JP5549536B2 - Analysis device, analysis method, and computer program - Google Patents

Description

  The present invention relates to an analysis apparatus, an analysis method, and a computer program. More specifically, the present invention relates to an analysis apparatus, an analysis method, and a computer program for analyzing a phenomenon by a modeling technique using process data and multimedia information such as an image.

  With the development of IT technology, multimedia information such as images and sounds can be easily handled as digital data. On the other hand, the development of data analysis technology has been remarkable in recent years, and coupled with a dramatic improvement in computer performance, the practical application of data analysis technology that can efficiently derive feature quantities and relational expressions from a large amount of data is rapidly progressing.

  The conventional modeling technique is basically based on correlation analysis using a continuous measurement amount of scalar values such as process data such as a multiple regression model. For example, Patent Literature 1 discloses a visual feedback control device that realizes predictive control by providing a model having an image predicting function in visual feedback control.

  Further, Patent Document 2 discloses a photographing unit for photographing sewage flowing out from an aeration tank in a sewage treatment process, an image processing unit for extracting a feature amount of an image photographed by the photographing unit, and an operation given to the sewage treatment process. A feature quantity predicting means for predicting a feature quantity after a predetermined time based on the quantity and the feature quantity extracted by the image processing means, and the feature quantity predicted by the feature quantity predicting means matches the target feature quantity. A sewage treatment process control device comprising: an operation amount correction unit that determines a correction amount of the operation amount; and an addition unit that adds the correction amount of the operation amount determined by the operation amount correction unit to the operation amount. Has been.

JP-A-6-266410 JP-A-8-52459 Japanese Patent No. 4150322 JP 2008-163381 A

  However, in the conventional modeling technique, when the independent variables and the dependent variables increase, the analysis requires a lot of time, and it is difficult to use the analysis results online. In addition, modeling techniques based on process data have limitations in estimating phenomena that have not been elucidated. On the other hand, it is also desirable to be able to estimate phenomena by analyzing quantitatively and objectively the phenomena that users judge based on qualitative and subjective criteria based on the human senses. . In other words, if a phenomenon can be analyzed using a quantitative / objective model, it can be predicted and estimated as a multimedia information such as an image or sound. This makes it possible to make evaluation judgments that make use of the high human perception ability.For example, an operation action that recognizes abnormal states and their precursors using such evaluation judgments and leads to a desirable control state that avoids abnormalities. Will be able to do.

  Accordingly, the present invention has been made in view of the above problems, and an object of the present invention is to use multimedia information to quantitate phenomena using images and audio, using quantitative and objective modeling techniques. It is an object of the present invention to provide a new and improved analysis apparatus, analysis method, and computer program that can be predicted and estimated as such multimedia information.

  In order to solve the above problems, according to an aspect of the present invention, an analysis apparatus for analyzing an analysis object generated in the process is provided in a process of generating the analysis object. The analysis apparatus includes: a first feature amount extraction unit that extracts an image feature amount representing a shape feature appearing in an image from an analysis target image of the analysis target; and process data from the process data in the process of generating the analysis target to the process data A second feature quantity extraction unit that extracts data feature quantities representing characteristics on the data that appears, an image feature quantity extracted by the first feature quantity extraction unit, and data extracted by the second feature quantity extraction unit A prediction unit that calculates a predicted image feature amount predicted for the analysis target generated in the process based on the correlation with the feature amount.

  The image feature amount is a contribution ratio of the analysis target object to the base image representing the characteristic shape with respect to the analysis target image, and the first feature amount extraction unit uses the independent component analysis to analyze the analysis target object. You may acquire the contribution rate to the base image with respect to an image. At this time, the prediction unit is generated by a process based on the correlation between the contribution ratio of the base image acquired by the first feature amount extraction unit and the data feature amount extracted by the second feature amount extraction unit. The estimated contribution rate to the base image for the analyzed object is calculated.

  The prediction unit calculates the Euclidean distance between the data feature quantity and each process data in the process of generating the analysis object, and predicts the image feature quantity associated with the process data that minimizes each Euclidean distance. It may be an image feature amount.

