JPH0210994B2 - - Google Patents

Description

[Detailed Description of the Invention] [Summary] A paper sheet validating machine development device that collects data on paper sheets including banknotes and performs various processes in developing a validating machine for the paper sheets. In order to solve the problem of reduced accuracy in classification logic caused by errors in attributes added to the data during collection, we automatically created classification logic from a database and analyzed the results of using that logic. It is disclosed that by configuring the system to compare the results indicated by the added attributes, abnormal data can be automatically removed in a short time and the accuracy of the discrimination logic can be improved.

[Industrial application field]

The present invention relates to a paper sheet identification machine development device, in particular, an apparatus that collects arbitrary data to create a database about the characteristics of paper sheets, and automatically performs a classification logic from the data on the database. The present invention relates to a development device for a paper sheet discriminator configured to remove abnormal data.

[Conventional technology]

Conventionally, recognition machines for paper sheets including banknotes have been developed. When examining the classification rate by the classification machine, in the case of conventional technology,
As shown in Figure 2, from the paper sheets to be identified,
A large amount of data similar to that used for discrimination by the discrimination machine under development is collected in a database 103 via a data collector 101 and a data editor 102, and then stored in a group dictionary table 104. The determination unit 106 uses the discrimination logic and the threshold table 105 for the recognition machine under development to check how accurately the recognition machine can classify paper sheets. To describe the above matter more specifically, for example, the following judgments were made.

(i) First, for a “randomly sampled sample”, the squares (zones/
data α (color, amount of reflection by color,
A dictionary table for each group (average and variance for each group) is created based on the category (denomination, direction of injection, etc.) determined for each specimen.

(ii) The values of the group-specific dictionary table created as above are used as "predetermined group-specific discriminant functions"
is applied to calculate the coefficients of the discriminant function.

(iii) Read the square data α for each banknote to be classified, calculate the discriminant function value corresponding to the group according to the category at that time, find the frequency of appearance, and calculate the classification rate using the threshold determined according to the frequency distribution. Measure whether the product is good or bad.

That is, a method has often been adopted in which a preferable discrimination mode is set "in advance," so to speak, and it is checked to what extent the discrimination mode can be correctly identified.
In other words, a decision logic is created in advance using an arbitrary small sample, and a large amount of collected data is passed through this logic to measure the classification rate/misclassification rate, confirm the performance, and trace the cause of performance deterioration. It was on. Therefore, it was not possible to determine whether the measured results were due to data errors or differential logic. Note that in FIG. 2, 110 corresponds to a data processing device.

[Problem that the invention seeks to solve]

Conventional discrimination logic is created by deriving a discriminant function from a small sample randomly selected from a database, creating a frequency table using this for the entire database, and calculating the desired discrimination rate from this frequency table. The discrimination logic was created using the value as the threshold. That is, as mentioned above, a dictionary table for each group is created using randomly selected samples, the coefficients of the discriminant function are calculated based on the contents of the table, and the good/bad classification rate is measured based on the coefficients. Because I was
If the results of the classification rate are unsatisfactory, it is necessary to re-extract the sample (of course, the dictionary table is updated and the group-by-group discriminant function needs to be recalculated), or the threshold value must be changed. I had no choice but to repeat trial and error. Therefore, the creation of a small sample, the derivation of a discriminant function, and the determination of a threshold value are independent tasks, resulting in a time-consuming problem. Furthermore, in order to derive a discriminant function using a small sample, a small sample is used that is not guaranteed to have the same accuracy as the database, and as a result, an inaccurate discriminant function may be derived. A problem arose. For this purpose, it is necessary to check whether the collected data itself is legitimate or not.
It was something that could not be checked.

The inventor of the present application previously applied for patent application No. 60-180086.
No., Special Application No. 1987-180087, Special Application No. 181446,
In Japanese Patent Application No. 60-181447, Japanese Patent Application No. 60-230702, etc., when developing a paper sheet recognition machine, it was considered that it would be possible to simulate the processing operations of various types of recognition machines. , has proposed a paper sheet identification machine development system.

