JPH0728766A - Pattern recognition system and monitor system - Google Patents

Abstract

PURPOSE:To provide the pattern recognition system which detects a weak reaction. CONSTITUTION:The character pattern inputted by an image scanner 1 and a picture preprocessing part 10 is recognized by a neural network. When a weak reaction detecting part 11 detects a weak reaction, a problem pattern storage part 12 stores a problem pattern, and a related information storage part 13 stores related information. A teacher information determining part 15 determines teacher information of each pattern. Teacher information is stored in a teacher information storage part 16, and a validity discriminating part 17 discriminates the validity of the recognition result based on teacher information with respect to each or patterns other than the problem pattern. Patterns whose recognition results are not valid are stored in the problem pattern storage part 12. A teacher information adding work support part 18 supports operator's teacher information adding work by outputting related information with respect to the problem pattern. If a performance reduction detecting part 19 detects the performance reduction of the neural network, an adjusting part 20 adjusts parameters of the neural network.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an improvement of a pattern recognition system for recognizing which one of a plurality of recognition result candidates a pattern corresponds to, and more particularly to detecting a weak reaction in the recognition.

[0002]

2. Description of the Related Art Conventionally, a pattern recognition system for recognizing various patterns has been used. This pattern recognition system includes a recognition unit such as a neural net configured based on a plurality of recognition result candidates, and this recognition unit recognizes which recognition result candidate the input pattern corresponds to. It is a thing. Here, in this specification, “corresponding” is a broad concept including not only complete identity but also feature similarity and part similarity. This pattern recognition system is used for various purposes such as recognizing patterns such as handwritten characters, voices, signals, and images, inputting them into a computer, and evaluating the state of large-scale systems such as chemical plants. There is.

By the way, in many cases, the pattern to be recognized by the pattern recognition system generally has characteristics of its population that change depending on conditions. For example, when the pattern to be recognized is a pattern that represents the state of a chemical plant, the pressure distribution applied to the pipes and valves changes due to the temporal condition of aging deterioration of various devices of the plant. In addition, the letters of the address notation of the mail change depending on the condition of the area. That is, since the postal mail in the center of the city has many business letters and direct mails, the address notation has many printed characters, while the suburban mail has many personal letters, so the address notation has many handwritten characters.

In order to cope with the change in pattern due to such conditions and to maintain the recognition performance of the pattern recognition system, the recognition means is adjusted so as to match the characteristics of the pattern input in the actual use environment. This is very important.
For this purpose, it is necessary to periodically evaluate the recognition performance of the pattern recognition system and adjust the parameters so that the performance for unrecognizable patterns is improved when the performance is relatively or absolutely decreased. is there. Here, the “parameter” means data such as a numerical value for controlling the configuration and behavior of the recognizing means. For example, when the recognizing means is a neural network, the coupling between the elements forming the neural network is performed. It is a parameter that controls the load and the reaction characteristics of each element.

[0005]

However, the conventional pattern recognition system as described above has the following problems. First, the conventional pattern recognition system mechanically outputs the code of the recognition result candidate having the most similar input pattern as a recognition result, even when a pattern that is not similar to any recognition result candidate is input. , Did not detect the fact that the pattern is unrecognizable.

That is, in a general pattern recognition system, pattern similarity is calculated for all recognition result candidates, and the code of the recognition result candidate having the maximum similarity mechanically becomes the recognition result. For example, when the recognition means is a neural network, each element in the output layer of the neural network (hereinafter referred to as “output layer element”) corresponds to each recognition result candidate, and the pattern of similarity with each recognition result candidate is similar. The degree is output as the reaction intensity of each output layer element. Then, the code of the recognition result candidate corresponding to the output layer element showing the maximum reaction strength becomes the recognition result. In this case, the reaction strength of the output layer element becomes the certainty factor for the recognition result.

When a pattern that is not similar to any of the recognition result candidates is input, the reaction strength, that is, the certainty factor, of the pattern with respect to each recognition result candidate is low and there is no great difference. In the calligraphy,
It is called "weak reaction" that the recognition means produces only a reaction strength below a predetermined standard.

However, in the conventional pattern recognition system as described above, even in such a case, the code of the recognition result candidate having the highest degree of similarity is the mechanical recognition result, so that the weak reaction should be detected. I couldn't. In the case of a weak reaction, the pattern that causes the weak reaction (referred to as “problem pattern” in the present specification) requires a process based on the weak reaction, such as adjustment of recognition means. In the pattern recognition system, the weak reaction was not detected, and as a result, the processing based on the weak reaction could not be performed.

Further, in the conventional pattern recognition system,
As a result of not detecting the weak reaction as described above, the above-described deterioration in performance of the recognition unit was not detected (such deterioration in performance may be detected by observing weak reactions a predetermined number of times or more, for example). Conceivable.). For this reason, it is difficult to know an appropriate adjustment time of the recognition means, and the deterioration of the performance of the recognition means may be revealed by the fact that the actual operation is hindered. Therefore, there has been a demand for a pattern recognition system that can be adjusted at an appropriate time.

Further, in the conventional pattern recognition system,
As a result of weak reaction not being detected, it was also impossible to automatically collect problem patterns during system operation. Therefore, in the adjustment of the recognition means in the conventional pattern recognition system, when a significant decrease in performance is found, a large amount of new sample patterns are collected in the use environment,
A method has been adopted in which the entire adjustment of various parameters that determine the behavior of the recognition means is redone from the beginning.

However, it is inefficient, costly, and difficult to perform the entire adjustment again. That is, in order to achieve high recognition performance in the entire adjustment redone,
The evaluation of the recognition performance and the adjustment of the parameters have to be repeated many times, and such a process increases the cost of the adjustment. Moreover, since such a processing requires a large-scale computer for processing a large amount of data, it cannot be performed at the site of use of the pattern recognition system, and simple adjustment is difficult.

In particular, in such indiscriminate sample collection, in order to collect a pattern having a low appearance frequency as a sample, the operator constantly monitors the pattern recognition system for a long time and determines the number of samples to be collected. It had to be increased further and the burden of adjustment was further increased.

Further, in the adjustment of the recognition means, generally, a sample and teacher information indicating a correct recognition result for the sample are used as a pair. Since the conventional pattern recognition system does not include means for automatically determining the teacher information, the teacher information must be added to the sample by the human operator inputting it for each sample. It was difficult to adjust efficiently.

