JP2005133115A - Method, apparatus and computer program for monitoring operation of blast furnace - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To monitor an operation state of a blast furnace, estimate the operation state, and determine an operation condition. <P>SOLUTION: This monitoring method comprises computerizing information on a spatially distributed state and changing state with time of various measurement data on a pressure or temperature of a furnace body of the blast furnace, as an image; consecutively calculating an independent component and an isolation matrix which are characteristic information of the image information, through an independent component analysis; defining a similarity index on the basis of the calculated independent component and consecutively searching a past operation state similar to the present operation state; confirming the course of the past operation state and the past operation condition when the case was operated, from various measurement data in the past similar cases and an operation diary record; and consecutively grasping the present operation condition of the blast furnace and predicting a future course of the operation condition. The method further comprises determining the future operation condition from those informations. In addition, the method comprises consecutively watching a time-series change of a calculated independent component, and precisely monitoring the abnormality of the operation state of the blast furnace, such as poor air permeability or a temperature drop in the furnace body. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a method, an apparatus, and a computer program for monitoring the operating state of a blast furnace.For example, during operation of a blast furnace, the spatial distribution and temporal change of measurement data such as pressure and temperature of the furnace body are converted into image information, By performing independent component analysis of the measurement data converted into image information, the independent component and the separation matrix, which are feature information of the image information, are sequentially calculated, and a similarity index is defined based on the calculated independent component, Search past operating conditions similar to the operating conditions, check the transition of blast furnace operating conditions and past operating conditions in past similar operating cases, and predict the transition of subsequent operating conditions to determine the operating conditions In addition, the present invention relates to a system that sequentially monitors the time series transition of the calculated independent component.

  Conventionally, there are those disclosed in Patent Literature 1, Patent Literature 2, and the like as methods related to monitoring and prediction of abnormal operation of a blast furnace. Each of these monitoring arrangement prediction methods collects measurement data from each sensor without reflecting the installation position information of each sensor on the blast furnace facility, and uses a preset set value or a simple physical model. The operation status is monitored and abnormal operation is predicted by comparison with the threshold value.

JP-A-5-156328 Japanese Patent Laid-Open No. 11-140520 "Fast and Robust Fixed-Point Algorithms for Independent Component Analysis" (Aapo Hyvarian, IEEE Transactions on Neural Networks.Vol.10, No.3, May 1999, P.626 / 634) Independent Component Analysis (2001, John Wiley & Sons. Inc.)

  However, the process called the blast furnace targeted by the present invention is an object to be handled as a distributed constant system process having a temporal change (dynamic characteristics). Therefore, the measurement data of various sensors installed with spatial distribution on the blast furnace equipment may not be collected and evaluated independently of each other, but installed on the blast furnace equipment where each sensor is attached. It should be collected and evaluated in relation to the location.

  In the conventional method, the installation positions of such sensors are not collected and evaluated in association with measurement data, and as a result, there is a problem that the operation state monitoring and prediction accuracy of the blast furnace is low.

  In addition, with the conventional method, it is difficult to quantitatively reveal the characteristic information of the blast furnace operation that is latent in the measurement data as a specific numerical value. Search the operation state similar to the operation state specified in advance, check the transition of the blast furnace operation state and past operation conditions in the past similar operation examples, predict the transition of the subsequent operation state, and operate There was a problem that the conditions could not be determined.

  The present invention has been made in view of such circumstances, and for the purpose of solving the above-described problems and monitoring and predicting the operating state of the blast furnace and determining the operating conditions, the pressure and temperature of the furnace body, etc. Spatial distribution information and temporal change information of measurement data is converted into image information, and independent components and separation matrices that are characteristic information of the image information are sequentially calculated by independent component analysis, and a similarity index is calculated based on the calculated independent components. The past operation state similar to the current operation state is defined and searched, and the time series transition of the calculated independent component is sequentially monitored.

  The operation monitoring method in the operation of the blast furnace according to the present invention is a two-dimensional plane in which measurement data of a measurement target amount from a plurality of sensors installed in the blast furnace is reflected on the installation position of each sensor or a three-dimensional plane. It is arranged on the surface of the solid, and the isoline on the surface is calculated for the spatial distribution state and temporal change of the measurement data, and the figure formed by the isoline or the feature information of the figure is evaluated as image information. A method for monitoring the operating state of a blast furnace, wherein an independent component analysis is performed on the image information to calculate an independent component and a separation matrix, a similarity index is defined using the calculation result, and accumulated image information This is characterized in that image information similar to image information designated in advance is searched.

  The operation monitoring method in another blast furnace operation of the present invention is constituted by a two-dimensional plane or a two-dimensional plane in which measurement data of a measurement target amount from a plurality of sensors installed in the blast furnace is reflected on the installation position of each sensor. It is arranged on the surface of a three-dimensional solid, the isoline on the surface is calculated for the spatial distribution state and temporal change of the measurement data, and the graphic or graphic feature information formed by the isoline is used as image information. A method of monitoring an operating state of a blast furnace to be evaluated, wherein an independent component analysis is performed on the image information to calculate an independent component and a separation matrix, and a similarity index is defined and stored using the calculation result. It is characterized in that image information similar to image information designated in advance is searched from image information, and the order of the similarity is evaluated based on the magnitude of the similarity index.

  The operation monitoring method in another blast furnace operation of the present invention is constituted by a two-dimensional plane or a two-dimensional plane in which measurement data of a measurement target amount from a plurality of sensors installed in the blast furnace is reflected on the installation position of each sensor. It is arranged on the surface of a three-dimensional solid, the isoline on the surface is calculated for the spatial distribution state and temporal change of the measurement data, and the graphic or graphic feature information formed by the isoline is used as image information. A method of monitoring an operating state of a blast furnace to be evaluated, wherein an independent component analysis is performed on the image information to calculate an independent component and a separation matrix, and a similarity index is defined and stored using the calculation result. It is characterized in that a time in a similar operation state is retrieved by retrieving image information similar to image information at a predesignated time from the image information.

  Another feature of the operation monitoring method according to another blast furnace operation of the present invention is that the measurement data of the measurement target amount from a plurality of sensors installed in the blast furnace is a two-dimensional plane reflecting the installation position of each sensor or A figure or figure that is placed on the surface of a three-dimensional solid composed of two-dimensional planes, calculates the isoline on the surface for the spatial distribution state or temporal change of measurement data, and forms the isoline The feature information of the blast furnace is evaluated as image information, and the operation state of the blast furnace is monitored. The image information is subjected to independent component analysis to calculate an independent component and a separation matrix, and the similarity is calculated using the calculation result. By defining an index and searching image information similar to the image information at a predesignated time from the stored image information, the time of a past similar operation state is searched, and various measurement data and operations at the time Diary From recorded, to check a transition or operating conditions executed to the time of the operation state of the time, characterized in that it determines a prediction to operational conditions the transition of the operational state at the time after which the above specified.

