JP4268500B2 - Process state similar case search method, state prediction method, and storage medium - Google Patents

Description

  In the present invention, during operation of a process such as a blast furnace, a method for searching past cases similar to the operation state, a method for predicting the future state of the process, and a program for causing a computer to execute the processing method are recorded. The present invention relates to a computer-readable storage medium.

  Conventionally, in the operation of a process such as a blast furnace, when an abnormal state occurs, an act of searching for a past case manually and responding to it has been performed. It is important to use past knowledge for future operational improvements, regardless of success or failure, but in the past there was no way to make full use of accumulated data, and it was common to rely on human memory. there were. Therefore, due to differences in personal experience, it became a problem because of different actions of opinion.

On the other hand, techniques using case-based reasoning are used in several fields.
Japanese Patent Laid-Open No. 3-132826 (Patent Document 1) proposes a technique that applies case-based reasoning that solves a current problem based on past problem-solving cases.
Japanese Laid-Open Patent Publication No. 7-271588 (Patent Document 2) proposes a tabular editor for case-based reasoning to support systematization of specialized knowledge.

  As a wastewater amount prediction system and a turbidity prediction system for a water purification plant process, Japanese Patent Application Laid-Open No. 2001-288882 (Patent Document 3) and Japanese Patent Application Laid-Open No. 2002-199956 (Patent Document 4) collectively represent a plurality of cases. A case is created, a case base is constructed using the created representative case, and a process is predicted.

Japanese Patent Laid-Open No. 3-132826 JP 7-271588 A JP 2001-288882 A Japanese Patent Application Laid-Open No. 2002-119956

  By the way, the process targeted by the present invention is a complex, non-linear and non-stationary process such as a blast furnace. A mathematical model for the process has been developed but is not yet sufficient. For this purpose, past operational knowledge and past cases are constructed and used as a knowledge base and case base. However, there is a problem that labor is spent collecting knowledge, building a case base, and updating it. In addition, if the update is not sufficient, there is a problem that the accuracy of search and prediction decreases.

Furthermore, in the methods described in Japanese Patent Application Laid-Open Nos. 2001-288787 and 2002-199956, cases are collected, and there is a problem that it is impossible to specify a time point similar to the current time point.
Moreover, since it is necessary to create a case base for prediction from a plurality of cases in advance, the update is a problem. For this reason, a learning section for cases has been proposed, but problems remain in maintaining permanent performance.

  The present invention has been made in view of such circumstances, and can search past cases of complicated, non-linear, and non-stationary processes at high speed and with high accuracy so that future prediction can be accurately performed. The purpose is to enable online prediction that does not require updating.

In order to achieve such an object, the process state similar case search method according to the present invention uses a computer to store process data including measured values of a plurality of process variables indicating the storage time and the order in which the process data is stored. In a method for searching past state similar cases of processes similar to the process state at a specified time from a time series database stored for each time of a fixed period together with a number ,
(B) for each of the plurality of process variables, a step of a series of process data from the specified time Tk until only time in the past predetermined number m, and out the time series from the database extract,
(B) Independent component analysis is performed on the extracted series of process data, and a plurality of independent variable values s composed of a series of data values at a predetermined time point m from the designated time Tk ; Calculating a separation matrix W ;
(C) The calculated series of a plurality of independent variable values s are used as input data, or the calculated series of the plurality of independent variable values s and the process variable selected in advance among the plurality of process variables. A series of process data is used as input data, and the data value Xi, which is the value of the input data at the time Tk and the past time from the time Tk, is normalized using upper and lower limit values Ximax and Ximin. Next , calculating a data value at time Tk in the search table by calculating a quantized input value Ixi by multiplying by a predetermined quantization number and then converting to an integer .
And storing (d) combined at least any one of the stored numbers of the time series database corresponding to the specified time Tk, the data value at time Tk of the retrieval table in the search table,
A process of creating a search table by repeating a series of steps including changing the specified time Tk, and for each of a plurality of process variables, the predetermined score from the specified time Ts different from the time Tk A step of obtaining a series of process data up to a point in time past m, and a series of process data from the extracted specific time Ts is multiplied by the separation matrix W calculated in the step (b) to obtain a specific time Ts calculating a German Tatsuhen figures, for the independent variable values for a specific time Ts that the calculated, then the process of step (c), a step of calculating the input data for a search for a particular time Ts, as a search key input data for the search of the specific time Ts, the data value at each past time Tk of the search table and the input data value, the value of the corresponding data available The similarity that is the difference between the data values is calculated, and the similarity is used to search for a similar data value up to the vicinity of a preset value of the similarity, thereby obtaining a similar case, and the past time the storage time or with string to a data value and obtaining a store number was engraved upon storage of the similar case is equivalent to a store number, extracts a set of process data similar case from the time series data base And a process .

