J). D(I, J) is an index value that indicates the degree of difference between the contents (description) of drawing I and drawing J.

Fig. 10 is a diagram for explaining the degree of difference in the first embodiment. In Fig. 10, for example, "drawing a1" is abbreviated as "a1". Similar abbreviations are used appropriately in this specification. Fig. 10 shows an L-dimensional data space (L-dimensional space) showing the distribution of content parameters of each drawing. In the L-dimensional space of Fig. 10, point I corresponding to drawing I (point showing content parameters of drawing I) is called P(I). The coordinates of P(I) are
P(I) = [I1, I2, ..., Ik, ..., IL]
The dissimilarity calculation unit 114 acquires I1 to IL from the previous drawing data pre-processing unit 112, and determines the coordinates of P(I) in the L-dimensional space.

Similarly, in the L-dimensional space of FIG. 10, a point J corresponding to drawing J is called P(J). The coordinates of P(J) are as follows:
P(J) = [J1, J2,..., Jk,..., JL]
The dissimilarity calculation unit 114 acquires J1 to JL from the previous drawing data preprocessing unit 112, and determines the coordinates of P(J) in the L-dimensional space.

As an example, the dissimilarity calculation unit 114 calculates the distance between point I and point J in the L-dimensional space as D(I,J). In the first embodiment, the dissimilarity calculation unit 114 calculates the Euclidean distance between point I and point J as D(I,J). That is, the dissimilarity calculation unit 114 calculates
D (I, J) = {(I1-J1) ² +...+(IL-JL) ² } ^1/2
= {Σ(Ik-Jk) ² } ^1/2 ...(4)
D(I, J) is calculated as follows.

As is clear from formula (4) above, D(I, J) takes a value equal to or greater than 0. The larger D(I, J) is, the lower the similarity in content between drawing I and drawing J. Conversely, the smaller D(I, J) is, the higher the similarity in content between drawing I and drawing J. When D(I, J) = 0, the similarity between drawing I and drawing J is at its maximum. Specifically, when D(I, J) = 0, drawing I and drawing J have the same content.

The dissimilarity calculation unit 114 calculates D(I, J) for all combinations of I and J. Then, the dissimilarity calculation unit 114 generates a table indicating the dissimilarity between each of the drawings a1 to MN as the dissimilarity table TB11. TB11 includes the i-th sub-table TB11-i in the dissimilarity table. TB11-i is a table indicating the dissimilarity between each past drawing belonging to drawing i. FIG. 11 shows TB11-1 as an example. TB11-1 shows the dissimilarity between each past drawing (drawings a1 to M1) belonging to drawing 1 (outline drawing). TB11-1 is a square matrix with M rows and M columns. TB11 is composed of a set of N sub-tables, TB11-1 to TB11-N. In this way, in the first embodiment, a sub-table for the dissimilarity is created for each drawing type.

Below, the operation of the dissimilarity calculation unit 114, the similarity set determination unit 115, and the non-learning drawing determination unit 116 will be explained with reference to TB11-1. However, the following explanations of each item related to TB11-1 also apply to TB11-2 to TB11-N in the same manner.

As is clear from the above formula (4), TB11-1(I,J) = TB11-1(J,I). Also, TB11-1(I,I) = 0. Therefore, as shown in FIG. 11, TB11-1 is generated as a symmetric matrix with diagonal elements of 0. Therefore, the following explanation will mainly focus on the case where I<J.

The dissimilarity calculation unit 114 sets the value of TB11-1(I,J) as TB11-1(I,J) = D(I,J). As an example, assume that D(a1,b1) = 8.1. In this case, the dissimilarity calculation unit 114 sets the value of TB11-1(1,2) as TB11-1(a1,b1) = 8.1.

(Similar set determining unit 115)
The similarity set determination unit 115 determines a set of past drawings (similar set) that are considered to have high similarity from among the drawings a1 to M1 based on the degree of difference shown in the TB11-1. The similar set determination unit 115 determines similar sets based on a threshold value of dissimilarity (for convenience, referred to as Dth). In the first embodiment, the case where Dth is set to 6.0 is illustrated. However, , Dth may be arbitrarily set by the user of the information processing system 100.