  The operation monitoring method in another blast furnace operation of the present invention is constituted by a two-dimensional plane or a two-dimensional plane in which measurement data of a measurement target amount from a plurality of sensors installed in the blast furnace is reflected on the installation position of each sensor. It is arranged on the surface of a three-dimensional solid, the isoline on the surface is calculated for the spatial distribution state and temporal change of the measurement data, and the graphic or graphic feature information formed by the isoline is used as image information. It is a method for monitoring the operating state of the blast furnace to be evaluated, and for the image information, an independent component analysis is performed to calculate an independent component and a separation matrix, and a time series transition of the independent component is set in advance. Compared with large and small, it is characterized in that the operating state of the blast furnace is monitored.

  The operation monitoring apparatus for blast furnace operation according to the present invention is a data collection means for collecting measurement data measured by various sensors installed on a blast furnace facility in the operation monitoring apparatus for outputting and monitoring the operation status of the blast furnace. And the spatial distribution state and temporal change of the collected measurement data are arranged on the surface of a three-dimensional solid composed of a two-dimensional plane or a two-dimensional plane reflecting the installation position of each sensor on the blast furnace facility, Image information converting means for calculating arbitrary isolines having the same measurement data, independent component analyzing means for performing independent component analysis of the image information calculated by the image information processing means, and independent component analysis of the image information Similar operation case search means for searching similar operation cases of blast furnace from results and / or independent component / separation matrix time series evaluation means for evaluating independent components and separation matrix in time series, and said image information means Independent component analysis means, similar operation case search means and / or independent component / separation matrix time series evaluation means, a recording means for recording an operation diary describing the measurement data and operation conditions, and the image information conversion means, independent It has a feature in that it has an output means for outputting the calculation results in the component analysis means, the similar operation case search means, and the independent component / separation matrix time series evaluation means.

  The computer program of the present invention is characterized in that it causes a computer to execute the processing procedure of any one of the operation monitoring methods of the present invention. Further, the present invention is characterized in that a computer functions as each means of the operation monitoring apparatus of the present invention.

  The method of the present invention is a method for measuring data measured by various sensors installed on a blast furnace facility on a two-dimensional plane reflecting a setting position of each sensor or a three-dimensional surface constituted by a two-dimensional plane. When the spatial distribution state of each data is represented as image information formed by these, the independent component and the separation matrix of the image information are calculated to define the similarity index, and the stored image information group By sequentially searching for images similar to the specified image, cases with highly similar images are sequentially identified, and from various measurement data and operation diary records in the cases, By confirming the operating conditions executed at the time of the case, it is possible to predict the current operating status of the blast furnace and the future operating status. Furthermore, it will be possible to determine future operating conditions based on these information. In addition, it is possible to accurately monitor the operating state of the blast furnace by sequentially monitoring the time series transition of the independent components to be sequentially calculated.

  DESCRIPTION OF EMBODIMENTS Hereinafter, an embodiment of an operation monitoring method, apparatus, and computer program in blast furnace operation of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of an operation monitoring apparatus in the blast furnace operation of the present embodiment.

(1. Blast furnace equipment and 2. Various sensors on the blast furnace equipment)
In the various sensors 2 on the blast furnace facility 1, physical quantities such as temperature or pressure, flow rate, flow velocity, particle size, density, composition, and the like are measured. FIG. 1 shows the position of a plurality of sensors installed on the outer surface of a blast furnace facility, taking a stave temperature sensor, a hearth wall temperature sensor, and a shaft pressure sensor as examples. In the present invention, the arrangement of the sensors on the blast furnace equipment may be unevenly spaced. Of various types of physical quantity measurement sensors, a pressure sensor will be described as an example.

(3. Data collection device)
In the data collection device 3, measurement data output from a plurality of pressure sensors arranged on the blast furnace facility is sampled and collected at a preset sampling period Δt. The sampling period Δt can be arbitrarily set at a time interval of several ms or more corresponding to the processing capability of the data collection device 3 and the processing capability of the data processing device 4 and the time interval required for operation monitoring and operation prediction. The pressure data collected by the data collection device 3 is sent to the data processing device 4 in real time.

  In this case, in the present invention, it is not necessary to limit the transmission form and transmission method of the measurement data to the data processing device 4. For example, the measurement data is transmitted to the data processing device 4 as an analog signal of current or voltage by a signal line. In addition, the data collection device 3 may convert the analog signal of the current or voltage into a digital signal, and transmit the digital signal to the data processing device 4. Further, the digital signal may be converted into various data compression methods. The transmission data amount may be reduced by compressing the data to be transmitted to the data processing device 4, and after transmission, the compressed transmission data may be restored to the digital signal before compression in the data processing device 4. The digital signal is separated from the blast furnace equipment via an information transmission network such as LAN, Ethernet (R), wireless LAN, and the Internet. May be transmitted placed the to the data processing apparatus 4 to the remote location was.

(4. Data processing device)
(5. Image information processing unit)
In the image information processing unit 5, the pressure data input from the data collection device 3 is converted into a two-dimensional plane or a three-dimensional solid composed of a two-dimensional plane reflecting each pressure sensor installation position information on the blast furnace equipment. Arbitrary isolines which are arranged on the surface and have the same pressure data are calculated, and a figure formed by the isolines is calculated and converted into image information.

  FIG. 2 illustrates the relationship between the coordinate system set in the blast furnace and a plurality of pressure sensor installation positions installed in the blast furnace equipment. In Fig. 2, the h-axis is defined in the blast furnace height direction, the r-axis is defined in the furnace radial direction, the θ-axis is defined in the furnace circumferential angle direction, and the ● marks indicate the installation positions of a plurality of pressure sensors arranged on the blast furnace equipment. Are arranged according to each coordinate axis. The installation position coordinates of each pressure sensor are known in advance.

  At this time, the measurement data at time t of the pressure sensor installed at the coordinates (h (i), r (i), θ (i)) is P (h (i), r (i), θ (i), It shall be expressed in t).

  Here, the pressure data input from the data collection device 3 is arranged on the surface of a three-dimensional solid composed of a plurality of rectangular planar elements reflecting the installation position information of each pressure sensor on the blast furnace equipment, Shows a method of calculating isoline of pressure data and converting it to image information.

  ● Based on the pressure data placed at the point marked with ●, the pressure data in the mutual space marked with ● is spatially interpolated to search for isolines. Here, the isoline is obtained by connecting points showing the same value from the spatially distributed pressure data with a line. At this time, the mutual intervals between the marks may be spatially unequal by an isoline search method described later, and need not be spatially equal.

  In order to search for isolines for spatially distributed pressure data, the method of using a triangular element composed of pressure sensor installation points is reliable. Has enormous freedom. Further, there arises a problem that the shape of the obtained isoline differs depending on the selection of the triangular element.

  Therefore, as a technique to reduce the isoline shape error due to element selection while lowering the degree of freedom of element selection, the "equivalence using a triangle element using a quadrilateral plane element four vertex average as a vertex" “Line Search Method” is exemplified.