The process state prediction method according to the present invention uses a computer to store process data composed of measured values of a plurality of process variables, together with a storage time and a storage number indicating the storage order, at every fixed period of time. In a method of predicting a future state of a process by searching a past state similar case of a process similar to a process state at a specified time from a series database,
(B) for each of the plurality of process variables, a step of a series of process data from the specified time Tk until only time in the past predetermined number m, and out the time series from the database extract,
(B) Independent component analysis is performed on the extracted series of process data, and a plurality of independent variable values s composed of a series of data values at a predetermined time point m from the designated time Tk ; Calculating a separation matrix W ;
(C) The calculated series of a plurality of independent variable values s are used as input data, or the calculated series of the plurality of independent variable values s and the process variable selected in advance among the plurality of process variables. A series of process data is used as input data, and the data value Xi, which is the value of the input data at the time Tk and the past time from the time Tk, is normalized using upper and lower limit values Ximax and Ximin. Next , calculating a data value at time Tk in the search table by calculating a quantized input value Ixi by multiplying by a predetermined quantization number and then converting to an integer .
And storing (d) combined at least any one of the stored numbers of the time series database corresponding to the specified time Tk, the data value at time Tk of the retrieval table in the search table,
A process of creating a search table by repeating a series of steps including changing the specified time Tk, and for each of a plurality of process variables, the predetermined score from the specified time Ts different from the time Tk A step of obtaining a series of process data up to a point in time past m, and a series of process data from the extracted specific time Ts is multiplied by the separation matrix W calculated in the step (b) to obtain a specific time Ts A step of calculating the independent variable value of the above, a step of processing the step (c) for the calculated independent variable value of the specific time Ts, and calculating input data for search at the specific time Ts; as a search key input data for searching for a specific time Ts, the data value at each past time Tk of the search table and the input data value, the value of the corresponding data available The similarity that is the difference between the data values is calculated, and the similarity is used to search for a similar data value up to the vicinity of a preset value of the similarity, thereby obtaining a similar case, and the past time a step of determining the storage time or store numbers with string to data values from the storage time of the similar case, the process data or storage number of the similar case, up to a specified earlier time, the designated And taking out a series of process data up to the previous storage number from the time series database , and making the taken out series of process data a future state of the process.

In addition, the process state similar case search method according to the present invention has a time series database for a blast furnace process, and the process variable values are the hot metal temperature, pulverized coal injection amount, solution loss carbon, heat flow ratio, charging pitch, Si amount. At least one of hot air temperature, furnace top temperature, and Q loss is selected.
In addition, the process state prediction method according to the present invention displays the future state of the process, and at that time, as the accuracy of the prediction, corresponds to the similarity of the similar cases used to obtain the series of process data taken out . It is displayed with reliability.

The computer readable storage medium according to the present invention has remembers a program for executing the respective steps in a computer according to any one of the invention.

According to the present invention, it is possible to quickly search for similar cases of complicated, non-linear and non-stationary processes. If there is a time series database, the similarity search apparatus can be systematically updated by the method of the present invention.
Furthermore, since a plurality of process variable values are converted into independent component variable values, the number of variables can be reduced. Therefore, the search table becomes small and the search speed can be improved.
In addition, this makes it possible to always maintain the performance of the apparatus for realizing the method of the present invention with high accuracy.
In addition, this makes it possible to always maintain the process state prediction performance with high accuracy.
Furthermore, accurately predicting the future state of the process provides important guidance for determining future operation actions and greatly contributes to the stabilization of operations.

Hereinafter, an embodiment of a method and apparatus for searching similar cases of the process of the present invention will be described.
FIG. 1 is a block diagram showing a configuration of an apparatus used for carrying out the method for searching similar cases of the process of the present invention.
In FIG. 1, the blast furnace process 10 is provided with a plurality of various sensors for measuring temperature, pressure, component, position, and the like.
Reference numeral 20 in FIG. 1 is a measurement / control device for the blast furnace process 10, which presents time-series data of various process information measured from the blast furnace process 10 to an operator of the process. The process control operation is performed. In the following description, the time series data of the process information is referred to as process variables, process data, or time series data as necessary.