The paper sheet identification machine development system collects a large amount of data at each surface position on paper sheets,
Data with undesired values are automatically excluded from the collected data, and a data group that will be read by the discriminator to be developed is extracted and generated, allowing various simulations to be performed.

Even in the above-mentioned paper sheet discrimination machine development system, the above-mentioned inaccurate discriminant function is used, or
It is desirable to solve undesirable problems such as the use of illegitimate data.

[Means for solving problems]

The present invention solves the above-mentioned problems by obtaining discriminant logic while obtaining a database, appropriately obtaining groups, and corresponding to the groups without adding any artificial judgment or manipulation. Based on the discriminant function, abnormal data on the database is excluded.

FIG. 1 shows a principle diagram of the present invention. Reference numeral 1 in the figure represents paper sheets including banknotes. Reference numeral 2 is a data collector, which has a function of reading information about set paper sheets using a plurality of sensors. 3 is a data editor that stores information that cannot be read by the sensor of the data collector (setting direction, type, lot, surrounding environment, etc. when paper sheets are set in the data collector 2). It has a function to add data to the data read. 4 is
a data base, the data processing device 110;
Stores data on paper sheets processed in .

5 is a primitive dictionary creation unit, which has a function of creating a primitive dictionary based on the sensor data for each additional information based on the information added to the paper sheet data on the database. ing. The original dictionary will be described later in FIG. 7, and stores information corresponding to individual meshes on paper sheets as a result of collecting information on a large amount of paper sheets.

7 is a group conversion table, and from the added information (setting direction, type of paper sheets, etc.),
This is a table for deriving the group to which it belongs.

Reference numeral 6 denotes a group dictionary creation unit, which has a function of synthesizing a group dictionary from the original dictionary created by the original dictionary creation unit 5 using the group conversion table 7.

Note that according to 5, 6, and 7 above, the dictionary creation unit 120
is configured.

Reference numeral 8 denotes a discriminant function frequency table creation unit, which has a function of creating a frequency table of each discriminant function for each group to which the paper sheets on the database belong. 10 is
This is a set value and indicates a desired classification rate. 9
is a threshold value creation unit, and has a function of deriving a threshold value that achieves a classification rate of a set value. Note that the threshold value creation processing unit 130 is constituted by the above-mentioned 8 and 9.

11 is a determination unit that determines which group the paper sheet data on the database belongs to from a discriminant function and a threshold value, and distinguishes between the determined group and a group derived from additional information. After comparison, if it is abnormal, it is stored in the abnormal database 13, and if it is normal, it is stored in the normal database 12.

[Effect]

In the data processing device 110, all data
The database (contents in 4) is used to automatically create identification logic, and the results of that logic are compared with additional information to determine abnormalities, so modifying the database is a series of tasks. This will be done automatically in a short time.

In addition, since the above-mentioned operation is repeated, even if a large amount of abnormal data is mixed in, the abnormal data can be excluded and the accuracy of the database can be maintained at a high level.

〔Example〕

An example will be explained below, but what is the classification rate and misclassification rate in the explanation? Classification rate = (number of groups identified as group x) / (total number of group x) Misclassification rate = (groups other than group x) (discriminated number)/(total number of groups x) (where x: any group).

FIG. 3 shows the configuration of an embodiment of the banknote data collector, in which FIG. 3A is a side view, FIG. 3B is a plan view, and FIG.
Figure C shows a configuration diagram of the roller. Reference numeral 14 in the figure is a feeding section, which supplies the banknote 18 to the data collection section 15 when the banknote 18 is inserted. The data collection unit 15 includes a roller 19 as shown in FIGS. 3B and 3C.
is generated, and the banknote 18' is conveyed on the conveyance path 20 as shown by the arrow. The conveyance path 20 is shown in FIG. 3B.
As shown, it has a width that is sufficiently large compared to the distance in the longitudinal direction of the banknote 18'. On the conveyance path 20, banknote entry detection sensors (S ₁ ) (S ₂ ) 21,
A passage detection sensor 22 is provided, and a data collection sensor 23 is arranged symmetrically with respect to the vertical center line shown in FIG. 3B. A track detection sensor 24 is also provided as shown.