  Alternatively, the prediction unit calculates the Euclidean distance between the data feature amount and each process data in the process of generating the analysis object, and extracts a plurality of process data in which each Euclidean distance is a predetermined value or less as similar process data. A weighted average of image feature amounts associated with similar process data may be used as the predicted image feature amount of the analysis object.

  Further, the second feature quantity extraction unit may normalize the value of at least one item data among the item data included in the process data in the process of generating the analysis object, and obtain the data feature quantity.

  The analysis apparatus may further include an estimation unit that estimates the shape of the analysis target based on the predicted image feature amount calculated by the prediction unit and the base image.

The analysis object is, for example, a rolling material in a rolling process, and a tail end image representing the shape of the tail end of the rolling material may be used as the analysis target image.

Moreover, in order to solve the said subject, according to another viewpoint of this invention, in the process which produces | generates an analysis target object, the analysis method which analyzes the analysis target object produced | generated in the said process is provided. The analysis method includes a first feature amount extraction step for extracting an image feature amount representing a shape feature appearing in an image from an analysis target image of the analysis target, and a process data in a process of generating the analysis target to the process data. Based on the correlation between the image feature quantity and the data feature quantity, the second feature quantity extraction step for extracting the data feature quantity representing the characteristic on the data that appears, and the analysis object generated in the process is predicted seen including a prediction step of calculating the predicted image feature amount, the image feature amount, for analyzing the target image of an analyte, a contribution rate to the underlying image representing a characteristic shape, a first feature quantity extraction The step acquires the contribution ratio of the analysis object to the base image with respect to the analysis target image using independent component analysis, and the prediction step is acquired by the first feature amount extraction step. Based on the correlation between the contribution rate and the data feature value extracted by the second feature extraction step of the base image, and calculates the contribution rate to the underlying image to be predicted for the analyte to be generated in the process It is characterized by that.

Furthermore, in order to solve the above-described problem, according to another aspect of the present invention, a computer is used to extract an image feature amount representing a shape feature appearing in an image from an analysis target image of the analysis target. A feature amount extraction unit; a second feature amount extraction unit that extracts data feature amounts representing characteristics on data appearing in the process data from process data in the process of generating the analysis target; and a first feature amount extraction unit Based on the correlation between the image feature amount extracted by the above and the data feature amount extracted by the second feature amount extraction unit, the analysis target generated in the process of generating the analysis target is predicted A prediction unit that calculates a predicted image feature amount, wherein the image feature amount is a contribution rate of the analysis object to the base image representing the characteristic shape with respect to the analysis target image, and the first feature amount extraction unit Is The contribution ratio of the analysis object to the analysis target image with respect to the analysis target image is acquired using the standing component analysis, and the prediction unit includes the contribution ratio of the base image acquired by the first feature value extraction unit and the second feature value. Based on the correlation with the data feature value extracted by the extraction unit, the analysis target is generated by calculating the predicted contribution rate to the base image for the analysis target generated in the process. There is provided a computer program for causing an analysis apparatus to analyze an analysis object generated in a process.

  As described above, according to the present invention, multimedia information can be used to predict and estimate phenomena, as multimedia information such as images and sounds, using quantitative and objective modeling techniques. Possible analysis devices, analysis methods and computer programs can be provided.

