JPH103464A - Process state recognition method, process state recognition device, data storage device, and learning device - Google Patents

Abstract

(57) [Summary] [Problem] Special data analysis and expert knowledge are required for process state monitoring. It is difficult to recognize the change tendency of the process data and to correct, change, and qualitatively express the change, which places a heavy burden on the specialist. A feature extraction process for generating a sequence of feature vectors based on a temporal change tendency of time-series process data, a vector search process for searching for a reference vector similar to the feature vector, a class corresponding to the reference vector A processing device 100 that executes a qualitative expression search process for searching for a qualitative expression corresponding to the sequence as a feature vector class sequence, a process for outputting the qualitative expression with similarity, and the like.
A reference vector dictionary 510 in which the input device 200 and the output device 300 are stored together with the corresponding class and the firing area radius, and a qualitative expression dictionary in which the feature vector class sequence is stored together with the corresponding qualitative expression 520.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a process status recognition method applied to a plant status monitoring system for monitoring the status and control status of a plant, a plant status control system for controlling the status of a plant, and the like. The present invention relates to a process state recognition device, a data storage device, and a learning device used directly for implementing a recognition method.

[0002]

2. Description of the Related Art In recent years,
The operation monitoring mode of the plant is changing from operation monitoring using analog panel instruments to using a distributed control system. Process monitoring of a plant by a distributed control system is usually performed by a few skilled operators who are familiar with the expertise in monitoring the condition of the process, but with that expertise, it is necessary to take all necessary measures to catch any abnormal signs. It is extremely difficult to constantly monitor process data because of the huge amount of process data managed by the distributed control system.

Further, it is considered that the recent tendency to decrease the number of skilled operators has spurred changes in plant operation monitoring forms. Even if an unskilled person learns specialized knowledge about the relationship between the time change tendency of process data and the process state, considerable time is required.

[0004] In view of the above points, it is generally conceivable to arrange these technical knowledge, for example, how to grasp the temporal change tendency for each process data in the form of documents or the like in advance. However, in general, the amount of such information is enormous, and when it is inappropriate to capture the changing tendency due to subtle differences that occur in each process, the user himself / herself arbitrarily corrects and changes the capturing method. Is difficult.

[0005] In addition, in order to reuse the way of grasping the temporal change tendency of the process data acquired in the past, the data must be recorded in some form such as a document. Requires a lot of effort. Even if a knowledge base is used to capture the changing trend of process data based on experience and know-how, if there is a subtle difference between processes, it is necessary to rearrange the target process each time. But these are a great burden for professionals. In particular, the burden on specialists to organize how to grasp the changing trend of process data first before other persons was extremely large.

[0006] As another conventional technique, (1) Japanese Unexamined Patent Application Publication No. 1-3000007 discloses a sudden change determination device, a level determination device, and an immobility determination device.
A process signal multiplex monitoring device that monitors a process signal variation characteristic from these three determination devices to monitor a plant abnormality. (2) Japanese Unexamined Patent Publication No. 6-17173 A process information display system that edits and displays and outputs a state in which a process state value changes over time as encoded information. (3) Japanese Unexamined Patent Publication (Kokai) No. 3-25601 The stability of the control system is determined based on the input / output signals of the control system.
A process control device that adjusts a control system according to the result of the determination, and also has a stability determining unit having a learning function of optimizing the result of the determination.

(4) Japanese Unexamined Patent Application Publication No. Hei 3-166601 A symbolizing unit for converting a feature amount of a time-series change pattern of a process state into an ambiguous symbol, a process control device using the symbolizing unit, and the like. (5) Japanese Unexamined Patent Publication No. 5-52606 A plant monitoring system that extracts historical data of process state quantities at a predetermined cycle, expresses the change pattern as a character string, and reports it to an operator. (6) Japanese Patent Application No. 7-87605 A process state recognition method, a process state recognition device, a data storage device, a learning device, and the like are known.

However, each of these prior arts has only a partial function of ideal plant state monitoring or process state recognition aimed at by the present invention, and has a process state recognition function and a qualitative expression of the process state. A learning function, a data storage function for learning and the like which are integrally combined have not yet been provided.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned various problems relating to monitoring of a change in the status of process data. It is possible to constantly monitor trends and detect abnormal signs at an early stage.Also, it is possible to change the setting of qualitative expressions to be notified by words specific to the target area, and to perform special data analysis and expert Provided is a process state recognition method capable of extracting a qualitative expression without requiring knowledge and outputting a qualitative expression with a similarity degree in a form desired by a user, and a process state recognizing device directly used for implementing the method. What you want to do.

Further, the present invention provides an apparatus used directly for implementing the above-mentioned process state recognition method, wherein even if a change tendency of process data is slightly different for each process variable, a reference used for grasping the change tendency is used. It is an object of the present invention to provide a learning device that can automatically refine the contents of a vector dictionary and a qualitative expression dictionary by a learning process, and a data storage device that stores data used in the learning process.

[0011]

According to a first aspect of the present invention, there is provided a process state recognition method for generating a sequence of feature vectors based on a temporal change tendency of time-series process data. A vector search process for calculating a similarity between the feature vector obtained by the feature extraction process and a plurality of reference vectors in the reference vector dictionary, and searching for a reference vector having a high similarity; A column of the obtained class (the similarity is equal to or larger than the threshold value and an arbitrary number is extracted from the highest similarity in descending order of similarity) is defined as a feature vector class sequence and corresponds to this feature vector class sequence. One having a qualitative expression search process for searching for a qualitative expression, and an output process for outputting a qualitative expression obtained by the qualitative expression search process with a similarity attached thereto A.