As described above, since the width of the conveyance path 20 is sufficiently large compared to the longitudinal length of the banknote 18', the upper end of the banknote 18' is at the upper end of the conveyance path 20 shown in FIG. 3B. The banknotes 18' can be placed in any desired position, from a state in which the banknotes 18' are set and transported so that they touch each other, to a state in which the banknotes 18' are set and transported in such a way that the lower end of the banknotes 18' touches the lower end of the transport path 20 shown in FIG. It is possible to set Corresponding to these set positions, which position on the banknote 18' is sensed by the data collection sensor 23 is determined.

As shown in FIG. 4, the top of the banknote 18' is divided into m×n grid areas 25, and the _grid areas arranged in the horizontal direction in _FIG . If the grid area groups arranged in the vertical direction in FIG. 4 are defined as zones Z ₁ , Z ₂ , Z ₃ . Thus, while the banknote 18' is being conveyed, the data collection sensor 23 collects data in the square areas on the two upper and lower tracks. And banknote 1
8' on the conveyance path 20, each of the square areas 25 within a predetermined range on the banknote 18' is
5 data can be collected. The data in the hatched square area 25 in FIG. 4 is extracted at the transport timing corresponding to zone Z ₂ while one of the sensors 23 is collecting data along the track T _i+1 . be done. At this time, the sensor 23
Which track corresponds to can be determined by which position of the track detection sensor 24 the edge of the banknote 18' passes through.

The banknotes that have been conveyed through the data collection unit 15 are stored in the illustrated storage unit 16 as shown in the illustrated banknote 18″.The data collected during the above-mentioned conveyance is sent from the control unit 17 to the banknote shown in FIG. The data is transferred to the data editor 3.

FIG. 5 shows the configuration of an embodiment of the data editing device. Reference numeral 26 in the figure is a control unit (processor);
27 represents a data transmission means that receives data from the data collector. Reference numeral 28 denotes a data input section, which, in correspondence with each data transferred from the data collector 2, inputs (i) data on what denomination the banknote 18' corresponds to;
(ii) Which direction does the data correspond to when the banknote 18' is set (there are two orientations for the front side and two orientations for the back side)? (iii) What environmental conditions were there at the time of data collection? (iv) Which banknote production lot does the banknote 18' correspond to? These additional data are added to the data input unit 2.
It is input from 8.

29 is a data display section, 30 is a ROM writer, 3
1 is the data store section (floppy disk 4
).

Data 32 transferred from data collector 2
As shown in the upper part of FIG. 6, (i) skew amount and setting direction information, (ii) length information of data 32, (iii) conveyance speed information of banknote 18', (iv) track position information, ( v) Has a data format with sense data. Then, as described above, additional data,
, , are added, and the entire information 33 is stored on a floppy disk (not shown) and stored in the database 4 on the data processing device 110.

Original dictionary creation unit 5 in data processing device 110
reads the paper sheet data on the database 4 one by one, classifies the data based on the additional information on the data and the track position information, and divides the data into the minimum division unit (unit... as shown in Figure 4). Shaded area 2 shown
5), the number of data N for each zone, the cumulative sum Σx of data for each zone, as shown in FIG.
The cumulative sum of squares Σx ² of data for each zone is maintained on the table.