  The various measured process data is stored in the process time series database 30 at regular time intervals. Actually, the storage cycle differs depending on the type of process data, but it is assumed here that the data is stored at a constant cycle. When the period is not constant, the long period data can be converted into the short period data by holding the value of the long period data in accordance with the short period data.

FIG. 2 shows a storage format of the process time series database 30.
Process variables stored in the process time series database 30 are hot metal temperature, pulverized coal injection (PCR), solution loss carbon (SLC), heat flow ratio, charging pitch, Si content, hot air temperature (Tf), furnace top Gas temperature, heat load (Q loss), etc. The storage number is a number indicating the order in which the time series data is stored. The storage time (storage DATE) is the time when the time series data is stored, and is expressed in year / month / day / hour / minute / second.

Reference numeral 35 in FIG. 1 denotes an independent component analyzer that analyzes an independent variable of a process variable and obtains an independent variable value. Independent component analysis allows a large number of process variables to be aggregated into a small number of independent components. Independent component analysis includes s as an independent signal vector (independent variable), x as a mixed signal vector, A as a mixing matrix, n as the dimension of the independent signal vector s, m as the dimension of the mixed signal vector x, and as the dimension of the mixing matrix A. This is a technique for calculating a separation matrix W for obtaining an independent signal vector s and an independent signal vector s from a mixed signal vector x constituted by the following (formula 1) when (m * n).
x = As (1 formula)

When the separation matrix W is obtained, the independent signal vector s is calculated by the following (Equation 2).
s = Wx (2 formulas)

  A number of independent component analysis methods have been proposed. Here, a method called high-speed ICA is used. For example, it was proposed in the paper “Fast and Robust Fixed-Point Algorithms for Independent Component Analysis” by Aapo Hyvarian et al. (Aapo Hyvarian, IEEE Transactions on Neural Networks. Vol. 10, No. 3, May 1999, P. ing. It is assumed that the mixed signal vector x is centered as preprocessing, that is, converted so that the average value becomes zero. FIG. 3 shows an outline flow of independent component analysis.

The basic algorithm of the high-speed ICA first whitens the mixed signal vector x (step 36 in FIG. 3).
Whitening uses eigenvalue decomposition of the covariance matrix (Σ [xx ^T ] = EDE ^T ) of the mixed signal vector x. Here, E is an orthogonal matrix of eigenvectors, and D is a diagonal matrix of its eigenvalues. A diagonal matrix D of eigenvalues is expressed by the following (Expression 3).
D = diag (d ₁ , ..., d _n ) (3 formulas)

The whitened signal vector z can be obtained by the following (Equation 4).
z = ED ^−1/2 E ^T x (4 formulas)
Here, D ^−1/2 is a matrix calculated by the following (formula 5).
D ^-1/2 = diag (d ₁ ^-1/2 , ..., d _n ^-1/2 ) ... (5 formulas)

Next, a separation matrix W is obtained from the whitened signal vector z (step 37 in FIG. 3).
One row of the separation matrix W is set to w ^T (referred to as a load vector), and the separation matrix W is obtained as follows.

1) An initial value of the load vector w is set (for example, a random number).
2) Perform the following calculation (6).

Here, g is set to a function such as the following (Expression 7) or (Expression 8).
g (u) = tanh (u) (7 formulas)
g (u) = u exp ( -u 2/2) ··· (8 type)
In (Equation 6), μ is a learning coefficient and is set to 1 (μ = 1). However, if there is a problem with convergence, for example, the value is smaller than 1 such as 0.1 or 0.01. Set.
3) The following (Equation 9) is calculated.

4) If the load vector w has not converged, return to 2) and repeat 2) and 3).

Next, different to weight vectors w are prevented from converging in the same direction, the output value w ^T ₁ z after each _{^{iteration, w T 2 z, ... w}} n T z is the i independently so that the decorrelation The orthogonalization of the component load vector w is performed by the following processing.
5) The following (Equation 10) is calculated.

The separation matrix W is obtained by the fast ICA algorithm, and the independent component variable value is calculated from the following (Equation 11).
s = Wz = WED ^−1/2 E ^T x (Expression 11)

The mixed signal vector x can be calculated by giving a process variable value as shown in the following equation (12).
x = {value of process variable 1, value of process variable 2, ...} ^T (12)
This separation matrix W is also used as it is for independent component variable value calculation at the time of a case search described later.