Further, since the conventional pattern recognition system does not have means for supporting the task of adding the teacher information, the human operator must collect the information about the sample required for the task of adding the teacher information. The burden of adding teacher information was further increasing.

The present invention has been proposed to solve the above-mentioned problems of the prior art, and an object thereof is to provide a pattern recognition system for detecting a weak reaction. Another object of the present invention is to provide a pattern recognition system which detects a performance deterioration of the recognition means.
Another object of the present invention is to provide a pattern recognition system that automatically collects problem patterns that cause weak reactions. Another object of the present invention is to provide a pattern recognition system which is inexpensive and can adjust the recognition means easily. Another object of the present invention is to provide a pattern recognition system that automatically determines teacher information corresponding to a problem pattern. Another object of the present invention is to provide a pattern recognition system that effectively supports the task of adding teacher information to a problem pattern.

Still another object of the present invention is to detect an abnormality in a target system to be monitored as a weak reaction,
It is to provide a monitoring system that reports an abnormality of the target system.

[0017]

In order to achieve the above object, the pattern recognition system according to claim 1 recognizes input means for inputting a pattern and which of a plurality of recognition result candidates the pattern corresponds to. In a pattern recognition system having a recognition means and an output means for outputting the result of the recognition, a weak reaction detection means for detecting a weak reaction in the recognition is provided.

The invention according to claim 2 is the pattern recognition system according to claim 1, characterized by further comprising performance deterioration detecting means for detecting a performance deterioration of the recognition means.

The invention according to claim 3 is the pattern recognition system according to claim 1, characterized by further comprising problem pattern storage means for storing a problem pattern which is the pattern causing the weak reaction. .

The invention according to claim 4 is the pattern recognition system according to claim 3, further comprising teacher information determining means for determining teacher information which is information indicating a correct recognition result corresponding to the problem pattern. Characterize.

According to a fifth aspect of the present invention, in the pattern recognition system according to the fourth aspect, the validity judgment for judging the validity of the recognition result for the pattern based on the teacher information corresponding to the pattern. It is characterized by having means.

The invention according to claim 6 is the pattern recognition system according to claim 4, further comprising adjusting means for adjusting the recognizing means based on the problem pattern and the teacher information corresponding to the problem pattern. Is characterized by.

According to a seventh aspect of the present invention, in the pattern recognition system according to the third aspect, the problem pattern,
It is characterized by comprising teacher information addition work support means for supporting the addition work of the teacher information to the problem pattern by outputting relevant information related to the problem pattern.

The invention according to claim 8 is, in the pattern recognition system according to claim 1, as the recognition means,
It is characterized by using a neural network.

The monitoring system according to a ninth aspect of the present invention is an input means for inputting a pattern representing a state of a target system to be monitored, and a recognition means for recognizing which of a plurality of recognition result candidates the pattern corresponds to. The plurality of recognition result candidates is a predicted state in the target system, a weak reaction detection means for detecting a weak reaction in the recognition, and a notification means for notifying the detection of the weak reaction It is characterized by that.

According to a tenth aspect of the invention, in the monitoring system according to the ninth aspect, a neural net is used as the recognizing means.

[0027]

The present invention having the above structure has the following functions. That is, in the invention of claim 1, since the weak reaction detecting means detects the weak reaction in the pattern recognition, it is possible to perform the processing based on the weak reaction.

Further, in the invention of claim 2, the performance deterioration detecting means detects the performance deterioration in the recognizing means, so that the recognizing means can be adjusted at an appropriate time.

According to the third aspect of the invention, the problem pattern storage means stores the problem pattern that causes a weak reaction, so that the problem pattern can be automatically collected.

Further, according to the invention of claim 4, since the teacher information determining means automatically determines the teacher information corresponding to the problem pattern, the burden of adjusting the recognizing means is reduced.

Further, in the invention of claim 5, the validity determining means confirms the validity of the recognition result for the pattern based on the teacher information corresponding to each pattern, so that a weak reaction does not occur but is erroneous. It is possible to detect patterns that cause recognition.

Further, in the invention of claim 6, the adjusting means comprises:
Since the recognition means is adjusted based on the problem pattern and the teacher information, the pattern recognition system can be adjusted easily and inexpensively in the field of use.

Further, in the invention of claim 7, the teacher information adding work supporting means effectively supports the teacher information adding work, so that the recognition means can be easily adjusted.

Further, according to the invention of claim 8, since the neural network is used as the recognition means, the pattern recognition can be performed without systematically logicalizing the recognition reference.

According to the ninth aspect of the invention, the abnormality of the target system is detected and reported as a weak reaction related to the pattern on the premise of the recognition result candidate representing the predicted state of the target system. It is possible to detect an abnormality in the target system without positively determining the condition of the state.

According to the tenth aspect of the present invention, since the neural network is used as the recognition means, the target system can be monitored without systematically logicalizing the recognition reference.

[0037]

Next, a plurality of embodiments of the present invention will be specifically described with reference to the drawings. It should be noted that each of the following embodiments is realized using a computer, and in the configuration of each embodiment, except for a portion such as an input / output device that should be realized exclusively by hardware, a predetermined program stored as a program is used. It is realized by operating the computer in steps. Therefore, the present apparatus will be described below assuming a virtual circuit block having each function of each embodiment.

The computer used in each embodiment has at least a CPU (central processing unit) and RA.
It has an M (mainly write / read type storage element) main storage device, and typically has an input device such as a keyboard and a mouse, an output device such as a CRT display device and a printer printing device, and a necessary device. An I / O control circuit is included.

As a software configuration of this computer, typically, a mode in which a program for realizing each embodiment is executed as an application (application) program on an OS (operating system) can be considered. However, only a dedicated program including an input / output procedure may be installed on the computer.

The storage mode of the program is also free, and it may be stored in a ROM (read-only memory) or stored in an external storage device such as a hard disk drive at the time of starting the computer. It may be loaded into the main memory at the start of processing. Alternatively, the program may be divided into a plurality of parts and stored in an external storage device, and only the necessary modules may be loaded (read) into the main memory at any time according to the processing content.

In each of the embodiments, the means for storing information is typically provided on an external storage device such as a hard disk device, but a main memory or a flash memory that has been tried in recent years (a large-capacity nonvolatile memory). Memory). Information that is not changed for the time being, such as dictionary data, may be stored in the ROM.