The group-specific dictionary creation unit 6 uses a group conversion table 7, as shown in FIG. Calculate the mean value M and variance value V of
The data is stored in a group dictionary table as shown in the figure. Furthermore, as mentioned above, the group numbers G ₁ , G ₁ , ...Gl correspond to the additional information and the track signal information.
The granting of this will be done as follows. That is, by examining the contents of the original dictionary table shown in FIG.
For example, the average value for each zone for the contents of the first table shown in Figure 7 and the second table shown in Figure 7.
If the average values of the contents of the first table for each zone tend to be similar across all zones, then the first table and the second table are Same group G ₁
be treated as belonging to the . When assigning group numbers in this way, it may be done manually by collecting those that are intuitively similar, or a table with the content to be approximated may be created using an appropriately prepared discriminant function. may be grouped together in the same group.

The average value M and variance value V at this time are derived as follows.

M ₁₁ = 1/N ₁ +N ₁ ′+…・( _N1 〓 x ₁ + _N1 ′ 〓 x ₁ ′+…) V ₁₁ = 1/N ₁ +N ₁ ′+…・( _N1 〓 x ₁ ² + _N1 ′ 〓 x ₁ ′ ² +…) −M ₁₁ ² N ₁ : Number of data when denomination 1, direction 1, track T ₁ , x ₁ : Data value when denomination 1, direction 1, track T ₁ , N ₁ ′: Number of data when denomination 1, direction 1, track T ₂ , x ₁ ′: Data value when denomination 1, direction 1, track T ₂ Discriminant function frequency table creation unit 8 prepares in advance. Assign the values of the group-specific dictionary to the discriminant F _a (M _a1 , M _a2 ,…, M _an , V _a1 , V _a2 ,…, V _an , X ₁ , X ₂ ,…, X _n ). Create a group-specific discriminant function. For example, function f ₁ of group G ₁
is f ₁ = F ₁ (M ₁₁ , M ₁₂ ,...M _1n , V ₁₁ , V ₁₂ ,..., V _1n , X ₁ , X ₂ ,..., X _n ) (Here, X ₁ to X _n are 1 ~ m zone sensor data). For example, when using Mahalanobis distance, the form of the above function Fa is F=(X ₁ −Ma ₁ ) ² /(γVa ₁ ) ² +(X ₂ −Ma
₂ ) ² / (γVa ₂ ) ² +……+ (Xn−Man) ² / (γVan) ² −
It will be something like 1. Note that in the above equation, γ represents a threshold value. Moreover, when using Euclidean distance, it is easy to write F=(X ₁ − Ma ₁ ) ² /γ ² +(X ₂ − Ma ₂ ) ² /
It becomes something like γ ² +...+(Xn-Man)/γ ² -1.

Next, as shown in FIG. 10, the data on the database 4 is read one by one and the groups shown in the group conversion table (FIG. 8) are
Calculate the discriminant function for the group, use the result to add 1 to the number of data for the value shown in the frequency table, and calculate the data.
A frequency table is created by performing the same processing on the entire base data.

The threshold creation unit 9 derives and holds thresholds for each group that satisfy the following conditions from the frequency table for each discriminant function and the set value S that indicates the classification rate from the outside.

The threshold value K ₁ of the discriminant function of group 1 in the case of the set value S is determined as satisfying S( _K1 〓 ^k=1 f _1k )/( _h 〓 ^k=1 f _1k ).

The determining unit 11 reads the data of paper sheets one by one from the database 4, and determines to which group the read data belongs based on the above-mentioned discriminant function and threshold value. For example, whether a piece of data belongs to group G ₁ depends on whether the condition f ₁ K ₁ is satisfied.

Next, as a result of the determination, it is determined that one piece of data belongs to any one group, and
If the determined group is different from the group derived from the additional data and the group conversion table, it is treated as abnormal data in the abnormal database 1.
3, and the others are stored in the normal database 12.

Here, when the ratio of abnormal data to normal data in one group, that is, the following error contamination rate α α = number of abnormal data/number of normal data, is large, the above derived discriminant function and threshold value are In addition to determining the
There is a risk that other groups that have been mixed in incorrectly will also be determined to be the own group.