According to a second aspect of the present invention, there is provided a process state recognizing device.
A processing device that executes a feature extraction process, a vector search process, a qualitative expression search process, and an output process according to claim 1, and executes an editing process such as a correction of the reference vector and the qualitative expression as necessary; An input device for performing an input operation associated with each processing operation, an output device used for the output process, and a reference vector to be compared with a feature vector in the vector search process are stored together with a corresponding class and a firing area radius. A reference vector dictionary and a qualitative expression dictionary in which the feature vector class sequence is stored together with a corresponding qualitative expression.

According to the first and second aspects of the present invention, the temporal change tendency of the time-series process data at the measurement points of the plant or the like is sequentially digitized, and the time-series trend data is calculated based on the feature vector obtained therefrom. A class string (attribute value string) of the change tendency is obtained, and a qualitative expression suitable for expressing the change tendency is searched from the dictionary based on the class string. Then, in response to the search result, a qualitative expression with a similarity indicating the temporal change tendency of the time-series process data and related information as necessary are output from the output device.

According to a third aspect of the present invention, there is provided a data storage device, for a qualitative expression output by the output device of the process state recognition device according to the second aspect, a notification date and time, a process variable of the qualitative expression, and a process variable of the process variable. The information on the feature vector generated based on the temporal change tendency of the time-series process data up to the notification date and time, the information on the qualitative expression, and the qualitative expression desired to be presented instead of the qualitative expression, Is accumulated. In this invention, related information such as the notification time of a qualitative expression with similarity output according to the inventions of claims 1 and 2 is accumulated as learning data.

According to a fourth aspect of the present invention, there is provided a learning apparatus which uses the notification history data accumulated by the data accumulation apparatus according to the third aspect and qualitatively matches a user's intention with respect to a temporal change tendency of time-series process data. A learning process for moving the reference vector and correcting the firing area radius is performed so that an expression can be obtained.

A learning device according to a fifth aspect uses the notification history data stored by the data storage device according to the third aspect, and qualitatively matches a user's intention with respect to a temporal change tendency of the time-series process data. A learning process for automatically generating the reference vector is performed so that an expression can be obtained.

A learning device according to a sixth aspect uses the notification history data accumulated by the data accumulation device according to the third aspect, and qualitatively matches a user's intention with respect to a temporal change tendency of the time-series process data. A learning process for automatically eliminating the reference vector is performed so that an expression can be obtained.

According to the fourth and sixth aspects of the present invention, a learning process is performed using the data stored by the data storage device of the third aspect, and a reference vector in a reference vector dictionary used when searching for a qualitative expression. Move, generate, delete, and update the firing area radius.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. First, based on FIGS.
Embodiment 2 of the invention will be described. FIG. 1 shows the configuration of a process state recognition device, wherein 100 is a processing device composed of a computer and has a process state recognition program.

Reference numeral 500 denotes a term dictionary.
Reference numeral 00 denotes a reference vector dictionary 510 and a qualitative expression dictionary 520. The reference vector dictionary 510 has the components of the reference vector and the firing area radius in the antecedent part as shown in FIG. 5 described later, and stores the attribute value of each reference vector (hereinafter also referred to as a class, indicating how the data changes). Are stored in the consequent part. The center of gravity of the vector component calculated from the reference vector belonging to each class is also stored.

The qualitative expression dictionary 520 includes FIG.
As shown in (c), a large number of combinations having the attribute value column (hereinafter also referred to as a feature vector class column) in the antecedent part and having a qualitative expression indicating the change tendency of the process data in the consequent part are stored. Have been. Note that, in the antecedent part of the qualitative expression dictionary 520, the length of the feature vector class
The user can arbitrarily set a setting reference (maximum number and the like) regarding the class sequence definition length, which indicates the number of classes constituting the feature vector class sequence.

Referring back to FIG. 1, when the time-series process data is input from the measurement point of the plant, the processing apparatus 100 performs a process of extracting a characteristic with respect to the change tendency of the data, and the change of the time-series process data is performed. The feature extracted as the tendency (the attribute value sequence represented by the class of the reference vector dictionary 510) is checked against the antecedent part of the qualitative expression dictionary 520, and if they match, the qualitative expression of the consequent part is output as a search result. Output to the device 300. The term dictionary 500
The data necessary for setting, changing, and editing the contents of the data is input by the input device 200.

Hereinafter, the processing in the processing apparatus 100 will be described in detail. First, details of the feature extraction process of the time-series process data will be described with reference to FIGS. First,
In order to discretize the variation (dy / dt) of the time-series process data obtained as a continuous amount, the time-series data y (t) of the process variable is calculated from the values at the time t and the time (t-Δt) at the time t. The difference value e is obtained by Expression 1. Here, Δt is a sampling interval for each process variable.

[0024]

## EQU1 ## e = y (t) -y (t-.DELTA.t)

Next, a process for obtaining a qualitative value for the change (dy / dt) based on the difference value e is performed.
First, a difference value e is set for the process data to be processed.
The absolute value, exceeds the absolute value of the threshold alpha ₁ set in the process variable table 400 shown in FIG. 2, and the difference value e gives rise [+] as a qualitative value if positive value .

If the absolute value of the difference value e exceeds the absolute value of the threshold value α ₂ set in the process variable table 400 and the difference value e is a negative value, it falls as a qualitative value [− ]give. Here, if the number of thresholds α is increased, the types of qualitative values can be increased. In cases other than the above-described rise and fall, the qualitative value is set to equilibrium [0].
For example, when the time-series process data Y (t) changes as shown in FIG. 3A, the qualitative value of the change (dy / dt) is as shown in the lower part of FIG.