A. First Embodiment (1) Configuration of the First Embodiment The first embodiment is a character reading device used for automatic classification of mail and the like, which is one type of the pattern recognition system of the present invention. First, FIG. 1 is a block diagram showing the configuration of the first embodiment. In FIG. 1, arrows indicate data flows, and broken arrows indicate control flows. Further, in FIG. 1, the intersection of the arrows is merely a graphic intersection and does not mean the relation between the data or control flows represented by the arrows. In the first embodiment, as shown in this figure, an image scanner 1 for inputting image data of a predetermined size including one or more characters, and an external interface 2 for recognition result output (corresponding to the output means). A keyboard 3 and a mouse 4, a display device 5, and a processing unit 6 for processing information. The processing unit 6 is provided with an I / O control circuit 7.

Further, the processing unit 6 uses the neural network driver 8 for calculating the behavior of the neural network for recognizing the pattern, the parameter storage unit 9 for storing the parameters for determining the specifications of the neural network, and the image data. The image preprocessing unit 10 cuts out a pattern for each character and inputs it to the neural network driver 8. The neural network driver 8 and the parameter storage unit 9 compose the recognition means. Further, the image scanner 1 and the image preprocessing unit 10 constitute the input means.

Further, the processing unit 6 stores a weak reaction detecting unit 11 (corresponding to the weak reaction detecting means) for detecting a weak reaction in pattern recognition, and a problem pattern storage for storing a problem pattern causing a weak reaction. It has a section 12 (corresponding to the problem pattern storage means) and a related information storage section 13 for storing related information related to the problem pattern.

Further, the processing unit 6 stores the dictionary data storage unit 14 for storing the dictionary data of the words consisting of the characters which are the recognition result candidates, and the recognition results for the character patterns continuous before and after the pattern for each pattern. A teacher information determination unit 15 that determines teacher information corresponding to the pattern based on the dictionary data, and a teacher information storage unit 16 that stores the teacher information. The dictionary data storage unit 14 and the teacher information determination unit 15 constitute the teacher information determination means.

Further, the processing section 6 determines, for patterns other than the problem pattern, based on the teacher information of the pattern.
There is a validity determining section 17 (which constitutes the validity determining means) for determining the validity of the pattern and regarding an invalid pattern as a problem pattern and storing it in the problem pattern storage section 12.

Further, the processing section 6 outputs related information relating to the problem pattern and receives the selection input of the teacher information to support the teacher information addition work supporting section 18, and based on a predetermined standard. A performance degradation detecting unit 19 (corresponding to the performance degradation detecting means) that detects a performance degradation of the neural network, and adjusts the parameters of the neural network based on the problem pattern and the teacher information when the performance degradation is detected. And an adjusting unit 20 (corresponding to the adjusting means). The keyboard 3, the mouse 4, the display device 5, the I / O control circuit 7, and the teacher information addition work support unit constitute the teacher information addition work support means.

To do this.

The neural network realized by the neural network driver 8 and the parameter storage unit 9 is an information processing mechanism simulating a neural network of a living body. The neural network used in this embodiment has an input layer, an intermediate layer and an output layer as shown in FIG. 2 (conceptual diagram of the neural network) (from the bottom of the figure), and each of these layers includes a plurality of elements. And each element is combined with all the elements in the layer immediately below. Circles in FIG. 2 represent individual elements, and arrows represent coupling between elements. In FIG. 2, for simplification of the figure,
Only a part of the coupling between elements is illustrated.

Of the respective layers, the input layer has J elements U _j having the same number as the dimension number J of the input pattern, and functions as an input terminal. That is, the pattern x is

## EQU1 ## When expressed as x = (x ₁ , x ₂ , ..., X _J ), the component x _{j of the} pattern x is input to each element U _j . Input layer element (hereinafter referred to as "input layer element")
The reaction properties of is the identity function, the reaction strength of element U _j is equal to x _j and the component x _j input.

Of the respective layers, the elements of the intermediate layer (hereinafter referred to as "intermediate layer elements") are arranged in a lattice pattern on the layers as shown in FIG. 3 (a plan view showing the arrangement of the intermediate layer elements). There is.
As shown in this figure, each element is arranged at a grid point corresponding to the physical coordinates defined on the layer. For example, the coordinate of the grid point at the lower left corner of the middle layer is (0,0), the coordinate to the right of that is (1,0), and the coordinate above is (0,1).
Is. The coordinates of the element are represented by the physical coordinates corresponding to the lattice point where the element is arranged. Therefore,
When X × Y elements are arranged vertically and horizontally on the intermediate layer, for example, the coordinates of the element in the upper right corner of the layer are (X−1,
Y-1).

The intermediate layer element U _i receives the reaction intensities of all the input layer elements. The reaction characteristic of the intermediate layer element is a bell-shaped reaction characteristic function g, that is,

[Equation 2] To adopt. Also, here, each symbol is

[Equation 3] Represents

That is, the neural network of the present embodiment is a neural network that has been attempted in recent years and includes an element in which the region of the input space in which the value of the reaction characteristic function is a predetermined threshold value or more is bounded. Although any distance can be used as the distance d between the vectors, here, the most general Euclidean distance, that is,

[Equation 4] To adopt.

FIG. 4 is a characteristic diagram showing an outline of the reaction characteristic function g, in which the horizontal axis represents the distance d and the vertical axis represents the reaction intensity a _i . Here, the reaction characteristic parameters r _i and p _i are parameters for adjusting the shape of the reaction characteristic, and are collectively called element parameters. They are,
Controls the sensitivity of the range of inputs to which the device responds. That is, when r _i is increased, the existence range of the pattern x in which the element strongly reacts is expanded, and when r _i is decreased, the existence range of the pattern in which the element strongly reacts is narrowed. Further, it is slowly varying characteristics with respect to d changes in by reducing the p _i, p _i
When is increased, the characteristic is such that it is sensitive to changes in d.

The solid line in FIG. 4 shows the shape of g when r _i = 1.0 and p _i = 1.0, and the broken line shows the shape when r _i = 0.5 and p _i = 5.0. The shaded area on the left side of the straight line perpendicular to the vertical axis indicates the range in which the element reacts at a predetermined threshold value δ or more in the case of the broken line. In this case, the threshold is 0.1. A region that reacts at a threshold value or more is hereinafter referred to as a reaction region.

In the case of an element having a hanging bell type reaction characteristic, the reaction region is a hypersphere on the J-dimensional input space centered on the coupling weight vector w _i , and is therefore called a reaction hypersphere. Here, the hypersphere corresponds to a sphere in a three-dimensional space in an arbitrary dimensional space.