Now, the frequency table F ₁ for the discriminant function f ₁ before removal,
The frequency table F _{2 for f 2 and} also the discriminant function after removal
Considering the frequency table F ₁ ′ for f ₁ , consider a case where group G ₂ is mixed into group G ₁ by mistake.

As shown in Figure 11, the frequency table F ₁ before removal is
Since the discriminant function was created considering part of group G ₂ as group G ₁ , group G ₁ and group
Separation from G ₂ is not clear. Further, in the frequency table _F2 at this time, as shown in FIG. 12, the group _G1 is included in the group _G2 .

A threshold value K ₁ that obtains a high discrimination rate in this state,
When K ₂ is set, while the erroneous components shown in FIG. 12 are removed, there is a risk that the legitimate components of group G ₂ shown in the shaded area in FIG. 11 will also be removed.

Therefore, the initial identification rate is set low (K ₁₁ ,
K ₂₁ ), create a discriminant function by removing only those that are definitely considered to be erroneous contaminations, and then gradually set the discrimination rate to a high value in the same way to remove the erroneous contaminations as shown in Figure 13. The correct discriminant function will be derived for each case. That is, the reliability of the removal is increased, and at the same time, the reliability of the derived discriminant function is also increased.

〔Effect of the invention〕

According to the present invention, since erroneous data on all databases is automatically removed, the accuracy of the database is improved, and the removal can be performed in a short time. Logic becomes more reliable.

[Brief explanation of drawings]

Fig. 1 is a diagram explaining the principle of the present invention, Fig. 2 is a diagram explaining a conventional example, and Figs. 3 to 6 are diagrams showing an embodiment from data collection to creation of a database. , FIG. 7 is a diagram for explaining the original dictionary table, FIG. 8 is a diagram for explaining the group conversion table, FIG. 9 is a diagram for explaining the group-specific dictionary table, and FIG. 10 is a diagram for explaining the creation of the frequency table. FIGS. 11 to 13 are diagrams illustrating abnormal data exclusion processing. In the figure, 1 is a paper sheet, 2 is a data collector, 3 is a data editor, 4 is a database, 5 is a primitive dictionary creator, 7 is a group conversion table, 8 is a discriminant function frequency table creator, 9 10 represents a threshold value creation unit, 10 represents a set value, and 11 represents a determination unit.

Claims (1)

[Scope of Claims] 1. A data collector 2 that reads information about set paper sheets 1 including banknotes using a plurality of sensors; A data editing device 3 that adds additional data related to data, a database 4 that holds information edited by the data editing device 3, and an information processing function that processes information in the database 4. The information processing function section includes at least a data processing device 110 having a data processing function section;
In a paper sheet validating machine development device that performs a simulation corresponding to the operation of a paper sheet validating machine to be developed using the contents of A threshold creation unit that derives a discriminant function for determining which of the N groups (N is an integer of 2 or more) to which the paper sheet belongs, G ₁ , G ₂ . . . , from the paper sheets on the database. 9, and determines the group indicated by the additional information and the group of the determination results as a result of determination by the discriminant function on the data on the database 4.
and is configured to remove the two groups as abnormal data if the two groups do not match according to the determination by the determination unit, and the information processing function unit removes the abnormal data from the group. A paper sheet discriminator development device characterized in that the content of the database 4 is classified based on the above, and the simulation is performed. 2. It has a function to set the discrimination rate using a discriminant function to a set value, and is configured so that the data determined as not abnormal as a result of the judgment is reused as a database, and automatic correction processing of the database is performed. is performed a predetermined number of times, and at this time, the classification rate is set to be the lowest for the first time and gradually increases thereafter, so that abnormal data are sequentially removed. An apparatus for developing a paper sheet discriminator according to claim 1.