Reference numeral 40 in FIG. 1 denotes a search table creation unit, which creates a search table 50 used for similar case search by independent component analysis performed on process variables from the process time series database 30.
The search table 50 is obtained by quantizing measured values of several process variables and independent values of process state values calculated from the measured values as input variables, and values at past points in time. Is used. The quantization of the input variable is performed by converting the input variable value (independent component value) to a value from 1 to N (N is a set value, for example, 25).

  In the case of the blast furnace process, it was obtained by independent component analysis of nine variables such as hot metal temperature, pulverized coal injection amount, solution loss carbon, heat flow ratio, charging pitch, Si amount, hot air temperature, furnace top temperature, Q loss, etc. The data of each of the five independent variables at each time point and the past two times is used as an input variable, and each variable value is quantized to a value from 1 to 25, for example, to create an index part of the search table 50. . As a variable to be searched in the search table 50, the date and time at that time or the storage number of the corresponding database is recorded.

  FIG. 4 shows an example of the search table 50. Here, the input variables are 15 variables (5 independent process variables, 3 times), and the number of quantization is 25. The search variable is a storage number or storage time (DB storage DATE) in the database (DB) 30. Input_1 to input_15 in FIG. 4 are input variables, which are quantized at the time (t), one time before (t-1), and two times before (t-2) regarding independent variables such as hot metal temperature, PCR, and SLC, respectively. It is a thing.

  FIG. 5 shows a procedure for creating the search table 50. In step 41, the process time series database 30 is accessed in time series order, and data values, storage times, or storage numbers of predetermined input variables are collected from the process time series database 30. In step 42, the independent component analysis of the input variable is performed to obtain the independent component variable. In step 43, the data value of the independent variable is quantized.

  If the input data (independent variable) is Xi, the upper and lower limits given in advance are Ximax and Ximin, and the quantization number is n, the quantized value Ixi of the input data Xi is calculated by the following equation (13). Is done.

Ixi = integerization (N * (Xi−Ximin) / (Ximax−Ximin) +1) (13)
However, i is 1 to 15 (i = 1 to 15). In addition, integerization means truncation after the decimal point.

In step 44, 15 quantized independent variable data values and storage times or storage numbers are set in the search table 50.
The process of FIG. 5 is repeated until a predetermined storage number or storage time in the process time series database 30.

  Next, reference numeral 60 in FIG. 1 denotes a similar case search unit, which searches for similar cases using the search table 50. In order to search the process state and the similar time point at the current time, the process data at the current time is fetched from the process measurement / control apparatus 20 and the similar case is searched by the procedure shown in FIG.

  In step 61 of FIG. 6, process data at the current time point for searching for similar cases is obtained from the measurement / control apparatus 20, and independent component variable values are calculated from the previously obtained separation matrix W.

  In step 62, the input data of the search table 50 is selected from the obtained process data (independent component variables), the quantization described above is performed, and 15 input values of the search table 50 are created.

  In step 63, the search table 50 is compared with the selected input data. As an example of a simple collation method, there is a method of collating the search table 50 from top to bottom. If there are items to be verified, the number of stored DB storage numbers or DB storage times (stored DATE) is recorded as many times as the number verified. In this case, the similarity of the case is set to 0.

  On the other hand, if there is no match in the search table 50 with the input data, the vicinity 1 of the input data is searched again. The neighborhood data is first searched up to (+1, −1) before and after the quantized data value of each input data.

  For example, if the hot metal temperature quantization data is 13, the data in the range of 12 to 14 is the collation range. For each input variable, search by expanding the range back and forth to find one that satisfies all, and if there is, record the DB storage number or DB storage time (storage DATE) as a similar time. In this case, the similarity is 1. Similarly, if there is nothing to search, the next neighborhood 2 is searched, and if there is something to be collated, the similar time point is recorded in the same manner, and the similarity is set to 2. A search is made up to the set neighborhood k, and if there is nothing to collate, there is no similar case.

In step 64, similar process data is extracted from the process time series database 30 in order of the DB storage number or DB storage time (storage DATE) at a plurality of similar times searched from the search table, and output to the screen or file. .
Note that reference numeral 70 in FIG. 1 denotes a similar case display unit that displays the search result.