It is explanatory drawing which shows the process outline | summary of the analyzer which concerns on embodiment of this invention. It is explanatory drawing which shows the system configuration | structure which collects the data in the rolling process which concerns on the same embodiment. It is a block diagram which shows the structure of the analyzer which concerns on the same embodiment. It is explanatory drawing which shows an example of a tail end image. It is a graph which shows the luminance distribution acquired from the tail end local image. It is a graph which shows the state which normalized the luminance distribution shown in FIG. FIG. 7 is a contour diagram showing the luminance distribution of FIG. 6. It is explanatory drawing which shows the image feature-value extraction process by independent component analysis. It is explanatory drawing which shows the relationship between the mixing matrix by an independent component analysis, and an independent component matrix. It is a base image which shows the tail end shape of a normal coil. It is a base image which shows the tail end shape of a coil with a sharp point on the WS side. It is a base image which shows the tail end shape of a coil with a sharp point on the DS side. It is a base image which shows the tail end shape of a coil in case there exists cooling water. It is a base image which shows the tail end shape of a coil when there is no cooling water. It is explanatory drawing explaining the meaning of the independent component vector acquired from the tail end local image. It is explanatory drawing which shows the calculation process of the data feature-value of the process data regarding leveling. It is explanatory drawing which shows an example of the histogram of the process data regarding leveling. It is a block diagram which shows the structure of the analyzer which performs modeling using a multiple regression model. It is explanatory drawing which shows the example of 1 structure of a modeling data storage part. It is explanatory drawing which shows the tail end local image and brightness | luminance matrix of WS sharpness and DS sharpness which are verified in Example 1. FIG. It is a graph which shows the leveling of the similar leveling coil whose leveling is similar to the WS sharp coil of FIG. It is explanatory drawing which shows an example of the brightness | luminance matrix of the similar leveling coil whose leveling is similar to the WS sharp coil of FIG. It is a graph which shows the independent component vector of the similar leveling coil whose leveling is similar to the coil of WS sharp of FIG. FIG. 17 is a graph showing leveling of a similar leveling coil whose leveling is similar to the DS-pointed coil of FIG. 16. It is explanatory drawing which shows an example of the brightness | luminance matrix of the similar leveling coil in which leveling is similar to the DS sharp coil of FIG. It is a graph which shows the independent component vector of the similar leveling coil in which leveling is similar to the coil of DS sharpness of FIG. It is a graph of the tail end local image which has WS sharp tendency from the Euclidean distance of an independent component vector, and those process data. It is a graph of the tail end local image which has DS sharpness tendency from the Euclidean distance of an independent component vector, and those process data. It is a graph of the tail end image which has a fishtail shape tendency from the Euclidean distance of an independent component vector, and those process data. It is a graph of the tail end local image which has a normal shape from the Euclidean distance of an independent component vector, and those process data.

  Exemplary embodiments of the present invention will be described below in detail with reference to the accompanying drawings. In addition, in this specification and drawing, about the component which has the substantially same function structure, duplication description is abbreviate | omitted by attaching | subjecting the same code | symbol.

<1. Overview of multimedia modeling>
First, based on FIG. 1, the outline | summary of the analyzer based on the multimedia modeling which concerns on embodiment of this invention is demonstrated. FIG. 1 is an explanatory diagram showing a processing outline of the analyzer according to the present embodiment.

  The analysis device according to the present embodiment is a device that performs modeling based on the correlation between the tail end shape of hot rolling finish rolling in a rolling process of a metal plate such as a steel plate and rolling process data. The analysis device captures the tail end of the rolling material 12 on the finish rolling delivery side of the rolling process with a camera such as a CCD camera to acquire a tail end image, and extracts an image feature amount (Y) based on the tail end image. The image feature amount is quantitative information representing the shape feature appearing in the image, and the image can be specified based on the image feature amount. In the present embodiment, an independent component vector acquired using ICA described later is an image feature amount.

  For example, as shown in FIG. 1, the tail end of the rolled material 12 has a tongue-like shape with both ends rounded when in a normal state. However, when the rolling force of the rolling machine 12 that sandwiches the rolling material 12 varies or the parallelism of the rolling roll 14 is low, the pressure applied to the rolling material 12 varies and the rolling material 12 tails. The ends have a single-pointed shape or a fishtail shape with sharpened ends. The analysis device extracts the feature of the shape of the tail end of the rolled material 12 as an image feature amount.

  Further, the analysis device extracts the data feature amount (X) from the process data expressed as a scalar value such as a set value or a sensor measurement value in the rolling process operation. The data feature amount is information representing characteristics on data appearing in the process data. Since the process data shows different values for each rolled material 12, the process data can be analyzed and its characteristics can be extracted as data feature amounts and used as information for finding the characteristics of the rolled material 12. The process data is time series data of set values in the rolling process and measured values such as plate temperature and plate thickness, and represents the state of the rolling material 12 according to the setting of the rolling process. The analysis device extracts a data feature amount based on the process data.