Next, with respect to the qualitative value relating to the change (dy / dt) of the time-series process data obtained by the above processing, the portion that has not changed continuously with the passage of time is determined by the following integration rules (1) to (1). A process of combining them into one is performed based on (3). The result of this processing is shown in FIG.

(1) One or more consecutive [+] s are held and used as one [+] while the number of consecutive times is used for use in a feature vector generation process described later. (2) One or more continuous [-] are retained and used as one [-] while the number of consecutive times is used for use in a feature vector generation process described later. (3) One or more consecutive [0] are held as one [0] while the number of consecutive times is retained for use in a feature vector generation process described later.

Next, the change of the sign of the qualitative value for the change (dy / dt) is taken as a characteristic point as shown in FIG. 3A, and the time-series process data as shown by a _{1 to} a ₆ in FIG. Is represented by a vector sequence divided by the feature points. This vector series is expressed on a Euclidean plane in which the horizontal axis represents time and the vertical axis represents the value of the time-series process data.

Here, the characteristic point is a sign change point of the following qualitative value.・ [+] → [-]: Change point from upward trend to downward trend ・ [+] → [0]: Change point from upward trend to invariant (equilibrium, constant) ・ [-] → [+]: Down Change point from trend to upward trend ・ [-] → [0]: Change point from downward trend to invariant (equilibrium, constant) ・ [0] → [+]: Change point from invariant (equilibrium, constant) to upward trend Change point ・ [0] → [-]: Change point from unchanged (equilibrium, constant) to downward trend

When a qualitative value sign change point as described above is detected, the time recorded as the current feature point is replaced with the time of the previous feature point, and the time (t−Δ
t) is stored as the time of the current feature point. At the same time, the value recorded as the current value is replaced with the previous value, and the value of the process data at the time (t−Δt) is stored as the current value.

The relationship between these data is as follows.・ Time of current feature point → time of previous feature point ・ Time (t−Δt) → time of current feature point ・ Current value → previous value ・ Process data value at time (t−Δt) → current value

Further, when it is determined that the feature point is present, a process for creating a feature vector (ΔX, ΔY) of the process data is performed based on the previous value (of the process value) and the current value.
Note that the latest feature vector is sequentially created based on the time of the current feature point and the current time that changes every Δt.

FIG. 4 shows a process of creating a feature vector. In the following description, it is assumed that the maximum number of class string definition lengths is set to 4 by the user, and the maximum number of feature vector class strings set in the antecedent part in the qualitative expression dictionary 520 is also 4. I do.

Here, one component of the feature vector, ΔX, is the difference ΔT between the time of the current feature point and the time of the previous feature point.
The ratio of t ′ to the maximum change set value ΔX ¹⁰⁰ (in the case of change in the + direction) or ΔX ^-100 (in the case of change in the − direction) registered in the process variable table 400, and the other component of the feature vector Is the maximum change setting value ΔY ¹⁰⁰ (+ direction) of the difference Δy between the value of the process data at the current feature point and the value of the process data at the previous feature point registered in the process variable table 400. Change),
Or ΔY ^-100 (for a change in the − direction). That is, the value of the process data is rescaled to the Δx component and the Δy component that form the coordinate system of the reference vector dictionary based on Equation 2.

[0036]

(Equation 2)

FIG. 4A shows a feature vector sequence expressed in the same manner as FIG. 3C, and FIGS. 4B and 4C show each vector a _1.
Diagram showing the relationship Delta] t 'and Δy per ~a _4, 4
(D) is a diagram showing the final feature vector (ΔX, ΔY). In this example, ΔX ¹⁰⁰ = 2, ΔY ¹⁰⁰ =
2.0, ΔY ^-100 = −2.0.

Next, the qualitative expression search processing using the term dictionary (qualitative expression dictionary 520) in FIG. 8C will be described in detail. In this process, a vector search process using the feature vector sequence generated by the feature extraction process for each process variable and the reference vector dictionary 510 in FIG. 5 is performed, and a change tendency of the time-series process data is accurately determined according to the search result. And a qualitative expression search process in a narrow sense for searching the qualitative expression expressed in the qualitative expression dictionary 520 in FIG. 8C.

First, the details of the vector search processing will be described. FIG. 5 shows an example of the reference vector dictionary 510 used in this processing. In the antecedent part of the dictionary,
A large number of reference vectors V _j (j = 1, 2,..., N) each including vector components ΔX and ΔY and a radius r of the circular firing region are defined, and in the consequent part, each reference vector changes. , And the center of gravity calculated by the reference vector belonging to the class is stored. Note that j is an index, and n is the number of reference vectors already defined in the reference vector dictionary 510.

FIG. 6 shows the feature vector components ΔX _ai , ΔY
_The process of deriving an attribute value sequence (feature vector class sequence) indicating a change tendency of process data from _ai (i = 1, 2,..., m) is shown. Note that i is an index, and m is the maximum value of the class column definition length.

In the vector search process, the similarity is determined by calculating the distance between the feature vector of the process variable to be processed and the component of the reference vector in the reference vector dictionary 510 in FIG. The reference vectors are searched from the reference vector dictionary 510 in descending order. As the reference vector, several classes classified into the same qualitative expression are grouped into one class (for example, a class “rapid rise” as a qualitative expression corresponding to class 1 in FIG. 5).
The class to which the reference vector belongs in the descending order of the similarity is added to the feature vector.

In FIG. 6, as a result of the similarity judgment using the reference vector dictionary 510 for the feature vectors a _{1 to} a ₄ , each of the feature vectors a ₁ , a ₂ , a ₃ , a ₄ becomes the reference vector V _12. , V ₃ , V ₁₆ , and V ₉ , and the classes 5, 2, 5, and 4 to which these reference vectors belong are added as high adaptability.