  7 and 8 show search results for similar cases of blast furnaces. The correspondence of hot metal temperature, charging pitch, pulverized coal injection amount, solution loss carbon at the present time and similar time is shown.

  7A and 7B and FIGS. 8A and 8B, the upper row is the current case and the lower row is the similar case. From these results, the trend before and after the current case is similar to that before and after the similar case, indicating that the similar case search functions well.

  In the method according to the present embodiment, it is very easy to update the search table 50. Although it is possible to execute it sequentially online, when a certain amount of update data is accumulated in the process time-series database 30, a value is added to the search table 50 using the data in the procedure of FIG. Is possible. It is also easy to delete the data before a certain point that cannot be used due to operational changes from the search table 50 by deleting the data of the storage date corresponding to the date and time or the point. It can be carried out.

In the method according to the present embodiment, not only an event similar to the current state of the process but also an event similar to a specific time point in the process time series database 30 can be searched.
In addition, a search corresponding to an input state that is not in the search table 50 can be performed by a neighborhood search of the search table 50, although the accuracy is somewhat reduced.

  FIG. 9 is a block diagram showing the configuration of an apparatus used for carrying out the process future state prediction method of the present invention. In the same manner as in FIG. 1, the search table 50 is searched for past cases similar to the time point at which the future state of the process is to be predicted, the similar time is searched, and the future state is predicted using the future case. To do. The procedure will be described with reference to FIG.

  In step 71 of FIG. 10, process data at the present time is obtained from the measurement / control apparatus 20. In step 72, the independent component variable is obtained. In step 73, the input data of the search table 50 is selected from the obtained process data, quantized, and the input data of the search table 50 is created. These steps 71 to 73 correspond to steps 61 and 62 in FIG.

  In step 74, the input data is calculated and set in the search table 50 to obtain a plurality of stored similar time points. Let the similarity at this time be IR. This step 74 corresponds to step 63 in FIG.

  In step 75, the process data is sequentially obtained from the process time-series database 30 using the future prediction time set from the similar time point, for example, the DB storage time up to 8 hours ahead as a key. In addition, it is possible to obtain a plurality of sets of future state prediction values by repeating the same operation at a plurality of similar points. Individual future prediction graphs at a plurality of similar points, averaged graphs thereof, and fluctuation ranges are displayed on the screen. In addition, the reliability corresponding to the similarity of the case is displayed together with the display.

  For example, when the similarity is 0, the reliability is high (H), when the similarity is 1, the reliability is slightly high (SH), and when the similarity is 2, the reliability is normal (M). When the similarity is 3, the reliability is displayed as caution (L). In FIG. 9, reference numeral 90 denotes a future state prediction display unit that displays the prediction result of those future states.

FIG. 11 shows the future prediction results of the blast furnace. The actual value and predicted value of the hot metal temperature two hours ahead correspond well, indicating the effectiveness of this method.
The method according to the present embodiment can also be applied to process variables that can be imaged corresponding to spatial positions such as temperature distribution and pressure distribution. In this case, the feature amount is obtained by independent component analysis, and the variable is used as an input variable, whereby similar case retrieval and future state prediction can be performed.

  FIG. 12 shows a case where a similar case search is performed on the distribution image in the furnace circumferential direction and height direction of the blast furnace. It can be seen that images similar to the original image (FIG. 12A) (FIGS. 12B to 12D) can be searched with high accuracy at a plurality of times.

(Other embodiments of the present invention)
The control operation by the apparatus in each embodiment described above can be realized by using a computer system as shown in FIG.
FIG. 13 is a block diagram showing an example of the configuration of a computer system arranged in the apparatus according to the above-described embodiment.
In FIG. 13, a computer system 100 includes a CPU 101, a ROM 102, a RAM 103, a keyboard controller (KBC) 105 of a keyboard (KB) 104, a CRT controller (CRTC) 107 of a CRT display (CRT) 106 as a display unit, and The disk controller (DKC) 110 of the hard disk (HD) 108 and the flexible disk (FD) 109 and the network interface controller (NIC) 112 for connection to the network 111 can communicate with each other via the system bus 113. Connected configuration.

The CPU 101 comprehensively controls each component connected to the system bus 103 by executing software stored in the ROM 102 or the HD 108 or software supplied from the FD 109.
That is, the CPU 101 performs a control for realizing an operation to be described later by reading a processing program according to a predetermined processing sequence from the ROM 102, the HD 108, or the FD 109 and executing it.