  The isoline search method will be described with reference to FIG. The pressure sensor installation position on the blast furnace equipment in FIG. 2 is associated with each point in advance so that one of the inner angles is constituted by a rectangular planar element that does not exceed 180 degrees. The element selection condition for the quadrangular planar element reduces the degree of freedom in element selection and facilitates element selection. In the case of a blast furnace facility, each pressure sensor position coordinate is known, and therefore, it is only necessary to make an association once.

  The blast furnace equipment has a three-dimensional shape in which conical shapes having a shaft angle, a Bosch angle (morning glory angle), etc. are partially combined, but the outer shape of the blast furnace equipment is a rectangular plane element whose one of the inner angles does not exceed 180 degrees. It can be expressed as a three-dimensionally combined shape, and the method described below can be implemented.

  FIG. 3 is a diagram showing that a shaft portion of a blast furnace can be expressed by a shape in which a rectangular plane element whose outer surface does not exceed 180 degrees is three-dimensionally combined.

  In order to make it easy to understand the meaning of the figure in the example of the shaft portion in FIG. 3, the points a and d are arranged with respect to the arrangement of the four vertices a, b, c and d of the rectangular plane element whose one of the inner angles does not exceed 180 degrees. An example in which the points b and c are arranged on the same furnace cross-section circle and the same furnace cross-section circle is shown. Pressure data measured by the pressure sensors located at the four vertices are P (h1, r1, θ1, t), P (h2, r2, θ1, t), P (h2, r2, θ2, t), P (h1, r1, θ2, t) In the case of the arrangement of the four vertices, any other special condition is not required as long as the four vertices constitute a rectangular plane element.

  Therefore, hereinafter, Pa = P (h1, r1, θ1, t), Pb = P (h2, r2, θ1, t), Pc = P (h2, r2, θ2, t), Pd = P (h1, r1 , θ2, t), and the explanation will be continued.

Here, the intersection of the diagonal lines of this quadrilateral plane element, that is, the pressure at the point “e” in FIG. Pe is an average value calculated from Pa, Pb, Pc, and Pd, and is defined as an arithmetic average, for example.
Pe = (Pa + Pb + Pc + Pd) ÷ 4 ・ ・ ・ Equation (1)

  Next, four triangular elements having apexes at the intersections o marked points e on the diagonal are defined inside the quadrangular plane elements. At this time, the pressure data on the side of each triangular element shall be obtained by spatially interpolating with the pressure data of the vertices constituting both ends of the side. In the interpolation, any method such as a primary interpolation method may be used.

As shown in FIG. 3, suppose that the value of the isoline to be searched now is P1,
Pa <P1 <Pd (2)
With this relationship, it is assumed that the isoline of P1 has entered the triangular element ade constituting the rectangular planar element abcd so as to cross the side ad. At this time, P1 exists as a pressure data point spatially interpolated on the side ad according to the condition of the expression (2), and is indicated by Δ on the side ad in FIG.

Next, the magnitude relationship between the pressure Pe at the remaining vertex e of P1 and the triangular element ade is compared.
Here, tentatively
Pa <P1 <Pe (3)
, P1 exists as a pressure data point spatially interpolated on the side ae, and is indicated by a Δ mark on the side ae in FIG.

Subsequently, it is assumed that the contour line P1 to be searched enters the triangle element abe so as to cross the side ae, and the magnitude relationship between the pressure Pb of P1 and the remaining vertex b of the triangle element abe is compared. Here, tentatively
Pa <P1 <Pb (4)
, P1 exists as a pressure data point spatially interpolated on the side ab, and is indicated by a Δ mark on the side ab in FIG.

  By connecting Δ marks on the side ad, the side ae, and the side ab obtained as described above with a straight line, it is possible to search for an isoline of P1 in the rectangular planar element abcd.

In the above example, instead of equation (4)
Pb <P1 <Pe (5)
Pb <P1 <Pc (6)
As an example, the pressure data points at this time are as shown by ◇ in FIG. 3, and by connecting them with straight lines, the isoline of P1 in the rectangular planar element abcd is shown with a broken line be able to.

In the above example, instead of formulas (3) and (4)
Pd <P1 <Pe (7)
Pd <P1 <Pc (8)
As an example, the pressure data points at this time are as indicated by □ in FIG. 3, and by connecting them with straight lines, the isoline of P1 in the rectangular planar element abcd is indicated by a broken line. be able to.

  As described above, the description has been given so far regardless of the magnitude relationship between the pressure data values Pa, Pb, Pc, and Pd located at the four vertices of the rectangular planar element and the isoline P1 to be searched. By the “isoline search method using a triangular element using a quadrilateral plane element four-vertex average as a vertex”, an isoline of P1 in the quadrangular plane element abcd can be searched as shown in FIG.

  Similarly, with respect to a rectangular plane element adjacent to the rectangular plane element abcd in FIG. 3 whose one of the inner angles does not exceed 180 degrees, four triangular elements whose vertexes are the intersections of the diagonal lines are defined, and equality is obtained by the same method. By repeating the line search, it is possible to continuously search for the isoline of P1 in the adjacent region of the rectangular planar element abcd in FIG.

  Further, by repeating the above process for all the rectangular plane elements set in advance on the blast furnace equipment, the search and drawing of the isoline on the blast furnace equipment is completed, and the pressure data is obtained from the blast furnace by the obtained isolines. A certain figure (hereinafter referred to as a contour figure) is formed on the facility. FIG. 4 shows an example in which the isolines of the pressure values P1, P2, and P3 in the vicinity of the point a on the blast furnace facility are indicated by solid lines, and the contour figure surrounded by the isoline of P1 is indicated by hatching.

  As described above, for the pressure data measured by a plurality of pressure sensors spatially arranged on the blast furnace equipment, a rectangular plane element whose one of the inner angles does not exceed 180 degrees is selected, and four vertices are formed at the intersections of the diagonal lines. Compared to the method of searching for isolines using only triangular elements, the method of searching for and drawing isolines using triangular elements with vertices at the intersection points It reduces the degree of freedom of selection and facilitates selection, and uses a triangular element whose vertex is the average value of each vertex of a quadrilateral plane element, so it is an effective method that can reduce isoline search errors that depend on element selection. is there. Since the triangular element is used in the final stage of the search, it goes without saying that there is no problem that the isoline to be searched intersects with another isoline on the way or the isoline is interrupted on the way. Yes.

  In addition, here, the surface of a three-dimensional solid composed of two-dimensional planes has been described as an example. However, the isoline search method can be implemented on a two-dimensional plane in which the surface of a three-dimensional solid is developed and is effective. It goes without saying that this is a technique. For example, FIG. 5 is an example in which the isoline of the shaft pressure searched by the above method is mapped to the two-dimensional development plane in the furnace height direction and the furnace circumferential direction. FIG. 5A shows a case where the shaft pressure distribution is good, and FIG. 5B shows a case where the shaft pressure distribution is bad. In the present invention, there is no need to limit the isoline search method, and the isoline search may be performed using other methods.

  As described above, in the image information processing unit 5, the pressure data input from the data collection device 3 is a three-dimensional surface composed of quadrilateral planar elements reflecting pressure sensor installation position information on the blast furnace equipment. And can draw isolines.