FIG. 5 is a diagram showing the relationship between the reaction hypersphere and the coupling load vector, which is drawn as a case where the input space is three-dimensional. Ie. The element U _i responds to δ or more only to the pattern included in this reaction hypersphere.

Among the layers, the output layer element U _k receives the reaction intensities of all the intermediate layer elements and outputs the sigmoid function f, that is,

[Equation 5] React according to the reaction characteristics of Here, each symbol represents a _k : reaction intensity of U _k w _ki : load on coupling from the device U _{i of the} intermediate layer to U _k a _i : reaction intensity of the device U _i of the intermediate layer These output layer elements function as output terminals of the neural network.

The parameters of the neural network in the first embodiment are the connection weights (loads on the connection) that represent the ease of passage of signals between the connections between the elements, and the elements that control the reaction characteristics of each element. It is a parameter.

(2) Operation of the first embodiment The first embodiment having the above structure has the following operation. The parameters of the neural network are
In advance, all the characters to be output as the recognition result are adjusted so as to recognize the characters based on the recognition result candidates. The dictionary data storage unit 14 also stores words made of characters that should be recognition result candidates. It is assumed that these words are place names that appear in the mailing address.

[Schematic Operation] When the image scanner 1 inputs image data, the image preprocessing unit 10 converts the image data into 1
A character pattern for each character is cut out and input to the neural network driver 8. The neural network recognizes the input pattern and outputs the recognition result from the external interface 2. Upon recognition, when the weak reaction detection unit 11 detects a weak reaction, the problem pattern storage unit 12 stores the problem pattern that causes the weak reaction, and the related information storage unit 13 stores information related to the problem pattern.

Further, the teacher information determining section 15 determines the teacher information of each pattern. Of these pieces of teacher information, the teacher information of the problem pattern is stored in the teacher information storage unit 16, and the validity determining unit 17 determines the validity of the recognition result for the patterns other than the problem pattern based on the teacher information. A pattern whose recognition result is not valid is regarded as a problem pattern and stored in the problem pattern storage unit 12.

The teacher information addition work support unit 18 supports the teacher information addition work by the operator by outputting relevant information about the problem pattern. When the performance deterioration detecting unit 19 detects the performance deterioration of the neural network based on a predetermined reference, the adjusting unit 20 adjusts the parameters of the neural network based on the problem pattern and the teacher information. The details of the above procedure will be described below.

[Pattern Input from Image Scanner] In pattern input, first, the image scanner 1 inputs digital image data from an object such as a mail in which characters are written. That is, although not shown, the image scanner 1 has a camera and an image processing circuit, and when the camera captures an object on which characters are written and generates an image signal, the image processing circuit causes the image signal to be output from the image signal. Remove the noise (dirt) of
An image signal for each predetermined frame size is cut out from the image signal, and the cut out image signal is converted into digital image data.

This image data is sent to the processing unit 6, the image preprocessing unit 10 of the processing unit 6 detects a character string portion from the image data, and the data of this character string portion is separated for each character data. . FIG. 6 is an example of the image data of one handwritten character separated in this way. This image is digital bitmap data in which the grayscale values of pixels in which black is 1 and white is 0 are arranged in the order of horizontal X pixels and vertical Y pixels.

Further, the image preprocessing unit 10 converts each image data for each character into a vector pattern and inputs it to the neural network driver 8. That is, in this conversion, the grayscale value of the pixel in the lower left corner in FIG. 6 is used as the first element and the value in the upper right corner is set as the XY element, and element extraction in the order of left → right is performed in the order of downward → upward. This is a process of generating an X × Y-dimensional vector by using each extracted element as a partial vector.

[Pattern Recognition] The neural network driver 8 receives a pattern in vector format every time it is input.
The behavior of the neural network is calculated based on the parameters stored in the parameter storage unit 9. The result of this operation is
The reaction strength of each element of the neural network with respect to the input pattern is determined, and the pattern recognition is performed thereby to generate the recognition result. That is, in the neural network, the output of the neural network is determined step by step from the lower layer by the reaction strength of the element with respect to the input being propagated to the upper layer through the coupling line in the order of input layer-middle layer-output layer.

In the neural network of the first embodiment, each output layer element has a one-to-one correspondence with a code representing each character (type) to be output as a recognition result, and is a character code which is a recognition result for a pattern. Is output when the reaction intensity of the element corresponding to the character code is the strongest among the output layer elements.

The recognition result obtained in this way is processed by the processing unit 6
In addition to being used for data processing described later, the data is sent via the external interface 2 to an external information processing means such as a sorter section in the automatic mail address reading / sorting machine.

[Detection of Weak Reaction] When the pattern recognition is performed as described above, the neural network driver 8 provides the weak reaction detection unit 11 with the reaction intensity of each element in the neural network, and the weak reaction detection unit 11 receives the reaction intensity. Based on the provided data, it detects a phenomenon that the reaction strength of the neural network is below a predetermined value. For example, in the neural network used in the first embodiment, when the pattern is not included in the reaction area of any intermediate layer element, none of the intermediate layer elements react and the output layer element does not react. That is,

[Equation 6] If does not hold, there is no intermediate layer element that reacts sufficiently strongly to the pattern, so it is appropriate to detect a weak reaction. Here, U _i is a reaction intensity of the intermediate layer element, x is a pattern, and θ _a is a minute positive definite value representing _a predetermined threshold value. This minute value θ _a changes depending on the characteristics of the pattern group to be recognized, but may be set to about 0.1, for example.

In the case of a weak reaction, the problem pattern storage unit 12 stores the pattern that causes the weak reaction as a problem pattern, and the related information storage unit 13 stores the related information related to the problem pattern. Here, the related information is, for example,
The response strength of each element of the neural network when the problem pattern is input, the pattern input before and after the problem pattern (hereinafter referred to as “related pattern”), the reaction strength of each element with respect to this related pattern, and the recognition result.

In the first embodiment, it is possible to provide an informing means for informing that when a weak reaction is detected. By doing so, it is possible to avoid confusion due to the use of a recognition result having a low certainty factor. .