Next, the contents of the above-described similarity determination will be described with reference to FIGS. 7, 13 and 14. Reference vector dictionary 51
By determining the similarity between the reference vector V _{j of} 0 and each feature vector a _i , a maximum of n (for example, 2) classes are assigned to each feature vector a _i in descending order of class suitability.

First, m (for example, 3) classes having the center of gravity closest to the feature vector a _i are selected. next,
The similarity _Sij to _ai of the reference vector _Vj belonging to each selected class is obtained. Next, for each class C _k (for example, described as C ₈ for class 8), the class suitability S _ik for a _i is determined. Finally, n (for example, 2) classes having a high degree of class suitability S _ik for a _i are selected, and these classes are assigned to the class to which a _i belongs (the feature vector class C ^1).
_i and C ² _i ). By this processing, a feature vector class sequence corresponding to the feature vector sequence is assigned.

(1) Derivation of similarity S _ij The similarity S of each reference vector V _{j to} the feature vector a _i
ij uses the components (ΔX _vj , ΔY _vj ) and r _j of each reference vector V _j defined in the reference vector dictionary 510 and the components of the feature vector a _i (ΔX _aj , ΔY _aj ). hand,
It is calculated by the following Equations (1) and (2). Here, r _j is a circular firing region radius, and one firing region radius is set for each reference vector V _j . When a component (coordinate) of the feature vector a _i exists in a region surrounded by a circle having a radius r _j from a point having the component of the reference vector as the center coordinate, the similarity S _ij of V _{j to} a _i is 0 or more. Value.
When the coordinates of a _i and the coordinates (ΔX _vj , ΔY _vj ) of V _j match, the similarity S _ij becomes the maximum value of 100.

[0046]

## EQU3 ## _{d ij (a i, V j} ) = {(ΔX ai -ΔX vj) 2 + (ΔY ai -ΔY vj) 2} 1/2 ... (1) S ij = int [100 × {1- (D _ij / r _j )}] (2)

In Equation 3, d _ij is the Euclidean distance between the reference vector V _j and the feature vector a _i as can be seen from FIGS. Further, it is assumed that a negative value is allowed for the similarity _Sij . The concept of similarity is shown in FIG.
(B).

(2) Derivation of class suitability S _ik For each class, the class suitability S _ik for the feature vector a _i is determined. The similarities S _ij of all the reference vectors belonging to the class k (described as C _k ) are compared. And
The maximum value of S _ij is defined as the fitness S _ik of class k. As a specific example, when the reference vectors belonging to the class 2 (referred to as C ₂₎ had one third, the similarity of each reference vector for the feature vector a _i is calculated as 20,50,0, the maximum degree of similarity the value 50 is assigned as the class fit of the C ₂ for the feature vector a _i (denoted as S _12).

(3) Derivation of feature vector class The top n (for example, 2) classes having high relevance are selected for each feature vector a _i , and the feature vector class is used in qualitative expression search. (A) Feature vector The class with the highest degree of fitness for each a _i is the feature vector class C ¹ _i , the class with the second highest degree of fitness is the feature vector class C ² _i, and the degree of fitness for those classes is Are described as S ¹ _i and S ² _i . Here, the existence of C ¹ _i is allowed even if the degree of conformity S ¹ _i is a negative value, but the existence of C ² _i is not recognized when the degree of conformity S ² i is a negative value. .

For example, for the feature vector a ₄ , the fitness of class 1 (denoted by C ₁ ) is 50, the fitness of class 2 (denoted by C ₂ ) is 15, and the fitness of class 3 (denoted by C ₃ ) is If fitness is 40, fit the other classes is 0, the value of the class 1 is C ¹ _4, class 3 is the value of C ² _4, fitness 50 class 1 is S ¹ ₄ values, fitness 40 class 3 is the value of S ² _4. On the other hand, for the feature vector a ₂ ,
Fit of (C ₇ hereinafter) 15, if the fit of the other classes is 0, Class 7 is C ¹ ₂ values, 15 is the maximum fitness becomes a value S ¹ _2. Since class other fit exceeds 0 does not exist, it does not exist a C ² ₂ value and S ² ₂ values.

(B) Processing when the feature vector class is a minute change When C ⁿ _i (n = 1 or 2) is a class representing a minute change,
C ⁿ _i at that time in the feature vector class sequence is not included.

[0052] (c) a feature vector class sequence above processing by the feature vectors Class C ¹ _i for each feature vector a _i, C ² _i is obtained, the behavior of the process data that has been represented by a feature vector sequence "feature vector It can be expressed as a "class sequence". By this processing, a plurality of feature vector class strings corresponding to the feature vector strings are obtained. According to the embodiment, feature vector class sequences (5254), (5354), (5251), and (5254) corresponding to the feature vector sequence (a ₁ , a ₂ , a ₃ , a ₄ ).
(5351) can be obtained.

Next, the qualitative expression search processing in a narrow sense will be described in detail. In this processing, the antecedent part that matches the feature vector class sequence of the time-series process data is searched from the qualitative expression dictionary 520, and the change tendency of the time-series process data is accurately determined from the consequent part corresponding to the antecedent part. Search for the qualitative expression shown in.

FIG. 8 shows the procedure of this processing procedure. First, a feature vector class sequence (FIG. 8 (b)) generated based on the feature vector sequence (FIG. 8 (a)) of the time-series process data obtained in the feature extraction process is stored in a qualitative expression dictionary 520 (FIG. c) Match against the antecedent part. And
If there is a match, the qualitative expression of the consequent part defined corresponding to the antecedent part is extracted.