The RAM 103 functions as a main memory or a work area for the CPU 101.
The KBC 105 controls an instruction input from the KB 104 or a pointing device (not shown).

The CRTC 107 controls display on the CRT 106.
The DKC 110 controls access to the HD 108 and the FD 109 that store a boot program, various applications, an editing file, a user file, a network management program, a predetermined processing program in the present embodiment, and the like.
The NIC 112 bidirectionally exchanges data with devices or systems on the network 111.

  Also, a software program for realizing the functions of the above-described embodiment for an apparatus connected to the various devices or a computer in the system so as to operate various devices to realize the functions of the above-described embodiments. What was implemented by supplying the code and operating the various devices in accordance with a program stored in a computer (CPU or MPU) of the system or apparatus is also included in the scope of the present invention.

  In this case, the program code of the software itself realizes the functions of the above-described embodiments, and the program code itself and means for supplying the program code to the computer, for example, the program code are stored. This storage medium constitutes the present invention. As a storage medium for storing the program code, for example, a flexible disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a magnetic tape, a nonvolatile memory card, a ROM, or the like can be used.

  Further, by executing the program code supplied by the computer, not only the functions of the above-described embodiments are realized, but also the OS (operating system) or other application software in which the program code is running on the computer, etc. It goes without saying that the program code is also included in the embodiment of the present invention even when the functions of the above-described embodiment are realized in cooperation with the embodiment.

  Further, after the supplied program code is stored in the memory provided in the function expansion board of the computer or the function expansion unit connected to the computer, the CPU provided in the function expansion board or function expansion unit based on the instruction of the program code Needless to say, the present invention includes a case where the functions of the above-described embodiment are realized by performing part or all of the actual processing.

It is a block diagram which shows embodiment of this invention and shows the structure of the apparatus used for implementation of the similarity search method of a process. It is a figure which shows embodiment of this invention and shows the data storage format of the process time series database which stored the actual value of the process. It is a flowchart which shows embodiment of this invention and demonstrates the procedure at the time of performing the independent component calculation of a process variable. It is a figure which shows embodiment of this invention and demonstrates the table for a search. It is a flowchart which shows embodiment of this invention and demonstrates the preparation procedure of the table for a search. It is a flowchart which shows embodiment of this invention and demonstrates the search procedure of the similar case in a similar search part. It is a figure which shows embodiment of this invention and shows an example of the result of having searched similar cases. It is a figure which shows embodiment of this invention and shows the other example of the result of having searched similar cases. It is a block diagram which shows embodiment of this invention and shows the structure of the apparatus used for implementation of the future state prediction method of a process. It is a flowchart which shows embodiment of this invention and demonstrates the procedure at the time of predicting a future state. It is a figure which shows embodiment of this invention and shows the result of having predicted the future state. It is the figure which showed embodiment of this invention and showed the example which performed the similarity search by the independent component variable of a process image. It is the block diagram which showed other embodiment of this invention and showed an example of the structure of the computer system arrange | positioned at the apparatus.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Blast furnace process 20 Measurement / control apparatus 30 Process time series database 35 Independent component analysis part 40 Search table preparation part 50 Search table 55 Independent component calculation part 60 Similar case search part 70 Similar result display part 90 Future state prediction display part

Claims (6)