[Determination of Teacher Information] Teacher information determination unit 15
Determines the teacher information for each pattern including the problem pattern. That is, when the character string is a fixed character string such as a place name, the appearance order of the characters has regularity. Therefore, the teacher information is a part of the recognition result for each character of the character string (preferably a majority ) Is obtained with sufficient confidence, for other parts of the string,
Even if the recognition result is not obtained or the recognition result is incorrect, it can be correctly determined. Specifically, when a part of the place name is missing, by searching the dictionary data in the dictionary data storage unit 14, the character that should be in the missing part is specified based on the relationship with the character immediately before and after, and the teacher information Can be

For example, FIG. 7 is an image of the address description part of the mail, and from the top of the eight character patterns in the rightmost row of this image, from the top, the recognition result of "Toju Tatsuta Tatsutacho" Suppose that is obtained. Here, the second character "□"
Was not recognized by the neural net, ie
This shows a problem pattern in which the response of all output layer devices was weak. In this case, when the teacher information determination unit 15 searches the dictionary data for a ward name that starts with the letters “foot”, only “Adachi” (ward) is obtained as the corresponding one. Therefore, the correct recognition result corresponding to this problem pattern, that is, the teacher symbol can be determined to be "standing".

If the teacher information cannot be uniquely determined by the above procedure, it is determined as a set of a plurality of candidates. For example, if the recognition result of the character just before the question pattern is "mountain" and the recognition result of the pattern immediately after is "prefecture", the teacher information is "shape" (Yamagata prefecture), "pear".
(Yamanashi Prefecture), “Mouth” (Yamaguchi Prefecture). The teacher information and the teacher information candidates determined as described above are stored in the teacher information storage unit 16.

[Determination of Validity of Recognition Result] The teacher information determination unit 15 determines teacher information based on dictionary data for each pattern of a character string that does not include a problem pattern. For example, when the dictionary data is searched by using "Tozumi Tatsuta", which is the recognition result of the 4th to 7th patterns, as the place name, of the recognition results of "Tozumi Tatsutacho, Ashiku-ku" in FIG.
It turns out that this place name does not exist in Adachi Ward. Search for a place name consisting of four letters in Adachi City
Only “Senju Ryuta” has the same constituent part as, and it can be determined that the teacher information for the fourth pattern is “thousand”.

The teacher information determined as described above is used to judge the validity of the recognition result. That is, the validity determination unit 17 compares the recognition result of each pattern with the teacher information corresponding to the pattern, and if the two do not match, the problem pattern storage unit 12 stores the pattern as a problem pattern. In this case, the teacher information and the related information corresponding to the pattern are also stored in the teacher information storage unit 16 and the related information storage unit 13, respectively.

[Detection of Degradation of Recognition Performance] As described above, the problem pattern input while the first embodiment is used for pattern recognition is accumulated in the problem pattern storage unit 12, The performance degradation detection unit 19 performs a performance degradation detection process for the neural network every time the is stored. Although a specific standard for detecting this performance degradation can be freely set, specifically, the performance is degraded when the number of problem patterns stored in the problem pattern storage unit 12 exceeds a certain number, for example. You should judge that.

When the performance degradation is detected, the performance degradation detecting unit 19 activates the adjusting unit 20. The activation of the adjusting unit 20 is performed when the first embodiment is not used for the recognition processing (for example, by mail). It is highly desirable to set it to run at night when the station's work is finished. Alternatively, only the alarm may be output when the performance deterioration is detected, and the adjustment may be performed when the operator inputs another adjustment instruction during the maintenance work.

Prior to the adjustment, the teacher information addition work support unit 18 supports the teacher information addition work. At this time, the operator confirms and confirms the teacher information determined by the teacher information determination unit 15, or The teacher information determined by the operator is added to the problem pattern for which the teacher information could not be determined.

[Teacher Information Addition Work] That is, the teacher information addition work support unit 18 presents each question pattern and related information of each question pattern to the operator by displaying it on the display device 5, and at the same time, the teacher information addition work is performed. By inputting a selection via the keyboard 3 or the mouse 4, the operator's task of adding teacher information is supported. In this case, the operator is also presented with the teacher information and the teacher information candidates as the related information, and the operator may determine the teacher information based on the presented related information and input the determined teacher information on the spot. The support of the teacher information addition work is continued until the teacher information corresponding to all question patterns in the question pattern storage unit 12 is confirmed or input.

When the operator commands the stop, the work of adding the teacher information may be stopped and the adjustment may be performed only on the problem pattern for which the teacher information is fixed or has been input.

Here, FIG. 8 is a display example of problem patterns and related information. In this display, a series of character patterns forming a character string including a problem pattern are displayed as an image together with formal items such as a pattern collection period. This image is divided into patterns for each number of letters,
The number of each pattern is displayed in parentheses like [1] when the teacher information for the pattern is confirmed and when the certainty factor of the recognition result is equal to or higher than a predetermined reference value. Those not enclosed in parentheses are problem patterns for which teacher information is undetermined, and are targets for which the operator should specify teacher information.

For example, the recognition result of the pattern number 1 is “foot”, and the reaction strength of the output layer element corresponding to the character type “foot” is strong at 0.892 (high reaction strength / confidence). Moreover, since this recognition result matches the teacher information based on the dictionary data, this pattern is not a problem pattern, and adjustment and operator judgment are unnecessary.

The recognition result of pattern number 2 is "x".
However, this is a problem pattern because it indicates that none of the elements in the output layer reacted sufficiently strongly, and adjustment and operator judgment (determination by confirmation operation) are necessary.

In this input / output screen, the pattern number pointed to by the mouse cursor, which is indicated by a diagonal black arrow, is the direct operation target.
In, the pattern is on the fourth pattern, and the fourth pattern and its related information are surrounded by a thick frame. The output layer element that responds most strongly to this pattern is the element corresponding to the Chinese numeral "10", and its reaction strength is 0.885.
Is. However, since the teacher information based on the dictionary data is the Chinese numeral “thousand” and the recognition result is inconsistent with this teacher information, this pattern is a problem pattern, and it is necessary to determine the teacher information and adjust the parameters.

The adjustment of the parameters of the neural network in the first embodiment is automatically performed by the adjusting unit 20, but this adjustment may be made manually by the operator. In this case, the adjustment is presented by the teacher information addition work. Relevant information will guide the adjustment. For example, as shown in the lower part of FIG.
The maximum reaction strength of the middle layer of the neural network for this pattern is 0.677, and the average reaction strength is 0.320. This value means that there is an element that responds to this pattern with a relatively high value in the intermediate layer, but the reaction intensity of this element is not properly transmitted to the output layer. Therefore, the operator can determine that the coupling load from the intermediate layer to the output layer should be adjusted in order to obtain correct recognition for this pattern. Further, when the maximum reaction strength of the intermediate layer is low, the operator can judge that it is necessary to adjust the coupling load of the input layer-the intermediate layer or increase the number of elements of the intermediate layer.