Next, a qualitative expression that matches the feature vector class sequence is searched. The feature vector class sequence is one in which feature vector classes of certain process data are continuously arranged from the present time to the past. However, one feature vector class can take a maximum of n (for example, ^{two of} C ¹ _i and C ² _i ). Therefore, while there is only one feature vector sequence, one or more feature vector class sequences can be generated as shown in FIG.

(A) Criteria for Generating a Feature Vector Class Sequence When generating a feature vector class sequence, a criterion is set in order to prevent generation of an unnecessarily large number of feature vector class sequences. For example, if only C ¹ _i is used, only one feature vector class sequence is generated. If C ¹ _i and C ² _i are considered only for one of the feature vectors, ^two types of feature vector class strings can be generated. In the example of FIG. 6, since C ¹ _i and C ² _i are considered for ^two feature vectors, four types of feature vector class sequences are generated. Here, C
An example is shown in which ² _i is taken into account in two places of class strings, and the number of generated feature vector class strings is limited to four.

Firstly, the selection criteria of the position of the feature vector classes take into account the C ² _i (index i of a _i), the difference between the fit S ¹ _i and S ² _i is smaller position. Next, for each of the generated feature vector class strings, if there is a class representing a minute change among the classes constituting each class string, a class string excluding the minute change class is replaced with the class string in place of the class string. Used for dictionary matching processing. For example, when the acquired feature vector class sequence is (9294) and the class representing the minute change is 9, the class sequence used for the matching process with the class sequence dictionary is (24). At the same time, the class sequence registered in the class sequence dictionary does not include a class representing a minute change.

(B) Matching process between feature vector class sequence and class sequence dictionary The feature vector class sequence generated / selected by such a method is used, and then a feature vector class sequence dictionary (hereinafter also referred to as a class sequence dictionary). ) Is performed. The class sequence dictionary defines a class sequence representing a behavior pattern of the process data. In the class sequence registered in this dictionary, a qualitative expression suitable for the behavior pattern is registered as shown in FIG. If the generated feature vector class sequence (arbitrary number from the present time or from the time specified by the user to the past) matches the class sequence registered in the class sequence dictionary, it is related to the class sequence. The qualitative expression is presented to the user.

(C) Qualitative Expression Search Processing The qualitative expression search processing procedure is as follows. First, one is extracted from the generated feature vector class sequence ((5254) in FIG. 8), and is set as a collation target. The following processing is performed similarly for all generated feature vectors. Since the feature vector sequence has been replaced with a class sequence by the vector search process,
The qualitative expression dictionary is collated from the current time side of the class string. When a plurality of qualitative expressions match the class sequence, a qualitative expression related to the long-term tendency (the longest feature vector class length in the qualitative expression dictionary) is selected.

(1) Collation with Class Sequence Dictionary The generated feature vector class sequence is collated with a class sequence dictionary (a class sequence dictionary is a collection of combinations of feature vector classes and qualitative expressions). First, a partial class sequence is derived from the feature vector class sequence. A partial class sequence is a sequence of classes that are arranged continuously from the present to the past in the feature vector class sequence, and the length (the number of classes constituting the class sequence) is maximum defined from only one. Even those with a long length are prepared. The maximum definition length is the maximum value (for example, 8) of the length of the class string that can be registered in the class string dictionary.

FIG. 8 shows an example in which the maximum definition length is 4, and the matching is performed on the feature vector class sequence (5254). 4 generated from the feature vector sequence (5254)
Each of the class strings of the two parts is compared with the contents of the class string dictionary, and a qualitative expression associated with the class string in the dictionary is assigned to the matched partial class string.

(2) Acquisition of a qualitative expression A qualitative expression associated with the longest partial class sequence among the partial class sequences matching the class sequence defined in the class sequence dictionary is obtained and output. (3) Calculation of the degree of conformity At this time, the degree of conformity to the data behavior of the qualitative expression (output) is calculated using the class conformity added to each class constituting the selected (longest) class sequence. And output it together with the qualitative expression.

The conformity of the qualitative expression for the selected class sequence is a value obtained by averaging the sum of the class conformances of the classes constituting the selected partial class sequence by the length (definition length) of the partial class sequence. (In the example of FIG.
5 "has been selected). When there are a plurality of antecedent parts of the qualitative expression dictionary 520 that match the feature vector class sequence, for example, the qualitative expression dictionary 520 whose definition length of the class sequence matches the longest feature vector class sequence (FIG. 8 (d)).

With the above processing, a qualitative expression representing a change tendency of the time-series process data is searched from the qualitative expression dictionary 520 from the characteristic vector sequence and the characteristic vector class sequence of the time-series process data generated in the characteristic extraction process. can do.

Next, the details of the term dictionary editing process will be described. In the term dictionary editing process of this embodiment, a basic change pattern of a process variable can be registered as an initial term dictionary content before use is started, and the already registered content is appropriately modified and changed. Is available. First, the reference vector dictionary 5
In the editing process for the reference 10, the reference vector dictionary 510 can be created by changing the settings of the components (ΔX _vj , ΔY _vj , r _vj ) of the reference vector and the class to which the vector belongs.

In addition, in this processing, the user can request correction of data in the reference vector dictionary 510 using a learning device described later, that is, a learning process. Then, in the editing processing for the qualitative expression term dictionary 520, the user changes the setting of the feature vector class sequence a _i and the corresponding qualitative expression, so that the qualitative expression dictionary 52
0 can be created.

As described above, in the first and second aspects of the present invention, the change tendency of the time series process data obtained from the measurement points in the plant is represented by the feature vector sequence (a ₁ , a
₂ ,...), And a similarity determination between the feature vector sequence and each reference vector of the reference vector dictionary 510 is performed to generate a feature vector class sequence of the time-series process data. By searching the qualitative expression dictionary 520 using the qualitative expression dictionary, it is possible to search for a qualitative expression that accurately represents the change tendency of the time-series process data.