  Here, the shaft pressure data has been described as an example. However, in the present invention, measurement data of other physical quantities such as temperature and flow velocity, and temporal change (time change rate) data of the measurement data are used. An isoline can be drawn by arranging the surface of a three-dimensional solid composed of quadrilateral plane elements reflecting the above-described various sensor installation position information or the developed two-dimensional plane of the surface of the three-dimensional solid.

(6. Independent component analysis unit)
Next, the independent component analysis unit 6 performs independent component analysis of the measurement data converted into the image information by the image information processing unit 5. Independent component analysis (hereinafter abbreviated as ICA) means that the observed signal consists of a linear sum of several statistically independent original signals, and both the original signal and its mixed state are unknown. This is a method for estimating the original signal.

  That is, the measurement data converted into the image information by the image information processing unit 5 is configured by linearly mixing the original signals of several statistically independent image information components reflecting the state in the furnace. Assuming that, in monitoring the operating state of the blast furnace, the present invention calculates the original signal of the statistically independent image information component of the measurement data converted into image information, that is, the independent component of the image information, and defines the similarity index. Thus, the similarity is quantitatively evaluated, and past similar operation cases are efficiently searched. In addition, the time series transition of the independent component is monitored.

  Here, the ICA algorithm will be described. Several algorithms have been proposed for the ICA algorithm depending on the difference in evaluation function for calculating the independent component and the method of selecting the convergence calculation method. Here, the algorithm called FastICA proposed by Aapo Hyvarian et al. In a paper (Non-Patent Document 1) and Aapo Hyvarien, Juha Karhunen, and Erikki Oja in a book (Non-Patent Document 2) is taken as an example. explain. In each equation shown below, each variable is indicated by an uppercase bold typeface for a matrix, a lowercase bold typeface for a vector, and a slant itself for a scalar.

First, the problems that ICA should solve are summarized. Assume that an m-dimensional input signal observed at a certain time t is a vector x _(t) , and the vector x _(t) is a statistically independent and linear combination of n-dimensional original signal vectors s _(t) . That is, it is expressed by the following formulas (9) to (11). Here, ^T represents transposition. The matrix A is a real matrix of m rows × n columns and is called a mixing matrix.

At this time, the ICA quantitatively calculates and evaluates the non-Gaussianity (non-normality ₎ of the observed signal vector x _(t) without having any knowledge about the original signal vector s _(t) and the mixing matrix A. Thus, the mixing matrix A and the vector y _(t) having n statistically independent components are simultaneously estimated. Hereinafter, the vector y _(t) is referred to as a restored signal vector.

At this time, if n ≦ m, a solution exists. That is, there exists a real matrix W of some n rows × m columns, and n-dimensional restored signal vectors y _(t) that are statistically independent from each other can be reconstructed by the following equation (12). At this time, the real matrix W is called a separation matrix.

In the case of the following expression (13) (I is a unit matrix of n rows × n columns), the restored signal vector y _(t) matches the original signal vector s _(t) .

However, the independence is maintained even if the order of the components of the restored signal vector y _(t) is changed, and the magnitude of each component does not affect the independence. Should be satisfied. In this case, P is called a permutation matrix in which the order of each component is replaced by a matrix of n rows × n columns having only one in each column and each row, and D is an n row × n column that determines the size of each component. A diagonal matrix is called a scaling matrix.

That is, the problem to be solved by the ICA is to obtain a statistically independent restoration signal vector y _(t) and a separation matrix W while allowing two arbitrary characteristics of the order and size of the components.

Next, the ICA solution will be explained in specific steps. First, preprocessing called centering and whitening is performed on the observed signal vector x _(t) .

The centering means that the average of the observed signal vectors is zeroed by subtracting the average value vector ((15) below ₎ of the observed signal vectors x _(t) . That is, if the observation signal vector after centering is x ^ _(t) , the following equations (16) and (17) are obtained.

This means that the original signal vector s _(t) is also zero-averaged by taking the expected values of both sides of equation (11) and subtracting from both sides. That is, if the average value vector of the original signal vector s _(t) is expressed by the following expression (18) and the original signal vector after centering is _expressed by s ^ _(t) , expression (11) can be expressed by the following expressions (19), 20).

  At this time, the following equation (21) is obtained. Even if equation (20) is used instead of equation (11), generality is not lost.

  Whitening is a process of normalizing the variance of signals and decorrelating them. The FastICA algorithm uses principal component analysis as follows.

The eigenvalue decomposition of the covariance matrix of the centralized observation signal vector x ^ _(t) (the following expression (22)) is expressed by the following expression (23), and the observation signal vector x ^ _(t) is expressed by the following expression (24): Converted.

Here, D is a diagonal matrix (the following equation (25)) in which the eigenvalues of the covariance matrix Σ _{x ^} are diagonal components in descending order, and E is an orthogonal matrix of the eigenvector of the following equation (26). z _(t) is a whitened observation signal and is represented by the following equation (27).

At this time, substituting equation (20) into equation (24) yields the following equation (28), and when substituting the following equation (29), equation (28) becomes the following equation (30), which is centralized and whitened: It can be seen that the observed signal vector z _(t) is a linear combination of the mixing matrix A˜ obtained by transforming the centered original signal vector s _(t) by the equation (29). Even if the equation (30) is used instead of the equation (11) or the equation (20), the generality is not lost.

At this time, the expression for obtaining the restored signal vector y _(t) after whitening is expressed by the following expression (31) using the separation matrix W.

  Although centering and whitening are not necessarily pre-processing required for ICA, convergence is achieved by reducing the size of the observed signal vector and reducing the amount of calculation, and limiting the space for ICA optimization to orthogonal space. Can be improved.

Distribution of independent random variables y _{1 (t)} and y _{2 (t)} with arbitrary identical distribution characteristics y _{1 (t)} + y _{2 (t)} by the Center Limit Theorem in probability theory The characteristics are closer to the Gaussian distribution than the distribution characteristics of the original two variables y _{1 (t)} and y _{2 (t)} , and the sum y _{1 (t)} + y _{1 (t)} +... + Y _{N (t)} Is guaranteed to follow the Gaussian distribution rule as the number N increases.

In other words, the farther the distribution of the restored signal vector y _(t) is from the Gaussian distribution, that is, the greater the non-Gaussianity, the closer the restored signal y _(t) becomes. Non-Gaussianity of the restored signal vector y _(t ) is evaluated by kurtosis, which is a fourth-order statistic of each component y _{i (t)} of y _(t) , as shown in the following equation (32). it can.

Therefore, if a separation matrix W that maximizes or minimizes the kurtosis is obtained, the components of the restored signal vector y _(t) are independent of each other. Note that when the observation signal vector is whitened and optimization in an orthogonal space is considered, the second term on the right side of Equation (32) is a constant term.

Accordingly, the independence evaluation function of the restored signal vector y _(t) is expressed by the following equation (34).

The FastICA algorithm proposed by Hyvarinen et al. Uses the kurtosis to extract independent components y _{i (t)} one by one in a batch mode. This solution is said to have good convergence and high speed.