Figure 12 is a diagram showing an example of the processing of the similar set determination unit 115. In the example of Figure 12, for convenience of explanation, TB11-1 in Figure 11 is also referred to as table 1200A. Table 1200A may also be referred to as a first generation dissimilarity table.

The similarity set determination unit 115 searches table 1200A for a component with the minimum dissimilarity under the condition I<J. Next, the similarity set determination unit 115 determines whether the discovered minimum dissimilarity (for convenience, referred to as Dmin) is equal to or less than Dth. If Dmin is equal to or less than Dth, the similarity set determination unit 115 determines the set of past drawings (set of past drawings) corresponding to Dmin as a similar set. In other words, if Dmin is equal to or less than Dth, the similarity set determination unit 115 classifies (groups) drawings a1 to M1 so that the sets of past drawings corresponding to Dmin belong to the same set.

In the example of FIG. 12, the similar set determination unit 115 searches table 1200A and identifies D(c1, e1) = 2.5 as the smallest component (see the hatched portion of table 1200A). Therefore, the similar set determination unit 115 sets Dmin as Dmin = 2.5. In table 1200A, Dmin < Dth. Therefore, the similar set determination unit 115 determines the set {c1, e1} as set A (see also FIG. 13 described below). Set A may be referred to as the first similar set.

Then, the similar set determiner 115 generates table 1200B by integrating elements c1 and e1 (components of set A) in table 1200A into set A. Table 1200B may be referred to as a second generation dissimilarity table. When generating table 1200B, the similar set determiner 115 sets the dissimilarity between set A and each element (each past drawing) outside set A, i.e., D(A, a1), D(A, b1), ..., D(A, M1), using a predetermined method.

In the first embodiment, the similar set determiner 115 sets the maximum value of the dissimilarity between each element belonging to set A (i.e., drawings c1 and e1) and one element not belonging to set A (for convenience, referred to as drawing X) as D(A, X).
D(A,X)=Max{D(c1,X),D(e1,X)}...(5)
In this manner, in the first embodiment, D(A, X) is set by the longest distance method of partitional hierarchical clustering. This also applies to set B, which will be described later.

As an example, consider the case where X=a1. As shown in table 1200B in FIG. 12, the similar set determiner 115 uses the above formula (5) to obtain
D (A, a1) = Max {D (c1, a1), D (e1, a1)}
=Max(3.3,10.3)
= 10.3
In the same manner, the similar set determiner 115 sets D(A, b1), ..., D(A, M1).

Next, the similarity set determination unit 115 searches table 1200B for a component with the smallest dissimilarity under the condition I<J. As a result of searching table 1200B, the similarity set determination unit 115 identifies D(b1, d1)=5.8 as the smallest component (see the hatched portion of table 1200B). Therefore, the similarity set determination unit 115 sets Dmin as Dmin=5.8. In table 1200B, Dmin<Dth. Therefore, the similarity set determination unit 115 determines the set {b1, d1} as set B (see also FIG. 13). Set B may be referred to as the second similarity set.

Then, the similar set determiner 115 generates table 1200C by integrating elements b1 and d1 (components of set B) in table 1200B into set B. Table 1200C may be referred to as a third-generation dissimilarity table. The similar set determiner 115 uses the above-mentioned formula (5) to set the dissimilarity between set B and each element outside set B, that is, D(B,A), D(B,a1), ..., D(B,M1). Note, however, that in calculating (B,A), D(B,a1), ..., D(B,M1), set A on the left side of formula (5) is replaced with set B.

As an example, the similar set determining unit 115
D(B,A)=Max{D(b1,A),L(d1,A)}
=Max(12.8, 11.4)
= 12.9
As can be understood from the calculation example of D(B, A), in the calculation in the similar set determination unit 115, a certain set (e.g., set A) is regarded as a virtual past drawing that represents a plurality of past drawings (drawings c1 and e1).