Next, an embodiment of the present invention will be described with reference to FIGS. FIG. 9 shows a configuration of a data storage device according to the third aspect of the present invention. After the qualitative expression indicating the change tendency of the time-series process data selected by the process state recognition device according to the second aspect of the present invention is output from the output device 300, the qualitative expression is incorrect by a user such as a driving operator. Suppose that it was notified. At this time, if a correction input indicating that the qualitative expression is incorrect is performed via the input device 200, the qualitative expression designated as incorrect, the date and time of notification, the process variable name, and the correct qualitative value desired by the user Expressions and the like are created by the processing device 100 and are additionally stored in the notification history data 600 used for the learning process.

Hereinafter, the maximum number of class string definition lengths as described above
Is presumed to have been set to 4 by the user
I will tell. FIG. 10 is an example of a correcting means of the term dictionary 500.
You. The processing in the process state recognition device according to the second aspect of the present invention.
Vector (a) generated by the logical device 100₁, a_Two,
a _Three, a_Four) Is a feature vector class sequence (5252).
The user, such as a driving operator,
If you select the correct qualitative expression you want to present,
Such correction processing is performed.

(1) Correction method 1 If the class sequence (5252) is not defined in the qualitative expression dictionary 520, a new class sequence (5252) is inserted into the qualitative expression dictionary 520, and the corresponding qualitative expression (Stationary 3) is set. (2) Correction method 2 Although the class sequence (5252) is defined in the qualitative expression dictionary 520, the set qualitative expression is (hunting).
And if the user wants to present a different qualitative expression (steady state 3), and if the user selects to change the qualitative expression in the qualitative expression dictionary 520, The set qualitative expression (hunting) is rewritten to a qualitative expression desired by the user (steady state 3).

(3) Modification Method 3 Although the class sequence (5252) is defined in the qualitative expression dictionary 520, the set qualitative expression is (hunting).
And if the user does not want to change the qualitative expression in the qualitative expression dictionary 520, the learning device described later The contents of the reference vector dictionary 510 are updated. FIG. 11 shows a part of the notification history data 600. By accumulating such notification history data 600, it can be used for updating the term dictionary 500 as described below.

Next, an embodiment of the present invention will be described with reference to FIG. FIG. 12 shows the configuration of the learning apparatus according to the fourth aspect of the present invention. The processing device 100 in the learning device is configured to transmit the notification history data 600 stored by the data storage device according to the third aspect of the present invention.
And a learning program for correcting the contents of the term dictionary 500 using the time-series process data obtained in the past from the measurement points of the plant stored in the database 700.

Hereinafter, the learning process performed by the processing device 100 will be described. When there is a request for a learning process in the term dictionary editing process performed by the processing device 100 of the process state recognition device according to the second aspect of the present invention, the data in the reference vector dictionary 510 is corrected, which is desirable for the feature vector sequence. Make sure the class is added.

Assuming that the time series process data of the process variable name registered in the notification history data 600 of the modification method 3 selected by the user in FIG. 11 is before the notification date and time also registered in the notification history data 600. , A (a ₁ , a ₂ , a ₃ ,...), A generated vector class sequence (hereinafter referred to as a feature vector sequence A) In the case where it is specified that the invention is a class sequence added by the inventions of the items 1 and 2, it is called as follows.) Is P (p ₁ , p ₂ ,
p ₃ ,...) On the other hand, if a feature vector class sequence that the user desires to express as the behavior of A is Q (q ₁ , q ₂ , q ₃ ,.
A specific reference vector movement processing procedure is as follows.

(1) A feature vector sequence A and a generated vector class sequence P are prepared. At this time, the number of vectors constituting A and the number of classes constituting P are greater than or equal to the maximum number of class sequence definition lengths already set by the user at the start of this processing. In this embodiment, the maximum number is four, and the number of vectors of A and the number of classes of P are four.
And

(2) Based on a desirable qualitative expression in the notification history data 600, a class string forming a pair with the qualitative expression is extracted from the qualitative expression dictionary 520 and displayed on the output device 300. (3) From among the displayed class strings, the user designates a class string Q of a desired expression as the behavior of the time-series process data (the behavior of A).

(4) The class strings P and Q are compared, and both classes are compared.
Perform the following processing for different class parts in the column.
Now. For example, P (p₁, p_Two, p_Three, p_Four) Is (5252)
, Q (q₁, q_Two, q_Three, q_Four) Is (5153),
Part of a different class (p_TwoAnd q_Two) And (p_FourAnd q _Four)
, The following learning process is performed.

(5) Among the reference vectors V _j belonging to each class of p _i and q _i , one having the highest similarity (the largest S) to the feature vector a _i is prepared, and a total of two are prepared. I do. Here, d ² is calculated by Equation 4 and similarity S is calculated by Equation 5. Note that in Equation 5, r _v is the radius of the ignition area possessed by each reference vector.

[0079]

D ² (a _i , V) = (ΔX _ai −ΔX _v ) ² + (ΔY
_ai −ΔY _v ) ²

[0080]

S (a _i , V) = int [100 × {1- (d / r _v )}]

(6) Regarding the prepared vector, let it be a reference vector V _p (ΔX _p , ΔY _p ) belonging to the class p _i and closest to the feature vector a _i , while belonging to the class q _i and The reference vector closest to the feature vector a _i is defined as V _q (ΔX _q , ΔY _q ). (7) For the similarities S _p and S _q with respect to a _i , the processing of Expression 6 is repeated in the manner of FIG. 15 until V _q (correct answer) becomes higher than V _p (error).