  The analyzer builds a model F (Y = F (X)) for estimating the state of the coil generated or manufactured in the rolling process based on the image feature (Y) and the data feature (X). To do. Based on the constructed model, the analysis device estimates, for example, the shape of the tail end of the rolled material 12 to be subsequently generated, detects an abnormality in the equipment of the rolling process from the estimated shape of the tail end, and generates it. The quality of the rolled material 12 to be produced can be predicted.

  Conventionally, a user visually recognizes the shape of the tail end of the rolled material 12 and makes a judgment based on a qualitative / subjective standard to detect the occurrence of an abnormality in a rolling process facility. However, in the analysis apparatus according to the present embodiment, information that has been conventionally determined based on the five senses can be determined based on a quantitative and objective model. In other words, if a phenomenon can be analyzed using a quantitative / objective model, it can be predicted and estimated as a multimedia information such as an image or sound. This makes it possible to make evaluation judgments that make use of the high human perception ability.For example, an operation action that recognizes abnormal states and their precursors using such evaluation judgments and leads to a desirable control state that avoids abnormalities. Will be able to do. Hereinafter, the configuration of the analysis apparatus according to the present embodiment capable of making such an evaluation determination, and the analysis processing performed thereby will be described in detail.

<2. System configuration>
[Data collection in rolling process]
First, based on FIG. 2, the system configuration | structure which collects the data in the rolling process 10 which concerns on this embodiment is demonstrated. FIG. 2 is an explanatory diagram showing a system configuration for collecting data in the rolling process 10 according to the present embodiment.

  The analysis apparatus according to the present embodiment detects an abnormality of equipment in the rolling process 10 and predicts the quality of the generated rolling material 12 based on the process data in the rolling process and the shape of the tail end of the rolling material 12. For this reason, it is necessary to obtain from the rolling process 10 process data and an image of the shape of the tail end to be analyzed by the analyzer.

  The rolling process 10 consists of hot rolling in which a heated rolling material 12 is sandwiched between rolls 14 and stretched to make it thinner. The state of the rolling material 12 in each step of the rolling process 10 can be grasped from the plate thickness, temperature, etc. of the rolling material 12 acquired by a plate thickness measurement sensor, a temperature sensor, or the like. Moreover, the state of the equipment in rolling the rolling material 12 can be grasped | ascertained by the rolling load, leveling, etc. in each roll 14 detected by the sensor. As shown in FIG. 2, the process data including the coil state and the equipment state acquired by these sensors is transmitted to the data collection server 20 via the network 15 as time series data. The data collection server 20 stores the received process data in the process data DB 22.

  The state of the rolling material 12 in the rolling process 10 is also monitored using moving images acquired by one or a plurality of cameras (for example, video cameras) 31 and 32 installed at a predetermined position. The moving image data acquired by the cameras 31 and 32 can be confirmed on monitors 33 and 34 installed at the site. The moving image data is transmitted to the moving image collection server 40 via the network 35. The moving image collection server 40 stores the received moving image data in the moving image DB 42. The analysis device acquires a still image (tail end image) of the tail end portion of the rolling material 12 used for analysis from the operation data stored in the moving image DB 42. The tail end image acquisition process will be described later.

  With such a system, the process data of the rolling process 10 to be operated and the moving image data for obtaining the tail end image are automatically stored in the data collection server 20 and the moving image collection server 40.

[Analyzer configuration and analysis processing]
Using the process data and moving image data collected by the system configuration as shown in FIG. 2, the analysis apparatus 100 executes an analysis process. The configuration of the analyzer 100 is shown in FIG. As shown in FIG. 3, the analysis apparatus 100 includes an image feature amount analysis unit 110, a modeling unit 120, a model unit 130, a base image DB 140, and a modeling data storage unit 150.