Next, the similar set determination unit 115 searches table 1200C for a component with the smallest dissimilarity under the condition I<J. As a result of searching table 1200C, the similar set determination unit 115 identifies D(B, a1)=8.1 as the smallest component (see the hatched portion of table 1200C). Therefore, the similar set determination unit 115 sets Dmin as Dmi=8.1. In table 1200C, Dmin>Dth. If Dmin is greater than Dth, the similar set determination unit 115 completes the set grouping process.

Figure 13 is an example of a dendrogram DG that represents the hierarchical structure of the set obtained in embodiment 1. By using the DG, the user can visually recognize the hierarchical structure of the set. The DG may be used as auxiliary information to allow the user to adjust Dth.

As can be seen from Fig. 13, when Dth is set to 8.1 or more, drawing a1 and set B are classified into the same set (for convenience, referred to as set B') by the similar set determining unit 115. As described above, B = {b1, d1}. Therefore,
B'={a1, b1, d1}={a1, {b1, d1}}={a1, B}
In this way, in the first embodiment, it is also possible to handle a new set (eg, set B') that is a nested version of an existing set (eg, set B).

10 again, consider the case of calculating D(B', e1). In the first embodiment, the similar set determiner 115 calculates D(B', e1) using the following formula (6):
D(B', e1)
=Max{D(a1, e1), D(b1, e1), D(d1, e1)}...(6)
That is, the similar set determination unit 115 may calculate L(B', e1) by calculating the maximum distance between each drawing belonging to the set B' and the drawing e1.
may be determined.

However, the above formula (6) is
D(B', e1)
=Max{D(a1, e1), Max{D(b1, e1), D(d1, e1)}}
=Max{D(a1, e1), D(B, e1)}...(6')
Therefore, the similar set determining unit 115 can calculate D(B', e1) by using the formula (6') instead of the formula (6).

In formula (6'), the number of arguments of the Max function is smaller than in formula (6). Therefore, by using formula (6') instead of formula (6), the calculation of D(B', e1) in the similar set determination unit 115 can be accelerated. In other words, by using formula (6'), D(B', e1) can be calculated in a shorter time than when formula (6) is used. In this way, in embodiment 1, various calculations can be performed faster (more efficiently) by handling a nested set structure.

(Non-learning drawing determination unit 116)
The non-learning drawing determination unit 116 determines non-learning drawings (past drawings not used for learning in the learning model generation unit 113) from among the drawings a1 to M1 according to the degree of difference between the drawings a1 to M1. In the first embodiment, the non-learning drawing determination unit 116 determines non-learning drawings from among the past drawings belonging to each set (e.g., set A and set B) determined by the similar set determination unit 115.

Figure 14 is a diagram showing an example of the processing of the non-learning drawing determination unit 116. In the first embodiment, the non-learning drawing determination unit 116 determines one representative drawing (representative drawing in a set) from among a plurality of drawings belonging to a certain set (for convenience, referred to as set X). The non-learning drawing determination unit 116 then determines each drawing, excluding the representative drawing, from among the plurality of drawings belonging to set X as a non-learning drawing. The determination of the representative drawing may be performed by any method. As an example, the non-learning drawing determination unit 116 may determine the drawing with the latest timestamp from among the plurality of drawings belonging to set X as the representative drawing.

Figure 14 illustrates an example in which the non-study drawing determination unit 116 (i) determines drawing c1 as the representative drawing in set A from among two drawings c1 and e1 belonging to set A, and (ii) determines drawing b1 as the representative drawing in set B from among two drawings b1 and d1 belonging to set B. In this case, the non-study drawing determination unit 116 (i) determines drawing e1 as the non-study drawing in set A, and (ii) determines drawing d1 as the non-study drawing in set B. Thus, in the example in Figure 14, drawings d1 and e1 are determined as non-study drawings from among drawings a1 to M1.

After completing the above-mentioned processing for drawing 1 (drawings a1 to M1), the dissimilarity calculation unit 114, similar set determination unit 115, and non-learning drawing determination unit 116 perform the same processing for drawing 2 (drawings a2 to M2). That is, the dissimilarity calculation unit 114, similar set determination unit 115, and non-learning drawing determination unit 116 perform the above-mentioned processing in the order of drawing 1 → drawing 2 → ... → ... drawing N. In this way, the non-learning drawing determination unit 116 determines which drawings from a1 to MN are to be excluded from the study, depending on the dissimilarity between the drawings a1 to MN.