[0082]

[6] _{ΔX p = ΔX p -α (1} -S p) (ΔX a -ΔX p) ΔY p = ΔY p -α (1-S p) (ΔY a -ΔY p) ΔX q = ΔX q + α _{(1-S q) (ΔX} a -ΔX q) ΔY q = ΔY q + α (1-S q) (ΔY a -ΔY q) r q-new = γ × r q r p-new = β × r p

In Equation 6, the processing end condition is S _q
> And S _p. Α, β, and γ are correction rates preset as learning parameters. As a result, the content of the reference vector dictionary 510 is automatically corrected to a new one by the learning process.

Therefore, the user does not need to directly update the reference vector dictionary 510, and the notification history data 600 is accumulated according to the invention of claim 3, and the reference vector is obtained by learning processing using the notification history data 600. The dictionary 510 can be improved. Therefore, the optimum class is added to the feature vector sequence by the similarity determination, and the qualitative expression according to the feature vector class sequence is updated by the qualitative expression dictionary 520 to another one.

Next, an embodiment of the invention will be described with reference to FIG. The configuration of this embodiment is also shown in FIG.
The learning process performed by the processing device 100 will be described below. When there is a request for a learning process in the term dictionary editing process performed by the processing device 100 of the process state recognition device according to the second aspect of the present invention, the data in the reference vector dictionary 510 is corrected, which is desirable for the feature vector sequence. Make sure the class is added.

Suppose that the time series process data of the process variable name registered in the notification history data 600 of the modification method 3 selected by the user in FIG. 11 is before the notification date and time also registered in the notification history data 600. A (a ₁ , a ₂ , a ₃ ,...) And P (p ₁ , p ₂ , p ₃ ) represent a feature vector sequence generated using the first and second aspects of the present invention. ,...), On the other hand, assuming that the feature vector class sequence that the user desires to represent as the behavior of A is Q (q ₁ , q ₂ , q ₃ ,...), A specific reference vector generation processing procedure is as follows. It is.

In the learning process, after selecting a reference vector to be corrected one by one from the error class and the correct answer class, if the following condition is satisfied before the learning is started, a reference vector generation process is performed. . As shown in FIG. 16, a reference vector (referred to as a correct vector) v _q (ΔX _q , ΔY _q ) and a feature vector (correct vector learning target) a _i (ΔX _a , ΔY _a ) belonging to the correct class, Distance d between
Is greater than or equal to the threshold φ ₁ · r _q, a new reference vector belonging to the correct class is generated at the position of the feature vector. Here, φ ₁ is the firing region radius r _q of the correct answer vector.
Is the ratio of the distance d to This distance d and the threshold φ
Equation ₁ is shown in Equation 7. After the new generation of the reference vector, the process ends without performing the learning process.

[0088]

D = {(ΔX _q −ΔX _a ) ² + (ΔY _q −ΔY _a ) ² } ^1/2 φ ₁ = 1.4

The newly generated vector has the same component (ΔX _q-new , ΔY _q-new ) as that of the feature vector (ΔX _a , ΔY _a ). and phi ₂ times the value of the ignition area radius r _q with a radius r _{q-new new} the correct vector v _q.

[0090]

[Equation 8] _{φ 2 = 0.4 ΔX q-new} = ΔX a ΔY q-new = ΔY a r q-new = φ 2 r q

Next, an embodiment of the present invention will be described with reference to FIG. The configuration of this embodiment is also shown in FIG.
The learning process performed by the processing device 100 will be described below. When there is a request for a learning process in the term dictionary editing process performed by the processing device 100 of the process state recognition device according to the second aspect of the present invention, the data in the reference vector dictionary 510 is corrected, which is desirable for the feature vector sequence. Make sure the class is added.

Assuming that the time-series process data of the process variable name registered in the notification history data 600 of the modification method 3 selected by the user in FIG. 11 is before the notification date and time also registered in the notification history data 600. from the process data, the feature vector sequence generated using the invention of claim 1 and _{2 a (a 1, a 2} , a 3, ...), the product vector class sequence _{P (p 1, p 2,} p 3 , ...), on the other hand, the feature vector class sequence that the user wishes to express as the behavior of A
Assuming (q ₁ , q ₂ , q ₃ ,...), The specific reference vector disappearance processing procedure is as follows.

After the correction (learning) of the correct answer vector and the error vector is completed, if the value of the firing region radius r becomes extremely small for each vector, or if it becomes extremely different from other reference vectors belonging to the same class, When approaching, the reference vector is deleted / updated.

(1) Disappearance based on firing area radius r
17 (b)) The reference vector (error vector v_p,
Correct answer vector v_q), The firing area after the learning process
Area radius (r_p, r_q) Is the threshold φ_ThreeIf the following (r
≤φ_Three), The reference vector is removed from the reference vector dictionary.
Good. Where φ _Three= 0.1.

(2) Similarity criterion S_{jj '}Disappeared / updated by
(FIG. 17 (c)) For each of the reference vectors also subjected to the learning process,
In the firing region of the error (correct answer) vector after the learning process,
Other reference vector v belonging to the error (correct) class _{p '}(v_{q '})
Exists and v_pV_{p '}Similarity S to_{pp '}(r_pBased on
Threshold value φ_Four(Exemplified in Equation 9 below)
(S_{pp '}≧ φ_Four), V_pUpdate of the components and v
_{p '}Is deleted from the reference vector dictionary.