As the neural network in the first embodiment, a neural network in which the number of elements is automatically increased / decreased, which has been tried in recent years, may be used. In this case, there is no or almost no operator intervention. It becomes possible to adjust the parameters.

When there are a plurality of candidates for the teacher information, the operator has to specify which candidate to adopt. In that case, the mouse button may be clicked (pressed) on the candidate to be selected.

If the teacher information is not provided at all or if the teacher information does not include correct teacher information, the mouse cursor is used to point to the "Teacher information designation" column at the lower right of the screen to convert so-called Kana-Kanji characters. And JIS
Specify the correct character type by the character specification method such as code input. The teacher information thus determined is stored in the teacher information storage means as the confirmed teacher information.

[Parameter Adjustment] Following the teacher information addition work, the adjusting unit 20 adjusts the parameters of the neural network. This adjustment is an operation for adjusting the position, size, and shape of the reaction area of the element so that the reaction area of the element includes the problem pattern, and is performed using the problem pattern and the teacher information corresponding thereto. . Here, the correction based on the error between the reaction intensity (that is, the output pattern) of the output layer element and the determined teacher information will be described.

Among the parameter adjustments, in the correction of the coupling weight to the output layer element, first, all the elements U _{k of the} output layer are corrected.
The error between the reaction of and the teacher information is obtained. The teacher information specifies a desired reaction strength of the output layer element with respect to the pattern. Here, the teacher information t is

Equation 7] _{t = (t 1, t 2} , ···, t k) a _{(0 <t k <1)} , the value is that component t _k to specify the desired reaction intensity of the output layer element U _k is there. In the first embodiment, teacher information is given such that only the component for the output layer element indicating the character type of the pattern has the value 1 and the components for the other elements have the value 0.

[0093] In this case, the error E _k in U _k is,

## EQU8 _## Defined as E _k = t _k −a _k .

Therefore, the coupling weight w _ki from all the elements U _i to U _k of the intermediate layer is according to the error E _k :

[Equation 9] w _ki ← w _ki + Δw _ki

[Equation 10] Δw _ki ← ηE _k a _i is modified. Here, “←” represents a substitution operation to a variable. Further, η is a coefficient indicating the correction rate of the coupling load, and normally it may be set to about η = 0.1.

Among the parameter adjustments, the following calculations are performed in the correction of the coupling load of the intermediate layer element and the element parameter. That is, first, the element that most strongly reacted with the pattern in the intermediate layer is determined. That is,

[Equation 11] The element U _m having the maximum value of is obtained. Next, ν is set on the intermediate layer with a radius R centered on the element U _m . That is, using the coordinates determined corresponding to the grid points on the intermediate layer described above,

[Equation 12] And Here, P _m and P _i are elements U _m and U, respectively.
The physical coordinates of _{i on} the intermediate layer are shown. Further, d is the Euclidean distance. Although the value of R will not be described in detail here, as an example, when the size of the intermediate layer is vertical Y and horizontal X, and min () is a function that takes the minimum value of the argument,

[Equation 13] And Here, T is the number of learning repetitions.

Ν in the above equation is a set of elements in the vicinity of the element U _m . Coupling weight vector w _i and element parameter p
_i and r _i are modified only in the elements belonging to ν. For the elements included in ν, the correction of the coupling load is

[Expression 14] w _i ← w _i + Δw _i

## EQU15 ## According to Δw _i ← α (x−w _i ). Here, x is the input pattern. Further, α is a coefficient indicating the correction rate of the coupling load. This value should be set to about 0.1.

For the elements included in ν, the element parameter r _i is modified by

## EQU16 ## r _i ← r _i + Δr _i

[Equation 17] Follow Here, γ _r is a positive coefficient that gives the correction rate of r _i , and is normally set to about 0.1. r _i of the correction amount Δr _i
Is obtained as the sum of the first term determined by the error between the reaction intensity of the output layer element and the teacher information and the second term called the diffusion term.

Here, the meaning of the equation 17 will be described.
w _ki E _k is a quantity representing the extent to which the reaction of the element U _i causes the error E _k . If w _ki E _k > 0, it is desirable to weaken the reaction of U _{i in} order to bring it close to 0, so r _i should be decreased in order to reduce the radius of the reaction hypersphere of U _i. . Therefore,

[Equation 18] And it is sufficient. Then, the first term is obtained by summing all the elements in the output layer.

The second term is a range within which no error appears in the output layer,
The second term, which has the effect of increasing the radius of the reaction hypersphere, is a term for causing one intermediate layer element to react with a pattern in a region as wide as possible. Γ >> ξ _r > 0 is set so as not to affect the error in the output layer. In general, ξ _r =
It should be about 0.001.

For the elements included in ν, the element parameters p _i are modified according to the following equation.

[0101]

[Formula 19] p _i ← p _i + Δp _i

[Equation 20] Here, γ _p is a positive coefficient that gives a correction rate of p _i , and is normally set to about 0.1. ξ _p has the effect of widening the region in which U _i reacts strongly within the range that does not affect the error. _Let γ _p >> ξ _p > 0. Normally, ξ _p = 0.00
1 is enough.

The problem pattern and the teacher information are erased after the adjustment is completed.

(3) Effects of the First Embodiment As described above, according to the first embodiment, the weak reaction detecting section 11 is used.
However, since the weak reaction in the pattern recognition is detected, it is possible to perform the processing assuming the weak reaction. Further, in the first embodiment, since the performance deterioration detecting unit 19 detects the performance deterioration in the neural network, it is possible to adjust the neural network at an appropriate time. It is possible to prevent problems from occurring.

Further, in the first embodiment, the problem pattern storage unit 12 stores the problem patterns, so that the problem patterns can be automatically collected. Therefore, for the purpose of adjustment, it is not necessary to collect a large amount of new samples in the use environment or to redo the entire adjustment of the parameters of the neural network from the beginning, and the adjustment can be made more efficient.
Adjustment costs are low, and frequent adjustments are easy. In particular, since it is not necessary to process a large number of samples required to redo the entire adjustment, the adjustment cost is further reduced.

In particular, in the first embodiment, the problem patterns are selectively collected, so that the operator constantly monitors the pattern recognition system for a long time or collects the number of collected samples in order to collect patterns with low appearance frequencies. There is no need to increase it, and the burden of collecting problem patterns is reduced.

