JP2003141215A - Operation analyzing device, method, computer program and computer readable storage medium, in manufacturing process - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make properly analyzable the operations and quality data of processes. SOLUTION: This operation analyzing device is provided with an operational factor space separating part 102 for separating an operational factor space using process operation data as a base vector into local areas, a local relational expression calculating part 104 for introducing local relational expressions for expressing relevance between operational factors and quality in the respective local areas, an activity function calculating part 103 for introducing activity functions for calculating contribution ratio of the respective local relational equations to the whole based on operational data, a mathematical expression model calculating part 105 for introducing mathematical expression models for expressing relation between the operational factors and quality as superposition of the local areas having the local relational expressions and the activity functions, a minimum error mathematical expression model selecting part 106 for selecting the minimum error mathematical expression model among the mathematical expression models calculated for a plurality of divided patterns and a learning error evaluating part 107 for further finely separating the operational factor space according to an error of the minimum error mathematical expression model and for judging processings in the respective means to be repeatedly executed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an operation analysis device, a method, a computer program, and a computer-readable storage medium in a manufacturing process, and in particular, in a general process in which quality is determined as an operation result, a plurality of operation factors and quality The present invention relates to a technique suitable for elucidating the cause of quality incompatibility by clarifying the relevance and finding operating conditions for obtaining a desired quality.

[0002]

2. Description of the Related Art Conventionally, in a process in which quality is determined as an operation result, an operation analysis method for analyzing the influence of operating conditions on quality is a correlation evaluated using a correlation coefficient between a single operation factor and quality. It is well known that an analysis method and a method of creating and evaluating a multiple regression model in which a plurality of operation factors are input and quality is output are evaluated.

Further, as a model for analyzing a more complicated relation between operating factors and quality, there is a document "JR Quinlan,
"Learning with continuous classes" Proceedings of
the5th Australian Joint Conference on Artificial
Intelligence. AI '92 ,, 1992, Pages 343-348 ”, a method using a decision tree is known.

Further, in the method disclosed in Japanese Patent Application Laid-Open No. 6-304723, the operating factors are physical property values such as carbon content of cast slab in the steel process, operating values such as casting width, and temperature values in the cooling zone. A quality control device using a multi-layer neural network is trained by using surface defects of steel sheets as quality data to perform quality control diagnosis.

[0005]

However, in the method using the correlation coefficient and the multiple regression model, the precondition is that the operation and quality data to be analyzed can be represented by a single linear model in all operation ranges. Since the correlation coefficient and regression model are derived and analyzed based on the above, when analyzing operation and quality data obtained from a process in which there are multiple quality mismatch factors each having different characteristics, the relationship between the two There was a problem that could not be captured correctly.

Further, in the method using the decision tree, the operation factor space having the operation data as a base vector is divided based on the operation condition, and a process of deriving a linear model of the operation factor and the quality in each local space is performed. By doing so, we are analyzing the operation data of a process that has multiple quality mismatch factors, but we assume that the boundary line of the local region is set strictly and different linear equations are established on both sides. However, there is a problem that it is not possible to perform a sufficiently accurate analysis on the data of a process in which there is a boundary region in which a plurality of disagreement factors are superimposed and which determines quality.

In the method disclosed in Japanese Unexamined Patent Publication No. 6-304723, a model in which a relationship between an operating factor and quality is learned is created using a multilayer neural network and applied to quality control diagnosis. The multilayer neural network has a problem that it is extremely difficult for a human to read based on what logic the control diagnosis was made, and the operator cannot judge the rationality of the control diagnosis result.

The present invention has been made in view of the above points, and operates a process in which there are a plurality of quality incompatibility factors, and there is a boundary region for determining quality by further superimposing these factors. The purpose is to be able to analyze quality data.

[0009]

The operation analysis device in the manufacturing process of the present invention is an operation analysis device which analyzes the operation and quality data in the manufacturing process and analyzes the relationship between a plurality of operation factors and quality. , Data input means for inputting process operation data and quality data, operation factor space dividing means for dividing the operation factor space having the process operation data as a base vector into a plurality of local regions, and operation factor and quality of each local region Local relational expression calculating means for deriving a local relational expression expressing relevance, and activity function calculating means for deriving an activity function for calculating a contribution ratio of each local relational expression to the whole based on operation data, Mathematical model calculation means for deriving a mathematical model representing the relation between the operation factor and the quality as a superposition of the local relation and the local region having the activity function, Among the mathematical models obtained for the division patterns of, the minimum error mathematical expression model selecting means for selecting the minimum error mathematical expression model having the smallest error, and the operation factor space if the error of the minimum error mathematical expression model is larger than the given convergence judgment factor. Is further subdivided, and learning error evaluation means for determining to repeatedly execute the processing by each of the above means is characterized.