Using a computer, process data consisting of measured values for each of a plurality of process variables is stored in a time series database that is stored at fixed time intervals along with the storage time and storage number indicating the storage order. In a method for searching past state similar cases of processes similar to
(B) for each of the plurality of process variables, a step of a series of process data from the specified time Tk until only time in the past predetermined number m, and out the time series from the database extract,
(B) Independent component analysis is performed on the extracted series of process data, and a plurality of independent variable values s composed of a series of data values at a predetermined time point m from the designated time Tk ; Calculating a separation matrix W ;
(C) The calculated series of a plurality of independent variable values s are used as input data, or the calculated series of the plurality of independent variable values s and the process variable selected in advance among the plurality of process variables. A series of process data is used as input data, and data values Xi that are values at the time Tk and past times from the time Tk of the input data are normalized using upper and lower limit values Ximax and Ximin. Next , calculating a data value at time Tk in the search table by calculating a quantized input value Ixi by multiplying by a predetermined quantization number and then converting to an integer .
And storing (d) combined at least any one of the stored numbers of the time series database corresponding to the specified time Tk, the data value at time Tk of the retrieval table in the search table,
A process of repeating the above-described series of steps while changing the designated time Tk to create a search table;
For each of the plurality of process variables, comprising the steps of obtaining a series of process data from the time Tk and another designated specific time Ts to a past time point by the predetermined number m,
A series of process data from a specific time Ts of the extracted, by multiplying the separation matrix W calculated in step (b), calculating a German Tatsuhen numbers in a particular time Ts,
With respect to the calculated independent variable value at the specific time Ts, the step (c) is processed to calculate input data for search at the specific time Ts;
Using the input data for search at the specific time Ts as a search key, the input data value and the data value at each past time Tk in the search table are collated for each corresponding data value, A similarity that is a difference between data values is calculated, and the similarity is used to search for a similar data value up to the vicinity of a preset value of the similarity as a similar case, and is associated with the data value of the past time was stored time or a step of determining the store number,
A step of extracting said engraved when stored in similar cases correspond to store numbers, a series of process data similar case from the time series data base,
State similar case retrieval method of the process which is characterized in that it consists of. Using a computer, process data consisting of measured values for each of a plurality of process variables is stored in a time series database that is stored at fixed time intervals along with the storage time and storage number indicating the storage order. In a method for predicting the future state of a process by searching for similar cases in the past state of a process similar to
(B) for each of the plurality of process variables, a step of a series of process data from the specified time Tk until only time in the past predetermined number m, and out the time series from the database extract,
(B) Independent component analysis is performed on the extracted series of process data, and a plurality of independent variable values s composed of a series of data values at a predetermined time point m from the designated time Tk ; Calculating a separation matrix W ;
(C) The calculated series of a plurality of independent variable values s are used as input data, or the calculated series of the plurality of independent variable values s and the process variable selected in advance among the plurality of process variables. A series of process data is used as input data, and the data value Xi, which is the value of the input data at the time Tk and the past time from the time Tk, is normalized using upper and lower limit values Ximax and Ximin. Next , calculating a data value at time Tk in the search table by calculating a quantized input value Ixi by multiplying by a predetermined quantization number and then converting to an integer .
And storing (d) combined at least any one of the stored numbers of the time series database corresponding to the specified time Tk, the data value at time Tk of the retrieval table in the search table,
A process of repeating the above-described series of steps while changing the designated time Tk to create a search table ;
For each of the plurality of process variables, comprising the steps of obtaining a series of process data from the time Tk and another designated specific time Ts to a past time point by the predetermined number m,
Multiplying the series of process data from the extracted specific time Ts by the separation matrix W calculated in the step (b) to calculate an independent variable value at the specific time Ts ;
With respect to the calculated independent variable value at the specific time Ts, the step (c) is processed to calculate input data for search at the specific time Ts;
Using the input data for search at the specific time Ts as a search key, the input data value and the data value at each past time Tk in the search table are collated for each corresponding data value, A similarity that is a difference between data values is calculated, and the similarity is used to search for a similar data value up to the vicinity of a preset value of the similarity as a similar case, and is associated with the data value of the past time was stored time or a step of determining the store number,
Extracting process data from the storage time of the similar case to a specified previous time, or a series of process data from the storage number of the similar case to the specified previous storage number from the time series database ; , And
A process state prediction method, characterized in that the series of extracted process data is set as a future state of the process.   The time series database targets a blast furnace process, and the process variable value is at least 1 from hot metal temperature, pulverized coal injection amount, solution loss carbon, heat flow ratio, charging pitch, Si amount, hot air temperature, furnace top temperature, Q loss. The process state similarity case search method according to claim 1, wherein one or more are selected.   The time series database targets a blast furnace process, and the process variable value is at least 1 from hot metal temperature, pulverized coal injection amount, solution loss carbon, heat flow ratio, charging pitch, Si amount, hot air temperature, furnace top temperature, Q loss. The process state prediction method according to claim 2, wherein one or more are selected. In the process state prediction method according to claim 2 or 4,
The future state of the process is displayed, and at the same time, the reliability corresponding to the similarity of the similar cases used to obtain the series of the extracted process data is displayed as the certainty of the prediction. state prediction method of to Help process. Each step of the state similar case retrieval method of the process according to claim 1 or claim 3, or claim 2, the state prediction of the process claimed in any one of claims 4 and 5 computer-readable storage medium characterized by the program remembers for executing the steps of the method on a computer.

Images