Further, in the first embodiment, the teacher information determining unit 1
Since 5 automatically determines the teacher information corresponding to the problem pattern, the burden of adjusting the neural network is reduced. Further, in the first embodiment, the teacher information addition work support unit 18 effectively supports the teacher information addition work, so that the burden on the adjustment of the neural network is further reduced.

Further, in the first embodiment, the validity judging section 17
However, since the validity of the recognition result of the neural net for the pattern is confirmed based on the teacher information determined by the teacher information determination unit 15 corresponding to each pattern, a weak reaction is not generated, but erroneous recognition is generated. Various patterns
It becomes possible to collect.

Further, in the first embodiment, the adjusting unit 20 automatically adjusts the parameters of the neural network based on the problem pattern and the teacher information, so that the adjustment can be performed easily and inexpensively in the field of use of the pattern recognition system. Becomes

As described above, according to the first embodiment, the weak reaction of the neural network with respect to the pattern and the performance deterioration are detected, the neural network can be adjusted at an appropriate time, and the adjustment work itself is also effective. Support and low cost
It is possible to provide an excellent pattern recognition system that adjusts a neural network simply and easily. Therefore, the pattern recognition system according to the first embodiment is suitable for a large-scale and complex recognition target, and particularly, it can quickly deal with a pattern in which the characteristics of the population of the pattern change depending on conditions. I will get it.

As a result, in the first embodiment, for example,
Widely applied to pattern recognition, which is difficult for human ability, such as voice recognition, condition monitoring of large-scale system such as electric power system / chemical plant whose characteristics change over time, prediction of power demand / stock and foreign exchange market conditions, surface flaw detection of parts, etc. can do.

B. Second Embodiment (1) Configuration of the Second Embodiment The second embodiment is a monitoring system to which the present invention is applied.
FIG. 9 is a block diagram showing the configuration of the second embodiment. In the second embodiment, as shown in this figure, a data collection device 22 that collects and inputs status data representing the status of the power system 21 (corresponding to the target system) that is a monitoring target, and the status data A monitoring unit 23 that monitors the state of the power system 21 by recognizing the state of the power system 21 on the basis of the state of the power system 21 and a position remote from the monitoring unit 23, and controls the power system 21 based on the result of the recognition. And a communication mechanism 25 that connects the monitoring unit 23 and the system control device 24.

Of these, the data collecting device 22 includes various sensors (not shown) and an AD converter for converting an output signal from the sensor into a digital signal as the status data, which is attached to each part of the power system 21. It has and.

The monitoring unit 23 also includes a pattern generator 26 for generating a pattern from state data, a neural network driver 27 for calculating the behavior of a neural network that recognizes the pattern, and parameters for determining the specifications of the neural network. Parameter storage unit 28 for storing
A weak reaction detection unit 29 (corresponding to the weak reaction detection means) for detecting a weak reaction in recognition of the neural network, and an alarm unit 30 for generating an alarm signal when the weak reaction is detected.
Have and. The data collection device 22 and the pattern generation unit 26 make up the input unit. Also, the neural network driver 27 and the parameter storage unit 28 that realize the operation of the neural network constitute the recognition means.

The system controller 24 also includes a display device 31 for displaying an alarm and a control mechanism (not shown) for controlling the power system 21. The alarm unit 30, the communication mechanism 25, and the system control device 24 of the monitoring unit 23 constitute the notification means.

The neural network in the second embodiment is
Similar to the first embodiment, it recognizes which of the plurality of recognition result candidates the input pattern corresponds to. The recognition result candidates in this neural network are various patterns that represent all the states that are predicted to occur on the power system 21. The parameters of this neural network are adjusted in advance so as to recognize the pattern based on the recognition result candidates and stored in the parameter storage unit 28.

(2) Operation of the second embodiment Monitoring of the state of the electric power system 21 in the second embodiment having the above-mentioned configuration is performed as follows. That is, the sensor group outputs various state quantities on the power system 21, for example, voltage, phase difference angle, and passing power as analog signals at predetermined sampling intervals. The AD converter standardizes these analog signals into values in a certain range and converts them into state data which are digital signals, and provides them to the monitoring unit 23. The pattern generation unit 26 of the monitoring unit 23 removes noise from the state data, generates a pattern that is a vector pattern representing the state of the entire power system 21 from the scalar value measured at each measurement point, and outputs the neural network driver 2
Provide to 7.

The neural network generates a state evaluation for the power system 21 as a recognition result for the pattern. Note that the neural network may generate not only the state evaluation of the power system 21 but also state prediction after a certain time. The state evaluation which is the recognition result is transmitted to the system control device 24 through the communication mechanism 25 together with the pattern, and is also provided to the weak reaction detection unit 29.

The neural network determines the state of the system for the pattern when the pattern is observed on the power system 21. That is, the determination of this state is expressed by the establishment failure of a plurality of evaluation items, and the simultaneous establishment of a plurality of items is also allowed. The output layer of the neural network outputs the determination of such a state, and therefore has the same number of elements as the number of each evaluation item, and when a certain evaluation item is satisfied, according to the certainty factor to the establishment, The element corresponding to the evaluation item produces a strong reaction strength.

When the recognition is performed, the weak reaction detecting section 29
As in the first embodiment, the case where the certainty factor of any evaluation item is equal to or less than a predetermined reference is detected based on the data regarding the reaction strength of each element of the neural network.
Since the neural network has learned all the predicted state patterns, if any other state pattern is observed, the power system 21 is considered to be in an abnormal state.

When a weak reaction is detected in this way, the alarm unit 30 transmits an alarm signal to the system controller 24 through the communication mechanism 25. The monitoring unit 23 may be configured to store the pattern that causes the weak reaction as a problem pattern when the weak reaction is detected.

The system controller 24 which has received the alarm signal
Since an alarm is issued through the display device 31 and the pattern at that time is displayed, the operator of the system control device 24 controls the power system 21 using the control mechanism provided in the system control device 24 with reference to this pattern. And can deal with abnormalities.

[Practical Example] Next, a practical example of abnormality detection in the power system 21 using the second embodiment will be described. In this example, the state quantity of the power system 21 is only the voltage at 25 observation points. FIG. 10 is a graph showing a 25-dimensional vector pattern which is a pattern showing the voltage in the actual example. The abscissa represents the observation point number, and the ordinate represents the unit voltage after converting the fluctuation range of the voltage and normalizing it in the range of [0, 1].

When a 25-dimensional voltage pattern is input, the neural network recognizes and evaluates the pattern based on the 12 evaluation items corresponding to each element in the output layer. The following table shows the contents of these 12 evaluation items.