At this time, the independence evaluation formula (34) of the restored signal vector y _(t) becomes a vector notation as in the following formula (35). And a partial differential equation (35) with w _i, when changing the w _i as the following equation (36), separating the vector w _i to extract the i-th independent component y _{i (t),} from the change characteristic Calculate

  In Equation (36), g is a gradient of a quartic equation, so it becomes a cubic equation. However, considering the stability of the algorithm, for example, use the sigmoid function tanh as shown in Equation (37) below. Has been proposed.

The problem to be solved by FastICA finally results in obtaining a separation vector w _i that maximizes or minimizes the left side of equation (35) under the limiting condition of equation (38) below.

Therefore, if this problem is solved by the Newton-Raphson method, the update equations of the separation vector w _i for extracting the i-th independent component y _{i (t)} are the following equations (39) and (40).

Here, w to _i are the updated separation vectors. μ is a learning coefficient and is set to μ = 1. However, if there is a problem with convergence, it is set to a value smaller than 1, such as 0.1 or 0.01.

In the convergence calculation in the Newton-Raphson method, for example, a random number is set as the initial value of the separation vector w _i , and Expressions (39) and (40) are repeatedly executed.

Then, when it becomes less than the magnitude (norm) is sized to the difference or the sum of the separate vector w _i of pre-update and post-update separation vector w to _i as shown by the following formula (43), to have converged It determined, and the separation vector w _i for obtaining the separation vector w to _i at this time.

Consider the sum in equation (43), as described with reference to equation (14), if before and after the update separation vector matches well w to _i = w _i, the direction opposite to the separation vector If it is updated, that is, because it can be a solution even if the w~ _i = -w _i.

FastICA calculates up to n m-dimensional separation vectors w _i that are elements of the separation matrix W one by one. First, assuming that i = 1, the separation vector w ₁ for extracting the _first independent component y _{1 (t)} is obtained. Similarly, the separation vector w for extracting the second independent component y _{2 (t)} is obtained. ₂ calculated, will ultimately determine the separation vector w _n for extracting the n independent component y _{n (t).}

At this time, when the preprocessing of whitening is performed on the observation signal vector x _(t) , in the calculation after the second independent component, in the update of the separation vector w _i of the i-th independent component, the following expression (44 By orthogonalizing each row vector of (), it is possible to prevent the i-1 separated vectors that have already been restored from being duplicated and extracted.

The i-th independent component separation vector w _i is orthogonalized by, for example, the following equation (45).

When the calculation of the _nth separation vector wn is completed, an n-row × m-column separation matrix W to be obtained is determined. At this time, a statistically independent restoration signal vector y _(t) to be obtained is obtained by Expression (31).

  FIG. 6 is a flowchart showing the FastICA algorithm processing proposed by Hyvarian et al., Which has been described as an example of the ICA algorithm. In the present invention, the FastICA algorithm illustrated in FIG. 6 is imaged on the surface of a two-dimensional plane or a three-dimensional solid composed of two-dimensional planes reflecting the installation positions of a plurality of sensors installed in the blast furnace. Apply to data. In the present invention, it goes without saying that an ICA algorithm other than FastICA may be used.

(Application of ICA to image information measurement data in blast furnace operation)
From the measurement data converted into image information in the operation of the blast furnace, the implementation method of the ICA in the present invention will be described below, taking as an example the two-dimensional flat image of the furnace pressure and the furnace height in the shaft pressure exemplified in FIG.

  First, with respect to the furnace height direction / furnace circumferential direction two-dimensional plane development image of the shaft pressure generated by the image information processing unit 5 illustrated in FIG. 5, 14 points on the furnace height coordinate axis as illustrated in FIG. Fourteen reference coordinates are provided on the angle coordinate axis, and shaft pressure data on a grid constituted by these reference coordinates is extracted.

At this time, a lattice number illustrated in FIG. 7 is assigned to each lattice, and the extracted shaft pressure data is stored in a 14 × 14 = 196-dimensional vector x _(t) in association with the lattice number. At this time, the vector x _(t) corresponds to the observed signal vector x _(t) in ICA, and m = 196.

Here, an example in which image information is extracted with a 14 × 14 equidistant lattice in the application of ICA to measurement data converted into image information has been shown. However, in the present invention, the lattice does not need to be an equidistant lattice, Further, the number of grids, that is, the dimension number m of the observation signal vector x _(t) may be another value.

  Here, t is time. That is, the shaft pressure data is collected in time series by the data collecting device 3 and the image information processing unit 5 sequentially executes image information processing, whereby a certain amount of image data can be generated and collected.

For example, as one example, when image data of shaft pressure is generated at a time interval ΔT = 5 [min] continuously for one year in a certain blast furnace, 12 months × 30 days × 24 hours × 60 minutes ÷ 5 minutes = 103 , 680 image data can be collected, and the independent component analysis unit 6 generates an image data vector x _(t) using the reference coordinates illustrated in FIG. 7 for the collected image data.

Further, the independent component analysis unit 6 calculates the separation matrix W of the image data vector x _(t) and the independent component signal vector y _(t) using the ICA algorithm.

FIG. 8 and FIG. 9 are illustrated as examples of the independent component analysis unit 6. FIG. 8 shows that the shaft pressure data for one year from January 1st, 2001, 00:00:00 to December 31, 2001, 23:55, at a time interval Δt = 5 [ min] represents image information, and the separation matrix W calculated by the independent component analysis unit 6 is shown. At this time, the dimension number n of the independent component signal vector y _(t) of the shaft pressure was set to n = 10. Therefore, the separation matrix W is a matrix of n × m = 10 rows × 196 columns.

FIG. 9 shows the independent component signal vector y _(t) at this time. Figure 9 is a diagram showing an independent component signal vector y _(t) 1st component y ₁ from _(t) 10th component y ₁₀ of _(t) in time series. The horizontal axis of FIG. 9 is time t [min], 0 [min] is January 1, 2001 00:00, and 5.256 × 10 ⁵ [min] is December 31, 2001 23:55. Corresponds to minutes.

Here, an example in which the dimension number n of the independent component signal vector y _(t) is set to n = 10 is shown, but other values may be used in the present invention.

Here, as an example of the embodiment, an image data vector x _(t) is generated for shaft pressure data for a past blast furnace for one year, and its separation matrix W and independent component signal vector y _(t) are obtained. In the present invention, the shaft pressure data collected every moment by the data collecting device 3 is sequentially converted into an image data vector x _(t) and accumulated together with the past shaft pressure data in the present invention to cope with the time transition. Alternatively, the sequential separation matrix W may be recalculated to obtain the sequential independent component signal vector y _(t) .

(7. Similar Operation Case Search Department)
The operation monitoring method of the blast furnace operation in the present invention clarifies what characteristics the current operation state has, and accurately retrieves the past operation state in which it is similar, Then, future operation conditions are determined with reference to the transition of the operation state at that time and the operation conditions executed at that time.