Then, the non-learning drawing determination unit 116 edits the above-mentioned TB2 according to the non-learning drawings determined by itself. Specifically, the non-learning drawing determination unit 116 edits TB2-i according to the non-learning drawings determined in drawing i. FIG. 14 shows an example of editing TB2-1 by the non-learning drawing determination unit 116. In the example of FIG. 14, the non-learning drawing determination unit 116 deletes the content parameters corresponding to the non-learning drawings in drawing 1 (i.e., drawings d1 and e1) from TB2-1. In addition, the non-learning drawing determination unit 116 assigns the drawing number of drawing c1 (representative drawing of set A) to set A, and assigns the drawing number of drawing b1 (representative drawing of set B) to set B.

The non-study drawing determination unit 116 generates the post-edit second table TB12 by editing TB2 as described above. Specifically, the non-study drawing determination unit 116 generates the i-th sub-table TB12-i in the post-edit second table by editing TB2-i as described above. FIG. 15 shows an example of TB12-1. TB12 is composed of a set of N sub-tables, TB12-1 to TB12-N.

(Learning model generation unit 113)
The learning model generation unit 113 acquires the TB 12 generated by the non-learning drawing determination unit 116 as teacher data. The learning model generation unit 113 then generates a learning model by performing a learning process (e.g., multinomial logistic regression analysis) similar to that of the reference embodiment. Each process after the generation of the learning model is similar to that of the reference embodiment.

(Effects of the First Embodiment)
Similarly to the reference embodiment, the information processing system 100 (information processing device 1) can improve the user's convenience in searching for drawings more than ever before. In addition, unlike the reference embodiment, the learning device 11 can generate a learning model using the TB 12. That is, unlike the reference embodiment, the learning device 11 can generate a learning model based on the content parameters of each past drawing excluding drawings that are not the learning target.

As described above, in the learning device 11, drawings that are highly similar to other past drawings are determined as drawings not to be learned. Then, the drawings not to be learned can be excluded and learning can be performed based on past drawings. In other words, learning can be performed based on a group of past drawings whose contents are low in similarity to each other. Therefore, the possibility of learning the above-mentioned patterns 3 and 4 can be reduced. As a result, the possibility of obtaining a biased learning model can be reduced.

In this way, the learning device 11 can prevent a decline in the quality of the learning model by performing learning while excluding drawings that are not the learning target (ensuring the quality of the learning model). Therefore, in the information processing system 100, a search can be performed by the drawing search device 12 using a high-quality learning model generated by the learning device 11. As a result, the information processing system 100 can achieve even higher search accuracy than the reference embodiment.

[Modifications]
(1) In the first embodiment, a distance in an L-dimensional space is exemplified as an example of a dissimilarity according to one aspect of the present invention. In particular, in the first embodiment, a Euclidean distance is exemplified as an example of a distance. However, the distance is not limited to the Euclidean distance. Other examples of the distance include a Manhattan distance or a Chebyshev distance. The distance may be any known type of distance (norm) that can be defined in an L-dimensional space.

In addition, please note that the dissimilarity according to one aspect of the present invention is not necessarily limited to distance. D(I, J) may be any index value calculated based on I1-IL and J1-JL, as long as it can express the difference in the contents of drawings I and J.

(2) In the first embodiment, the case where D(A, X) is set by the longest distance method is exemplified. However, the method for setting D(A, X) is not limited to the longest distance method. For example, D(A, X) may be set by other methods such as the shortest distance method, the center of gravity method, or the group average method.

When the longest distance method is used, D(A, X) can be set to a larger value than when the other methods described above are used. Therefore, by using the longest distance method, it is expected that the similar set determination unit 115 can complete the classification process of similar sets with a smaller number of calculations. For this reason, the longest distance method can be said to be a suitable method for reducing the calculation cost of the learning device 11.