Another feature of the operation analysis apparatus in the manufacturing process of the present invention is that the local relational expression is a linear polynomial between operation and quality data.

Another feature of the operation analyzing apparatus in the manufacturing process of the present invention is that the activity function is based on a normal distribution function having a center at the center of gravity of a local region. There is a point.

Another feature of the operation analysis apparatus in the manufacturing process of the present invention is that the quality condition corresponding to a specific operation condition is a condition, and a plurality of the above-mentioned items are used as means for explaining the relationship between them. It is provided with an analysis result display means for presenting the local relational expression and the contribution ratio of each local relational expression.

Another feature of the operation analysis apparatus in the manufacturing process of the present invention is that it is applied to a steel process and the quality data is the number of surface defects and internal defects of the product.

The operation analysis method in the manufacturing process of the present invention is a method for analyzing the operation and quality data in the manufacturing process to analyze the relation between a plurality of operation factors and quality, and to analyze the process operation data and the quality data. Data input procedure for inputting, operation factor space division procedure for dividing the space whose process operation data is the basis vector into multiple local regions, and local relational expression expressing the relation between the operation factor and quality in each local region A local relational expression calculation procedure, an activity function calculation procedure for deriving an activity function for calculating the contribution rate of each local relational expression to the whole based on operation data, and the local relational expression and the activity function. A mathematical model calculation procedure for deriving a mathematical model that expresses the relationship between operation factors and quality as a superposition of local regions, and the number obtained for multiple division patterns Among the models, the minimum error formula model selection procedure for selecting the minimum error formula model with the least learning error, and if the error of the minimum error formula model is larger than the given convergence judgment factor, the operation factor space is further subdivided. It is characterized in that a learning error evaluation procedure for determining to repeatedly execute each of the above procedures is executed.

The computer program of the present invention is characterized in that the computer functions as each means of the operation analysis device. Further, another computer program of the present invention is characterized by causing a computer to execute each procedure of the operation analysis method. The computer-readable storage medium of the present invention is characterized in that the computer program is stored.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION Preferred embodiments of an operation analysis device, a method, a computer program, and a computer-readable storage medium in the manufacturing process of the present invention will be described below with reference to the drawings.

FIG. 1 is a diagram showing the construction of an operation analysis device in the manufacturing process of this embodiment. 1 shown in FIG.
Reference numeral 01 is a data input unit, and operation data in the manufacturing process and quality data corresponding to the operation are input to the operation analysis device. The above-mentioned operation data is, for example, the fluctuation level of the molten metal surface in the continuous casting process in the steel process, the heating furnace temperature in the hot rolling process, and the like, and is given as a continuous value. When p operation factors u1, u2, ..., Up are given for N cases, the input operation data is a matrix with N rows and p columns. Further, the quality data is, for example, the number of surface defects per automobile steel plate coil in the steel process, and is given as a continuous value. When N cases of quality data are given corresponding to the operation data, the input quality data is an N-dimensional vector. Given the input operation data, which is a matrix of N rows and p columns, and the quality data, which is an N-dimensional vector, the quality data is u1 from the linear algebra theory.
It can be regarded as N points distributed in a p-dimensional operation factor space having .up. Therefore, if the quality is represented by the symbol y, it can be considered that the operation factor and the quality generally have the relationship of the formula (1) shown in the following Expression 1 via the mapping function f (·).

[0018]

[Equation 1]

In the present invention, the local relational expression and the activity function are derived on the assumption that the equation (1) is further expressed by the superposition of the local relational expression and the local region having the activity function. As a specific superposition of local regions, for example, a linear sum as shown in Expression (2) below can be used. Here, Σ represents the sum of terms, and M represents the number of local regions.

[0020]

[Equation 2]

Reference numeral 102 shown in FIG. 1 denotes an operation factor space division unit, which has means for dividing the operation factor space into local regions. When the space is already divided into some local areas, the local area is further subdivided. As a specific division method, for example, the model error of the local relational expression in the local region in the previous division pattern is evaluated,
The local area with the largest error is further divided into two. At this time, in the p-dimensional operation factor space, there can be p division patterns orthogonal to the respective axes u1 to up.
The processes up to the calculation of the mathematical model are executed for each of the p divisions.

Reference numeral 103 shown in FIG. 1 is an activity function calculation unit, which calculates an activity function based on the space division pattern obtained by the operation factor space division unit 102. The activity function is
It is possible to use an arbitrary function satisfying the normal condition expressed by the equation (3) shown in the following Expression 3.