[0096]

_{100 S pp '= int [100} × {1- (d pp' / r p)}] of the number 9 φ ₄ = 0.95 × maximum similarity value _{ΔX p-new = ΔX p +} η {r p ' _{/ (r p '+ r p} )} (ΔX p'
_{-ΔX p) ΔY p-new =} ΔY p + η {r p '/ (r p' + r p)} (ΔY p '
_{-ΔY p) r p-new =} max (r p ', r p)

The values of φ ₁ , φ ₂ , φ ₃ , φ ₄ , η are common to all process variables. This makes it possible to automatically recognize the time-varying tendency of the time-series process data, greatly reducing the burden on the user.

[0098]

As described above, according to the first and second aspects of the present invention, the antecedent part that matches the class sequence of the time-series process data derived by the processing device is searched from the qualitative expression term dictionary, and the subsequent By outputting the qualitative expression described in the section as a search result, the user is provided with a qualitative expression suitable for expressing the change tendency of the time-series process data. In addition, it is possible to output a plurality of qualitative expressions with respect to the behavior of one time-series process data, and in addition, to each of the qualitative expressions, an expression conformity indicating a degree of matching of the data behavior to the expression is added. Thus, the reliability of the output is improved.

Further, according to the third aspect of the present invention, the qualitative expression output from the process state recognition device according to the second aspect of the present invention can be correctly performed by a user such as an operation operator pointing out the correctness of the process state recognition result. The qualitative expression to be notified, the notification time at which the notification was made, and the like are stored as learning data.

The invention of claims 4 to 6 is the invention of claim 3
By executing the learning process using the learning data accumulated according to the invention of the present invention, the contents of the term dictionary are improved and refined, which is useful for the recognition of the subsequent process state.

Therefore, according to these inventions, an enormous amount of knowledge for grasping a change tendency necessary for monitoring a large number of process data can be automatically obtained without storing in a document or the like. The recognition result can be provided in an easy-to-understand expression. For this reason,
The burden on the user is reduced, and special data analysis and expert knowledge are not required. In addition, the learning function can absorb know-how in estimating the expression that indicates the tendency of change in process data, and can absorb subtle differences that occur in each plant. If there is an error in the qualitative expression search, the qualitative expression And the contents of the term dictionary used for the search can be easily corrected.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a process state recognition device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a process variable table.

FIG. 3 is an explanatory diagram of a feature extraction process performed by the process state recognition device.

FIG. 4 is an explanatory diagram of a process of deriving a component of a feature vector by the process state recognition device.

FIG. 5 is an explanatory diagram of a reference vector dictionary.

FIG. 6 is an explanatory diagram of a feature vector class sequence deriving process by the process state recognition device.

FIG. 7 is an explanatory diagram of a distance d between a feature vector and a reference vector.

FIG. 8 is an explanatory diagram of a qualitative expression search process performed by the process state recognition device.

FIG. 9 is a configuration diagram of a data storage device according to an embodiment of the third invention.

FIG. 10 is an explanatory diagram of a term dictionary correction method.

FIG. 11 is an explanatory diagram of accumulated notification history data.

FIG. 12 is a configuration diagram of a learning device according to an embodiment of the present invention.

FIG. 13 is an explanatory diagram of a concept of similarity.

FIG. 14 is an explanatory diagram of a process of deriving a feature vector class.

FIG. 15 is an explanatory diagram of a reference vector moving process by learning.

FIG. 16 is an explanatory diagram of a reference vector generation process by learning.

FIG. 17 is an explanatory diagram of a process of disappearing a reference vector by learning.

[Explanation of symbols]

 REFERENCE SIGNS LIST 100 processing device 200 input device 300 output device 400 process variable table 500 term dictionary 510 reference vector dictionary 520 qualitative expression dictionary 600 notification history 700 database

Claims (6)

[Claims] 1. A feature extraction process for generating a sequence of feature vectors based on a temporal change tendency of time-series process data, and a plurality of references in the reference vector dictionary and the feature vector obtained by the feature extraction process A vector search process for calculating a similarity to a vector and searching for a reference vector having a high similarity; and a sequence of classes corresponding to the reference vectors obtained by the vector search process as a feature vector class sequence,
A qualitative expression search process for searching for a qualitative expression corresponding to the feature vector class sequence; and an output process for outputting a qualitative expression obtained by the qualitative expression search process with a similarity attached thereto. Process state recognition method. 2. A feature extraction process, a vector search process, a qualitative expression search process, and an output process according to claim 1, and an editing process such as a correction of the reference vector and the qualitative expression as necessary. A processing device; an input device for performing an input operation associated with each of the processing operations; an output device used for the output process; and a reference vector to be compared with a feature vector in the vector search process. A process state recognition device, comprising: a reference vector dictionary stored with a radius; and a qualitative expression dictionary in which the feature vector class sequence is stored with a corresponding qualitative expression. 3. A qualitative expression output by the output device of the process state recognition device according to claim 2, a notification date and time, a process variable of the qualitative expression, and a time-series process up to the notification date and time of the process variable. Information on a feature vector generated based on a temporal change tendency of data, information on the qualitative expression, and a qualitative expression desired to be presented instead of the qualitative expression are stored as notification history data. Data storage device. 4. A qualitative expression that matches a user's intention with respect to a temporal change tendency of time-series process data using notification history data accumulated by the data accumulation device according to claim 3. A learning apparatus for executing a learning process for moving the reference vector and correcting the firing area radius. 5. A qualitative expression that matches a user's intention with respect to a temporal change tendency of time-series process data using notification history data accumulated by the data accumulation device according to claim 3. A learning device for executing a learning process for automatically generating the reference vector. 6. A qualitative expression that matches a user's intention with respect to a temporal change tendency of time-series process data using notification history data accumulated by the data accumulation device according to claim 3. A learning device for executing a learning process for automatically eliminating the reference vector.