[0124]

[Table 1]

FIG. 11 is a diagram showing an example of the evaluation result. In this figure, the vertical axis represents the evaluation item and the horizontal axis represents the certainty factor of establishment of each evaluation item. The certainty factor is the reaction strength of the element of the output layer corresponding to the evaluation item. The evaluation of this figure is based on the pattern "daytime", "slightly decreased voltage",
The three items of "Southeastern" are affirmed, in other words,
It means "the system is in the daytime and the voltage is slightly lower in the southeastern part".

(3) Effects of the Second Embodiment As described above, according to the second embodiment, the abnormality of the power system 21 is detected as a pattern on the basis of the recognition result candidate representing the predicted state of the power system 21. Since it is detected and reported as a weak reaction related to the above, it is possible to detect an unexpected abnormality of the power system 21 without positively determining the condition of the abnormal state of the power system 21.

Further, according to the second embodiment, since the neural net is used as the recognition means, it is possible to detect the abnormality of the power system 21 without systematically logicalizing the recognition reference.

C. Other Embodiments The present invention is not limited to the above-described embodiments, but can be variously modified without changing the gist thereof, and thus includes the following other embodiments. Is.

For example, in each of the above embodiments, the neural net is used as the recognizing means, but any recognizing means can be used as long as it is a means for recognizing a pattern. Basic statistical pattern recognition method (composite similarity method as an example)
It is also possible to use a recognition means based on.

Further, the pattern recognition system of the present invention is
Performance degradation detection means, problem pattern storage means, teacher information determination means, teacher information addition work support means, validity determination means,
It can also be implemented without providing all or part of the adjusting means. In the character recognition device of the first embodiment, the pattern recognition target is a character. However, the present invention is not limited to the character, but a large-scale and complicated target in which the statistical properties of the population of the pattern may change. For example, it is suitable for various applications such as voice recognition, sonar acoustic analysis, chemical plant monitoring, power demand forecasting, stock / exchange market forecasting.

Further, although the monitoring system of the second embodiment targets the power system for monitoring, the monitoring system of the present invention can target all targets such as a chemical plant. In addition, the monitoring system, the performance deterioration detecting means,
The problem pattern storage means, the teacher information determination means, the teacher information addition work support means, the validity determination means, and the adjustment means may be provided for implementation. In this way, it is possible to detect abnormality while flexibly changing the neural network that is the recognition means in response to changes in the power demand trend.

Although each of the above-described embodiments is realized by using a computer, all or part of the functions may be realized by a dedicated electronic circuit. For example, it is possible to use a neural network dedicated arithmetic unit that has been tried recently in order to realize a neural network as a recognition means.

[0133]

As described above, according to the pattern recognition system of the present invention, processing that is premised on weak reaction in recognition, in particular, a timely and efficient recognition means that responds to characteristic changes of the population of patterns. Can be adjusted.
Further, according to the monitoring system of the present invention, it is possible to detect an unexpected state of the monitoring target without positively determining the abnormal condition.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a first embodiment of the present invention.

FIG. 2 is a conceptual diagram showing a structure of a neural network in the embodiment.

FIG. 3 is a plan view showing an element array of an intermediate layer of the neural network in the example.

FIG. 4 is a characteristic diagram showing an outline of a reaction characteristic function of an intermediate layer element of the neural network in the example.

FIG. 5 is a diagram showing a relationship between a reaction hypersphere and a coupling load vector in the example.

FIG. 6 is an example of image data of a handwritten character for one character in the embodiment.

FIG. 7 is an example of an image of an address portion of a mail in the embodiment.

FIG. 8 is a display example of a problem pattern and related information in the embodiment.

FIG. 9 is a block diagram showing the configuration of a second embodiment of the present invention.

FIG. 10 is a graph showing a voltage pattern in an actual example of the same embodiment.

FIG. 11 is a graph showing an example of evaluation results in an actual example of the same embodiment.

[Explanation of symbols]

 1 ... Image scanner 2 ... External interface 3 ... Keyboard 4 ... Mouse 5, 31 ... Display device 6 ... Processing unit 7 ... I / O control circuit 8, 27 ... Neural network driver 9, 28 ... Parameter storage unit 10 ... Image pre-processing Part 11, 29 ... Weak reaction detection part 12 ... Question pattern storage part 13 ... Related information storage part 14 ... Dictionary data storage part 15 ... Teacher information determination part 16 ... Teacher information storage part 17 ... Validity determination part 18 ... Teacher information addition Work support unit 19 ... Performance deterioration detection unit 20 ... Adjustment unit 21 ... Power system 22 ... Data collection device 23 ... Monitoring unit 24 ... System control device 25 ... Communication mechanism 26 ... Pattern generation unit 30 ... Warning unit

Claims (10)

[Claims] 1. A pattern recognition system comprising: an input unit for inputting a pattern; a recognition unit for recognizing which of a plurality of recognition result candidates the pattern corresponds to; and an output unit for outputting the result of the recognition. A pattern recognition system comprising weak reaction detecting means for detecting a weak reaction in the recognition. 2. The pattern recognition system according to claim 1, further comprising a performance deterioration detection unit that detects a performance deterioration of the recognition unit. 3. The pattern recognition system according to claim 1, further comprising problem pattern storage means for storing a problem pattern that is the pattern that has caused the weak reaction. 4. The pattern recognition system according to claim 3, further comprising teacher information determining means for determining teacher information that is information representing a correct recognition result corresponding to the problem pattern. 5. The pattern recognition system according to claim 4, further comprising a validity determining means for determining validity of the recognition result for the pattern based on the teacher information corresponding to the pattern. . 6. The pattern recognition system according to claim 4, further comprising adjustment means for adjusting the recognition means based on the problem pattern and the teacher information corresponding to the problem pattern. 7. A teacher information addition work support means for supporting the addition work of the teacher information to the problem pattern by outputting the problem pattern and related information related to the problem pattern. The pattern recognition system according to claim 3. 8. The pattern recognition system according to claim 1, wherein a neural network is used as the recognition means. 9. An input unit for inputting a pattern representing a state of a target system to be monitored, and a recognition unit for recognizing which of a plurality of recognition result candidates the pattern corresponds to, The recognition result candidate is a predicted state in the target system, and is provided with a weak reaction detection unit that detects a weak reaction in the recognition, and a notification unit that notifies the detection of the weak reaction. . 10. The monitoring system according to claim 9, wherein a neural network is used as the recognition means.