[0023]

[Equation 3]

Specifically, for example, based on the normal distribution function μi having the center at the center of gravity of the local region expressed by the equation (4) shown below, the equation (5) shown below is obtained. The regular membership function defined by can be used as the activity function.

[0025]

[Equation 4]

Here, c _ij represents the center point of the local area, and σ _ij represents the standard deviation of the normal distribution function. FIG. 2 shows an example of the normal distribution function and the activity function when the one-dimensional operation factor space is divided into four local regions. Further, FIG. 3 shows an example in which the two-dimensional operation factor space is divided into three local regions.
Focusing on the boundary area of the local area, the activity functions in the areas on both sides of the boundary line overlap smoothly, so unlike a decision tree, multiple local relational expressions are superimposed to determine quality. You can express the situation that you are.

Reference numeral 104 shown in FIG. 1 denotes a local relational expression calculating unit, which calculates an undetermined coefficient of the local relational expression based on the functional system assumed in the local region. Arbitrary functions can be assumed for the functional system that is the basis of calculation. As a specific example, for example, a linear polynomial expressed by Equation (6) shown below can be used as the local relational expression.

[0028]

[Equation 5]

Here, w _ij represents an undetermined coefficient of the local linear model. Further, here, a method of calculating the undetermined coefficient in the equation (6) on the assumption of the linear polynomial will be described. When p-row N-column matrix data is input to the operation data and N-dimensional vector data is input to the quality data as the input data, it is assumed that the equation (7) shown in the following Equation 6 holds in each local region. There is.

[0030]

[Equation 6]

To determine the undetermined coefficient w _ij , a weighted error evaluation function based on the activity function and the expression (1
The undetermined coefficient is determined so that 1) becomes the minimum. this is,
This means that the error in the area with low activity is weighted so that it is evaluated small.

[0032]

[Equation 7]

The undetermined coefficient that minimizes the equation (11) is equal to the coefficient that satisfies the equation (12) shown in the following equation (8).
It is calculated by the matrix calculation of 3).

[0034]

[Equation 8]

Reference numeral 105 shown in FIG. 1 denotes a mathematical model calculation unit for operation and quality, which uses the activity function and the local relational expressions obtained by the activity function calculating unit 103 and the local relational expression calculating unit 104. The mathematical model of Expression (2) is constructed.

Reference numeral 106 shown in FIG. 1 is a minimum error mathematical model selection unit. In dividing the operation factor space, when there are a plurality of division degrees of freedom, a plurality of division patterns are created, an activity function and a local relational expression are calculated for each, and then a mathematical model is constructed. With respect to each mathematical model, the model error is evaluated by the formula (15) shown in the following Expression 9, and the model with the smallest error is adopted. Where y
_input is the input quality data.

[0037]

[Equation 9]

Reference numeral 107 shown in FIG. 1 is a learning error evaluation unit, which compares the error of the minimum error model obtained by the minimum error mathematical model selection unit 106 with a given error determination factor to obtain data with sufficient accuracy. Determine if a mathematical model that can explain As an error determination method, for example, a method of comparing the absolute value of the error with the error determination factor, a method of comparing the change amount of the error with respect to the increment of division with the error determination factor, and an evaluation function considering the number of divisions and the model error are used. For example, a method of aborting the division when the evaluation function increases with respect to the increase in the division calculated. Either way,
If the convergence is evaluated to be inadequate, the process from 102 to subdivide the operating factor space is repeated. 4 is 2
It is a figure which shows typically the mode of division | segmentation in the operation factor space of dimension. Since there are two types of divisions orthogonal to the axis of u1 or u2 with respect to the initial division state, the activity function, the local relational expression, and the mathematical model are configured for each division pattern, and the error is evaluated. . In the example of FIG. 4, u2
Since the division pattern orthogonal to the axis has the smallest error, the division pattern is adopted. Next, 2-1 and 2
-2, the model error of each local relational expression is evaluated, and the local region 2-1 having a large error is determined as a target of subdivision. Since there are two possible division patterns for the local region 2-1,
The above processing is repeated thereafter.

Reference numeral 108 shown in FIG. 1 is an analysis result display unit, which displays the finally obtained area division pattern of the mathematical model, the local relational expression in each local area, and the activity function distribution to display the data. Display information to analyze the underlying relations and their contributing situations.

According to the operation analysis apparatus according to the present embodiment described above, the operation factor space is divided into local regions and the relational expression for each local region is derived, so that there are a plurality of quality mismatch causes having different characteristics. The data obtained from the process can be analyzed correctly. In addition, even when analyzing process data in which there is a boundary region in which multiple disagreement factors are superimposed to determine quality, an activity function that smoothly connects the boundaries of each region is used compared to a decision tree. Since the mathematical model is constructed by using the mathematical model, the influence of the overlapped region can be appropriately evaluated.