That is, in the present invention, the independent component signal vector y _(t) and the separation matrix W calculated by the independent component analysis unit 6 are evaluated by the independent component / separation matrix evaluation unit 7, and the current operation state is determined using the result. Clarify what features it has.

  Furthermore, by searching exactly in which past operation state is similar in the similar operation example search unit 8, referring to the transition of the operation state and the operation conditions executed at that time, Determine future operating conditions.

Here, a method of searching for similar image data using the independent component signal vector y _(t) implemented by the independent component / separation matrix evaluation unit 7 and the similar operation example search unit 8 will be described.

The independent component analysis unit 6 calculates an n-dimensional independent component signal vector y _(t) . Here, the time t is assumed to be a discretization time t = 1, 2, 3,..., T discretized at an appropriate time period.

Between the n-dimensional independent component signal vector y _(t) , the n-row × m-column separation matrix W, and the m-dimensional image data vector x _(t) used for these calculations, the relationship of equation (12) Is established. Equation (12) shows that the image data vector x _(t) and the independent component signal vector y _(t) have a one-to-one correspondence via the separation matrix W, and the independent component signal vector y _{(t )} Match each other.

That is, when the independent component signal vector corresponding to the image data vector x _s of the image to be searched (hereinafter referred to as a teacher image) is _expressed by the following equation (46), the independent component signal vector of the image data vector x _(t) of the same image It is a necessary and sufficient condition that y _(t) becomes the following formula (47).

  The relationship between the teacher image and the image having high similarity can be considered as the similarity between the independent component signal vectors is higher as the similarity is higher, that is, the following equation (48).

  Therefore, as a method for efficiently searching for images having high similarity to the teacher image, attention is paid to the similarity of the independent component signal vectors, and a similarity index that is 1 is defined for the same image. An example of a method for searching for similar images of a teacher image using the is shown below. Here, by defining the similarity index in the form of a combination of several necessary conditions that the teacher image and the similar image satisfy, the similarity index increases as the value of the similarity index is larger and closer to 1. Make it correspond.

(Requirement 1)
A necessary condition in the relationship between the teacher image and the similar image is that the error square sum (norm) between the independent component signal vector y _s of the teacher image and the independent component signal vector y _(t) of the similar image is small. The definition and calculation method of the similarity index f ^* _{1 (t)} corresponding to the necessary condition 1 are illustrated in Expression (49) and Expression (50).

From this definition, the similarity index f ^* _{1 (t)} approaches ₁ as the image is similar to the teacher image, and the similarity index f ^* _{1 (t)} approaches 0 as it is not similar.
(Requirement 2)
A necessary condition for the relationship between the teacher image and the similar image is that the relationship shown in the following Expression 25 is satisfied.

  Since the separation matrix W has an inverse matrix, the following equation (51) is obtained from equation (12).

It can be seen that the image data vector x _(t) is a linear combination of statistically independent n-dimensional independent component signal vectors y _(t) . Necessary condition 2 is a condition that pays attention to the degree of coincidence of the independent components that becomes the most dominant when the image data vector x _(t) is constituted by a linear combination of the independent component signal vectors y _(t) . The definition and calculation method of the similarity index f ^* _{2 (t)} corresponding to the necessary condition 2 are exemplified in Expression (52) to Expression (55).

From this definition, the similarity index f ^* _{2 (t)} approaches 1 as the image is similar to the teacher image data, and the similarity index f ^* _{2 (t)} approaches 0 as it is not similar.

In the present invention, the similarity index f ^* _{1 (t)} corresponding to the necessary condition 1 and the similarity index f ^* _{2 (t)} corresponding to the necessary condition 2 are used as a similarity index by the expression (56). An exemplary similarity index f ^* _(t) is defined and calculated.

From this definition, considering the necessary condition 1 and the necessary condition 2 at the same time, the similarity index f ^* _(t) approaches 1 as the image is more similar to the teacher image, and the similarity index as the image is not similar f ^* _(t) approaches zero.

Furthermore, in the method for searching for similar image data in the present invention, when the independent component signal vector y _s of the teacher image is given, the similarity index f ^* _{( ) is calculated} from the independent component signal vector y _(t) of the image data in the search target range. _t) is calculated, and the independent component signal vector y _(t) is rearranged with respect to the discretization time according to the magnitude of the value of the similarity index f ^* _(t) . As a result, the independent component signal vectors y _(t) in the search target range can be rearranged in order of high similarity to the teacher image, and the time can be specified, and the similar image can be specified from the time.

  FIG. 10 shows a flowchart of the similar image data search method described so far.

Next, a similar image search embodiment according to the similar image data search method shown in FIG. 10 will be described. The shaft pressure image when the shaft pressure distribution illustrated in FIG. 5A is good (09:50 on December 20, 2001) is used as a teacher image, and a similar image is obtained from the image data for one year in the embodiment. Search for. That is, the similarity index f ^* _(t) is calculated from the independent component signal vector y _(t) of the image data for one year exemplified in the above embodiment, and the value of the similarity index f ^* _(t) is close to 1. Three images from the time image are sequentially set as a first candidate, a second candidate, and a third candidate for similar images. FIG. 11 shows the rearrangement result based on the similarity index f ^* _{(t) at} this time, and FIG. 12 shows three similar images specified from the time of each candidate. In FIG. 12, the searched three similar images are similar to the teacher image in the order of search candidates, and this search method is effective.

Similarly, the shaft pressure image when the shaft pressure illustrated in FIG. 5B is poor (06:50 on December 15, 2001) is used as a teacher image, and image data for one year of the embodiment is used. Search for similar images. That is, the similarity index f ^* _(t) is calculated from the independent component signal vector y _(t) of the image data for one year exemplified in the above embodiment, and the value of the similarity index f ^* _(t) is close to 1. Three images from the time image are sequentially set as a first candidate, a second candidate, and a third candidate for similar images. FIG. 13 shows the rearrangement result based on the similarity index f ^* _{(t) at} this time, and FIG. 14 shows three similar images specified from the time of each candidate. In FIG. 14, the three similar images searched are similar to the teacher image in the order of search candidates, and this search method is effective.

  From the results of FIGS. 10 to 14, the similar image data retrieval method according to the present invention can distinguish two characteristic image examples of the developed image of shaft pressure and retrieve similar images of each image.

The similar image data search method according to the present invention extracts the huge amount of information of each image as an independent component signal vector y _(t) that is a feature quantity of a small number of dimensions, and what kind of features it has. It is possible to clarify and efficiently search for an image similar to a teacher image from a large amount of image data, and to quantitatively evaluate the similarity as a similarity index f ^* _(t). It has the feature that it can do.

At this time, the separation matrix W can be calculated in a short time by using, for example, the FastICA algorithm proposed by Hyvarian et al. Also, the similarity index f ^* _(t) is calculated from the independent component signal vector y _(t) of the image data to be searched, and the independent component signal vector y is determined according to the value of the similarity index f ^* _{(t). The} process of rearranging _(t) with respect to the discretization time and searching for similar image candidates in order from an image with a similarity index f ^* _(t) value close to 1 is completed in a short time.