[Software implementation example]
The control block (particularly the control device 10) of the information processing system 100 may be realized by a logic circuit (hardware) formed on an integrated circuit (IC chip) or the like, or may be realized by software.

In the latter case, the information processing system 100 includes a computer that executes the instructions of a program, which is software that realizes each function. The computer includes, for example, one or more processors, and a computer-readable recording medium that stores the program. The object of the present invention is achieved by the processor reading the program from the recording medium and executing it in the computer. The processor can be, for example, a CPU (Central Processing Unit). The recording medium can be a "non-transient tangible medium," such as a ROM (Read Only Memory), as well as a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, or the like. The system may also include a RAM (Random Access Memory) that expands the program. The program may be supplied to the computer via any transmission medium (such as a communication network or a broadcast wave) that can transmit the program. One aspect of the present invention can also be realized in the form of a data signal embedded in a carrier wave, in which the program is embodied by electronic transmission.

[Additional Notes]
The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

1 Information processing device 10 Control device 11 Learning device (model generating device)
100 Information processing system 111 Past drawing data acquisition unit 112 Past drawing data preprocessing unit (search target drawing content parameter acquisition unit)
113 Learning model generation unit (learning unit)
114 Dissimilarity calculation unit 115 Similar set determination unit (determination unit)
116 Drawing not to be studied decision unit (decision unit)
a1 to MN drawings (past drawings, drawings to be searched)
ND Drawing (New Drawing, Target Drawing)
A, B Set d1, e1 Drawings (drawings not subject to study)

Claims (3)

A model generation device that generates a learning model for searching at least one drawing corresponding to a target drawing from a plurality of search target drawings,
a search target drawing content parameter acquisition unit that acquires search target drawing content parameters related to the description contents of each of the plurality of search target drawings by analyzing the plurality of search target drawings;
a dissimilarity calculation unit that calculates a dissimilarity between any two different search target drawings included in the plurality of search target drawings based on the search target drawing content parameters of each of the plurality of search target drawings;
a determining unit that determines non-learning drawings from among the plurality of search target drawings according to the degree of difference;
A learning unit that generates the learning model based on the search target drawing content parameters of each of the plurality of search target drawings excluding the non-learning target drawings ,
The determination unit is
determining a set consisting of pairs of the plurality of search target drawings whose dissimilarity is equal to or less than a threshold value;
determining the non-learning drawings from among the plurality of search target drawings belonging to the set;
The determination unit sets the maximum value of the dissimilarity between (i) each of the plurality of search target drawings belonging to the set and (ii) a certain search target drawing that does not belong to the set as the dissimilarity between the set and the certain search target drawing, a model generating device. The search target drawing content parameters include L parameters from a search target drawing first content parameter to a search target drawing Lth content parameter,
One of the two different search target drawings is called drawing I, and the other is called drawing J.
The model generating device according to claim 1, wherein the dissimilarity calculation unit calculates the distance between a point corresponding to drawing I and a point corresponding to drawing J in an L- dimensional data space indicating the distribution of the search target drawing content parameters as the dissimilarity between drawing I and drawing J. A model generation method for generating a learning model for searching at least one drawing corresponding to a target drawing from a plurality of search target drawings, comprising:
a search target drawing content parameter acquisition step of acquiring search target drawing content parameters related to the contents of each of the plurality of search target drawings by analyzing the plurality of search target drawings;
a difference calculation step of calculating a difference in contents between any two different search target drawings included in the plurality of search target drawings based on the search target drawing content parameters of each of the plurality of search target drawings;
a determining step of determining non-learning drawings from among the plurality of search target drawings according to the degree of difference;
A learning process for generating the learning model based on the search target drawing content parameters of each of the plurality of search target drawings excluding the non-learning target drawings ,
In the above determination step,
determining a set consisting of pairs of the plurality of search target drawings whose dissimilarity is equal to or less than a threshold value;
determining the non-learning drawings from among the plurality of search target drawings belonging to the set;
A model generation method in which, in the determination process, the maximum value of the dissimilarity between (i) each of the plurality of search target drawings belonging to the set and (ii) one search target drawing that does not belong to the set is set as the dissimilarity between the set and the one search target drawing .

Images