[0041]

[Examples] In the following, the number of surface defects of plated steel sheets for automobiles was determined by using three factors, namely, the fluctuation level in the mold in the continuous casting process in the steel process, the drawing speed of the slab, and the heating furnace temperature in the hot rolling process as operating factors. An example in which an index normalized by the coil weight is used as the quality data will be described.

The data to be analyzed are the data for 420 coils collected by the process computer, and the operation data was obtained by averaging the time-series data at the operation timing corresponding to each coil as a representative value. Further, in the analysis, the operation results were normalized to a value of 0 to 1 in the existing data range for scaling. In the following description, the operation factor u1 corresponds to the fluctuation level of the molten metal in the mold, u2 corresponds to the slab drawing speed, and u3 corresponds to the heating furnace temperature.

FIG. 5 is a diagram showing the state of the operation factor space division number and the model error when the above operation and quality data are analyzed by this operation analysis apparatus. The horizontal axis shows the number of divisions of the space, and the vertical axis shows the error standardized by the number of data points obtained by squaring the difference between the actual value of quality and the model predicted value. Number of divisions from 4 to 5
Since the amount of decrease in the error is small even if the number is increased to 1, it is considered that the error has converged with the number of region divisions of 4, and the mathematical model was evaluated in this case. The area division pattern of the mathematical model shown in the analysis result display section and the local relational expression in each local area are as follows.

Local area 1: (u1, u2, u3) =
Region centered on (0.5, 0.25, 0.25) y = 1.7079 + 2.7834u1-1.0689u2-0.87
13u3

Local area 2: (u1, u2, u3) =
Region centered at (0.5, 0.25, 0.75) y = -0.1959 + 2.7234u1-0.2624u2 + 2.44
80u3

Local area 3: (u1, u2, u3) =
Region centered on (0.5, 0.75, 0.25) y = -1.4710 + 2.4671u1 + 6.5374u2-0.4
656u3

Local area 4: (u1, u2, u3) =
Region centered on (0.5, 0.75, 0.75) y = -4.5674 + 3.4088u1 + 7.9233u2 + 4.3
592u3

From the above results, the following characteristics regarding this process can be read. The coefficient of u1 (fluctuation level) is a positive value in all regions, and the number of flaws increases with an increase in fluctuation of the level.

The coefficient of u2 (cast strip drawing speed) has a negative value in regions 1 and 2 where u2 is small, and a positive value in regions 3 and 4 where u2 is large. Therefore, as the drawing speed increases, the number of flaws decreases when the speed is low, and the number of flaws increases when the speed is high.

The coefficient of u3 (heating furnace temperature) has a negative value in regions 1 and 3 where u3 is small, and has a positive value in regions 2 and 4 where u3 is large. Therefore, as the heating furnace temperature increases,
Defects decrease in the low temperature region and increase in the high temperature region.

From the coefficient values of the linear local relations in each region, u1 (melt level fluctuation amount) in regions 1 and 2 and u2 (cast slab drawing speed) in regions 3 and 4 have the greatest influence on the number of defects. Is a factor.

In FIG. 6, the number of flaws in the entire operating range is calculated from the mathematical model obtained by the analysis, and u1, u
Select two items from each operation factor of 2 and u3 and quality y
It is a three-dimensional plot of the relationship with. From this figure, it is possible to visually evaluate the influence of the above-mentioned operating factors for each area on the flaw. For comparison, FIG. 7 shows the state of simple correlation between each operation factor and the number of flaws. From this figure, it is difficult to obtain the analysis obtained in this example.

In this embodiment, the analysis device is realized as a program on a computer, but the calculation device,
It may be configured by hardware in which a memory or the like is combined.

The operation analysis apparatus of the present invention may be composed of a plurality of devices or one device.

Further, the above-described embodiment is composed of a CPU of a computer, an MPU, a RAM, a ROM or the like, and is carried out by operating a program recorded in the RAM or the ROM. Therefore, the program code itself of software for realizing the functions of the above-described embodiments, means for supplying the program code to a computer, for example, a storage medium storing the program code are included in the scope of the present invention.