Therefore, in the present invention, the shaft pressure data collected from time to time by the data collecting device 3 is sequentially converted into an image data vector x _(t) and accumulated together with the past shaft pressure data. W is recalculated, the sequential independent component signal vector y _(t) is obtained, the sequential similarity index f ^* _(t) is calculated, the independent component signal vector y _(t) is rearranged, and the accumulated image data It is possible to search for similar images sequentially from among them.

  At this time, for example, if the image data of the blast furnace operation state at the current time is used as a teacher image and the time of the image having high similarity to the current time is specified, the operation of the time is determined from various measurement data and operation diary records at the time. The transition of the state and the operation condition executed at the time can be confirmed, and the current operation state of the blast furnace and the transition of the future operation state can be predicted.

  Furthermore, it is possible to determine future operating conditions with reference to this information.

In FIG. 10, attention is paid to the absolute value y _smax of the maximum component of the independent component signal vector y _s of the teacher image as the necessary condition 2, but the necessary condition 2 is expanded and the relationship shown in the following _Expression 29 is noted. Then, the similarity index f ^* _{k + 1 (t)} is defined as in the equations (59) and (60), and these are combined as necessary, and the similarity index f as in the equation (61) ^* Needless to say, it is effective to define and use _(t) .

  In the present invention, the definition of the similarity index and the calculation procedure thereof may be methods other than those exemplified in equations (49) to (61) and FIG.

(8. Independent component / separation matrix time series transition evaluation unit)
The n-dimensional independent component signal vector y _(t) calculated by the independent component analysis unit 6 is evaluated by the independent component / separation matrix time series evaluation unit 9, and the operation monitoring unit 10 monitors the operation state of the blast furnace.

In the example of the above embodiment, FIG. 9 shows an independent component signal vector y ₍ 1) of a shaft pressure data image for one year from January 1, 2001, 00:00:00 to December 31, 2001, 23:55. from the first component y ₁ of _{t) (t)} illustrates the time series 10th component y ₁₀ a _(t).

Each independent component represents a statistically independent signal component that is latent in the shaft pressure data image. For example, a rectangular waveform expressed by the _fifth independent component y _{5 (t)} indicates a blast furnace storm from that time.

Moreover, the time when the blast furnace had poor air permeability and the operation was not good (t = 0.82 × 10 ^{5 to} 1.55 × 10 ⁵ [min]: February 26, 2001 22:40 to April 20 At 17:20), the waveform of the _second independent component y _{2 (t)} fluctuates greatly.

Therefore, as illustrated in FIG. 15, when the time series transition of the _second independent component y _{2 (t)} is monitored and the control value set in advance is exceeded, the air permeability of the blast furnace is poor. It can be monitored that it is in operation. In FIG. 15, for the sake of explanation, the upper management value y _2u = 1.0 and the lower management value value are set as the management values set in advance, taking the time series transition of the _second independent component y _{2 (t)} as an example. Although the case where y _2l = −1.0 is exemplified, in the present invention, other values may be used as the management value. Moreover, you may monitor the time-sequential transition of independent components other than the 2nd. Furthermore, you may monitor the time-sequential transition of the variable which combined the some independent component.

(9. Recording part)
In the recording unit 9, an operation diary describing calculation results, measurement data, and operation conditions in the image information processing unit 5, the independent component analysis unit 6, the similar operation case search unit 7 and the independent component / separation matrix time series evaluation unit 8. Is recorded as a file in text format, etc., and made into a database.

  In the present invention, it is not necessary to limit the database format, and other database formats such as an XML format may be used. And it is also possible to input the information recorded by the recording unit 9 into a file, for example, and evaluate the operating state of the blast furnace offline. In the present invention, it is not necessary to limit the offline data input form, and the data may be input from a database such as an XML format.

  In addition, in the present invention, it is not necessary to limit the transmission form and transmission method of the calculation result to the recording unit 9. For example, the calculation result is converted into a digital signal and the digital signal is transmitted to the recording unit 9. In addition, the digital signal may be further compressed by various data compression techniques to reduce the amount of transmission data, and may be transmitted to the recording unit 9 for recording. The digital signal may be transmitted and recorded to a recording unit 9 installed at a remote location away from the blast furnace equipment via an information transmission network such as Ethernet (R), wireless LAN, and the Internet.

(10. Output unit)
In the output unit 10, the calculation results of the image information processing unit 5, the independent component analysis unit 6, the similar operation case search unit 7, and the independent component / separation matrix time series evaluation unit 8, for example, FIG. 5, FIG. 8, FIG. 9, FIG. 11, FIG. 12, FIG. 13, FIG. 14 and FIG.

  At this time, in the present invention, it is not necessary to limit the transmission form and transmission method of the calculation result to the output unit 10. For example, the calculation result is converted into a digital signal and the digital signal is transmitted to the output unit 10. In addition, the digital signal may be compressed by various data compression techniques to reduce the amount of transmission data and transmit it to the output unit 10. After transmission, the output unit 10 compresses the compressed transmission data. The digital signal may be restored and output to the previous digital signal, and further, for example, the digital signal is remote from the blast furnace facility via an information transmission network such as LAN, Ethernet (R), wireless LAN, and the Internet. It may be transmitted to the output unit 10 installed in the output. At that time, in the method of the present invention, it is not necessary to limit the data compression method.

  The data processing apparatus 4 according to the embodiment described above is configured by a CPU or MPU of a computer, a RAM, a ROM, and the like, and can be realized by operating a program recorded in the RAM or ROM. Therefore, it can be realized by recording a program that causes a computer to perform the above functions on a storage medium and causing the computer to read the program. As the storage medium, a CD-ROM, DVD, flexible disk, hard disk, magnetic tape, magneto-optical tape, nonvolatile memory card, or the like can be used.

  In addition, the functions of the above-described embodiments are realized by executing a program supplied by a computer, and the program code is shared with an OS (operating system) or other application software running on the computer. Needless to say, such program code is included in the embodiment of the present invention even when the functions of the above-described embodiment are realized.

  Although the present invention has been described above by taking shaft pressure data as an example from various measurement data, in the present invention, measurement data of other physical quantities of other parts and temporal changes of the measurement data (time (Change rate) may be used.

  Further, in the above embodiment, the monitoring target is a blast furnace, but the operation monitoring method in the present invention is a reactor (beer brewing tank, petroleum refining tower, nuclear reactor, heat exchanger that cannot directly detect the internal state quantity. Etc.).