[0056]

As described above, according to the present invention, the operation factor space division is performed, the activity function of the local region and the local relational expression are derived, and the mathematical model expressing the relation between the operation and the quality is derived. Therefore, it is possible to appropriately analyze the operation data of the manufacturing process in which there are a plurality of operation incompatibility factors and there is a region in which they are superposed to determine quality. For this reason, for example, if a quality mismatch occurs under a certain operating condition, the local relational expression and the activity function under that operating condition are calculated, and which operating factor and how are changed based on the coefficient of the highest active local relational expression. By doing so, it is possible to obtain a guideline as to whether a desired quality condition can be obtained. Further, in the case of operating conditions in which there are a plurality of local relational expressions having the same degree of activity, the operating condition for obtaining the desired quality is searched from each of the local relational expressions, and the effect is the improvement amount as a whole. Can be evaluated.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of an operation analysis device according to the present embodiment.

FIG. 2 is a diagram showing a local activity function distribution when a one-dimensional operation factor space is divided into four.

FIG. 3 is a diagram showing a local activity function distribution when a two-dimensional operation factor space is divided into three.

FIG. 4 is a schematic diagram showing how a two-dimensional operation factor space is subdivided.

FIG. 5 is a diagram showing a behavior of a model error with respect to the number of operating region divisions in the analysis of the example.

FIG. 6 is a diagram in which the number of flaws in the entire operation range is calculated from the mathematical model obtained by the analysis of the example, two items are selected from each operation factor, and the relationship with the quality y is plotted in a three-dimensional diagram. Is.

FIG. 7 is a diagram showing a single correlation between an operation factor and the number of flaws in the operation data analyzed in the example.

[Explanation of symbols]

101 Data input section 102 Operation factor space division unit 103 Activity Function Calculation Unit 104 Local Relational Expression Calculation Unit 105 Operation and quality mathematical model calculation unit 106 minimum error formula model selection unit 107 learning error evaluation unit 108 Analysis result display section

Claims (9)

[Claims] 1. An operation analysis device for analyzing operation and quality data in a manufacturing process to analyze the relationship between a plurality of operation factors and quality, comprising data input means for inputting process operation data and quality data. Operation factor space dividing means for dividing the operation factor space whose process operation data is the base vector into a plurality of local regions, and local relational expression calculation for deriving a local relational expression expressing the relation between the operation factor and quality in each local region Means, an activity function calculating means for deriving an activity function for calculating the contribution rate of each local relational expression to the whole based on the operation data, and a superposition of the local relational expression and the local region having the activity function As the mathematical model calculation means for deriving a mathematical model expressing the relation between the operation factor and the quality, and the mathematical model obtained for a plurality of division patterns, Minimum error formula model selection means for selecting a small minimum error formula model, and if the error of the minimum error formula model is larger than the given convergence determination factor, the operation factor space is further subdivided and the processing by each of the above means is performed. An operation analysis device, comprising: a learning error evaluation means that determines to repeatedly execute the operation. 2. The operation analysis device according to claim 1, wherein the local relational expression is a linear polynomial between operation and quality data. 3. The operation analysis device according to claim 1, wherein the activity function is a normal membership function configured based on a normal distribution function having a center at the center of gravity of a local region. . 4. An analysis result display that presents a plurality of the local relational expressions and a contribution ratio of each of the local relational expressions as a means for explaining the relationship between the specific operating condition and the corresponding quality condition as a condition. A means is provided.
The operation analyzer according to any one of 1 to 3. 5. The operation analysis apparatus according to claim 1, wherein the quality data is applied to a steel process and the quality data is the number of surface defects and internal defects of a product. 6. A method for analyzing operation and quality data in a manufacturing process to analyze the relationship between a plurality of operation factors and quality, comprising a data input procedure for inputting process operation data and quality data, and a process operation. An operation factor space division procedure that divides the space whose data is the basis vector into multiple local regions, and a local relational expression calculation procedure that derives a local relational expression that expresses the relation between the operation factor and quality in each local region, An activity function calculation procedure for deriving an activity function for calculating the contribution rate to the whole of the local relational expression based on the operation data, and an operation factor as a superposition of the local relational expression and the local area having the activity function. Among the mathematical model calculation procedure for deriving a mathematical model that expresses the relationship of quality and the mathematical model with the smallest learning error among the mathematical models obtained for multiple division patterns. A minimum error formula model selection procedure for selecting a formula model, and if the error of the minimum error formula model is larger than a given convergence determination factor, the operation factor space is further subdivided and it is determined to repeatedly execute each of the above procedures. An operation analysis method, characterized by executing a learning error evaluation procedure. 7. A computer program that causes a computer to function as each unit of the operation analysis apparatus according to claim 1. Description: 8. A computer program for causing a computer to execute each procedure of the operation analysis method according to claim 6. 9. A computer-readable storage medium having the computer program according to claim 7 or 8 stored therein.