It is a block diagram which shows the structure of the operation monitoring apparatus which is embodiment of this invention. It is a figure explaining the relationship between the coordinate system set to the blast furnace, and the several pressure sensor installation position on blast furnace equipment. Using the shaft part of the blast furnace as an example, the outer surface of the blast furnace equipment is represented by a three-dimensional combination of quadrilateral planar elements whose inner angles do not exceed 180 degrees. It is a figure explaining that the isoline of the pressure P1 in the quadrilateral plane element abcd can be searched by the “isoline search method using”. A blast furnace facility is represented by a shape that three-dimensionally combines quadrilateral plane elements that do not exceed 180 degrees, and “isoline search method using triangular elements that use the average of the four-vertex plane elements as vertices” on its surface. FIG. 7 is a diagram for explaining a contour figure formed by drawing an isoline of pressure searched by using an isoline and P1 obtained. It is a figure explaining the furnace height and the furnace peripheral direction two-dimensional plane expansion | deployment image of the shaft pressure computed in the image information-ized process part 5 for the example of the shaft part of a blast furnace. It is a figure explaining the flowchart which shows the process of FastICA algorithm which is one of the calculation methods of an independent component analysis (ICA). It is a figure explaining the example of the reference coordinate system when taking out the image data vector used for an independent component analysis from the shaft pressure image illustrated in FIG. It is a figure explaining the example of a calculation of the separation matrix computed in the independent component analysis part 6. FIG. It is a figure explaining the example of a calculation of the independent component calculated in the independent component analysis part 6. FIG. It is a flowchart explaining the example of the algorithm which defines a similarity index using an independent component in the similar operation example search part 7, and searches for a similar image. It is a figure explaining the result of having calculated a similarity index with the algorithm which searches for a similar image, and searching for the candidate with high similarity. It is a figure explaining the similar image obtained from the search result of FIG. It is a figure explaining another result which calculated the similarity index with the algorithm which searches for a similar image in the similar operation example search part 7, and searched the candidate with high similarity. It is a figure explaining the similar image obtained from the search result of FIG. It is a figure explaining the method to monitor the operation state of a blast furnace by evaluating the time series transition of an independent component in the independent component and separation matrix time series evaluation part 8. FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Blast furnace equipment 2 Various various sensors on blast furnace equipment 3 Data collection device 4 Data processing device 5 Image information processing part 6 Independent component analysis part 7 Similar operation example search part 8 Independent component / separation matrix time series evaluation part 9 Recording part 10 Output section

Claims (9)

The measurement data of the measurement target amount from multiple sensors installed in the blast furnace is arranged on the surface of the two-dimensional plane or the three-dimensional solid composed of the two-dimensional plane reflecting the installation position of each sensor, and the measurement data space It is a method for monitoring the operating state of a blast furnace in which an isoline on the surface is calculated with respect to a general distribution state and a temporal change, and the figure formed by the isoline or the feature information of the figure is evaluated as image information. ,
For the image information, an independent component analysis is performed to calculate an independent component and a separation matrix, a similarity index is defined using the calculation result, and an image similar to image information specified in advance from accumulated image information An operation monitoring method in blast furnace operation characterized by retrieving information. The measurement data of the measurement target amount from multiple sensors installed in the blast furnace is arranged on the surface of the two-dimensional plane or the three-dimensional solid composed of the two-dimensional plane reflecting the installation position of each sensor, and the measurement data space It is a method of monitoring the operating state of a blast furnace in which an isoline on the surface is calculated with respect to a general distribution state and a temporal change, and the figure formed by the isoline or the feature information of the figure is evaluated as image information. ,
For the image information, an independent component analysis is performed to calculate an independent component and a separation matrix, a similarity index is defined using the calculation result, and an image similar to image information specified in advance from accumulated image information An operation monitoring method in blast furnace operation, wherein information is retrieved and the order of the similarity is evaluated according to the magnitude of the similarity index. The measurement data of the measurement target amount from the sensors installed in the blast furnace is arranged on the surface of the two-dimensional plane or the three-dimensional solid composed of the two-dimensional plane reflecting the installation position of each sensor, and the measurement data space It is a method for monitoring the operating state of a blast furnace in which an isoline on the surface is calculated with respect to a general distribution state and a temporal change, and the figure formed by the isoline or the feature information of the figure is evaluated as image information. ,
For the image information, an independent component analysis is performed to calculate an independent component and a separation matrix, a similarity index is defined using the calculation result, and similar to the image information at a time specified in advance from the accumulated image information. A method for monitoring operation in blast furnace operation, wherein the time of a similar operation state is searched by searching for simple image information. The measurement data of the measurement target amount from multiple sensors installed in the blast furnace is arranged on the surface of the two-dimensional plane or the three-dimensional solid composed of the two-dimensional plane reflecting the installation position of each sensor, and the measurement data space It is a method for monitoring the operating state of a blast furnace in which an isoline on the surface is calculated with respect to a general distribution state and a temporal change, and the figure formed by the isoline or the feature information of the figure is evaluated as image information. ,
For the image information, an independent component analysis is performed to calculate an independent component and a separation matrix, a similarity index is defined using the calculation result, and similar to the image information at a time specified in advance from the accumulated image information. Search for similar image information in the past, and check the transition of the operation status at that time and the operation conditions executed at that time from various measurement data and operation diary records at that time. Then, the operation monitoring method in the blast furnace operation, wherein the operation condition is determined by predicting the transition of the operation state after the specified time. The measurement data of the measurement target amount from multiple sensors installed in the blast furnace is arranged on the surface of the two-dimensional plane or the three-dimensional solid composed of the two-dimensional plane reflecting the installation position of each sensor, and the measurement data space It is a method for monitoring the operating state of a blast furnace in which an isoline on the surface is calculated with respect to a general distribution state and a temporal change, and the figure formed by the isoline or the feature information of the figure is evaluated as image information. ,
For the image information, the independent component analysis is performed to calculate the independent component and the separation matrix, and the operation state of the blast furnace is monitored by comparing the time series transition of the independent component with the size of the control value set in advance. An operation monitoring method in blast furnace operation characterized by the above.   The blast furnace operation according to any one of claims 1 to 5, wherein the independent component and the separation matrix are sequentially updated using measurement data sequentially collected from a plurality of sensors installed in the blast furnace. Operation monitoring method. In the operation monitoring device for outputting and monitoring the operation status of the blast furnace,
Data collection means for collecting measurement data measured by various sensors installed on the blast furnace facility;
The spatial distribution state and temporal change of the collected measurement data are arranged on the surface of a two-dimensional plane or a three-dimensional solid composed of a two-dimensional plane reflecting the installation position of each sensor on the blast furnace equipment, and the measurement data Image information generating means for calculating an arbitrary isoline having the same value;
Independent component analysis means for performing independent component analysis of the image information calculated by the image information processing means;
Similar operation case search means for searching for similar operation cases of a blast furnace from the independent component analysis results of the image information and / or independent component / separation matrix time series evaluation means for evaluating the independent components and the separation matrix in time series,
Recording means for recording an operation diary describing the calculation results, measurement data, and operation conditions in the image information conversion means, independent component analysis means, similar operation case search means and / or independent component / separation matrix time series evaluation means,
An operation monitoring apparatus in blast furnace operation, comprising: output means for outputting calculation results in the image information conversion means, independent component analysis means, similar operation case search means, and independent component / separation matrix time series evaluation means.   The computer program for making a computer perform the process sequence of the operation monitoring method of any one of Claims 1-6.   The computer program for functioning a computer as each means of the operation monitoring apparatus of Claim 7.

Images