(1) Image feature amount analysis unit 110
The image feature amount analysis unit 110 acquires the tail end image of the rolling material 12 from the moving image data stored in the moving image DB 42 of the moving image collection server 40, and extracts the image feature amount of the tail end image. The image feature quantity analysis unit 110 corresponds to a first feature quantity extraction unit. First, the image feature quantity analysis unit 110 acquires a tail end image as an analysis target image from moving image data. In the rolling process 10, about 600 to 800 materials are rolled per day, and the image feature amount analysis unit 110 acquires a tail end image of each rolled material. Thereby, for example, a tail end image 200 of the rolled material 12 as shown in FIG. 4 is obtained.

  In each image of FIG. 4, “WS” is a work side on the left side, and “DS” on the right side is a drive side on which a mill motor or the like for driving a rolling mill is installed. Here, the image included in the moving image data is an image obtained by observing the rolling material 12 before being wound by the coiler at a fixed point, and the image feature amount analysis unit 110 determines the tail end portion from the image of the rolling material 12 flowing on the rolling mill. A still image (tail image) including the shape of the image must be extracted. Therefore, the image feature amount analysis unit 110 acquires the tail end image by the following tail end image acquisition processing.

(Tail edge image acquisition process)
One moving image data is created for each rolling material. The image feature amount analysis unit 110 first cuts out a still image near the tail end from the moving image data of the rolling material 12 using the still image capturing device 44. The still image capturing device 44 receives a signal output when the tail end of the rolling material 12 passes through the rolling roll 14 and executes a capture process to acquire a still image near the tail end from the moving image data.

  Next, the image feature amount analysis unit 110 extracts an image feature amount from the tail end local portion (hereinafter referred to as “tail end local image”) of the still image near the tail end acquired by the still image capture device 44. Therefore, the luminance matrix which is numerical data representing the luminance of the tail end local image is normalized. This is because the size (number of rows and number of columns) of the luminance matrix of the tail edge local image extracted for each rolling material varies, and by normalization, for example, 20 × 50 The purpose is to unify the size of the luminance matrix of the image. When the luminance matrix of the tail end local image is represented by a three-dimensional graph, for example, FIG. The protruding part becomes brighter. The scales in the sheet width direction and the rolling direction represent the number of columns (number of pixels) constituting the image.

  The image feature amount analysis unit 110 normalizes the matrix of the tail end local image in order to extract the image feature amount later. For example, as shown in FIG. 6, it may be a 20 × 50 normalization matrix, a 20 × 40 normalization matrix, or a 10 × 20 normalization matrix, and the degree of normalization is appropriately determined. be able to. When the normalization matrix shown in FIG. 6 is two-dimensionalized, a contour diagram representing a luminance change as shown in FIG. 7 is obtained. In FIG. 7, a portion having a high luminance value is represented by a light color, and a portion having a low luminance value is represented by a dark color.

  Further, the image feature amount analysis unit 110 converts the two-dimensional luminance matrix represented by the contour diagram into a one-dimensional column vector. For example, as shown in FIG. 8, the contour diagram of the tail end local image represents a two-dimensional (for example, 20 × 40) luminance matrix. The image feature amount analysis unit 110 arranges each row of the luminance matrix in a column and converts it into a one-dimensional (for example, 800 × 1) column vector. The conversion from the luminance matrix to the column vector is performed for each tail end local image. The above is the preparation processing for extracting the image feature amount from the tail end local image.

(Image feature extraction processing using ICA)
When a predetermined number of one-dimensional column vectors are acquired from the tail end local images of each rolling material, the image feature amount analyzing unit 110 analyzes these tail end local images and extracts image feature amounts. That is, the image feature amount analysis unit 110 functions as a first feature amount extraction unit. In this embodiment, an independent component analysis (ICA), which is a set for converting a signal in which unknown signal sources are mixed into a statistically independent signal, is used to obtain an image feature amount from a tail end local image. Extract. Image feature extraction using ICA can be performed by the methods described in Patent Documents 3 and 4, for example.

The signal matrix X is decomposed into a product of the mixing matrix A and the independent component matrix s using ICA. That is, there is a relationship represented by the following formula 1 among the signal matrix X, the mixing matrix A, and the independent component matrix s.
X = A · s (Formula 1)

  In the present embodiment, for example, as illustrated in FIG. 8, the image feature amount analysis unit 110 uses the 800 × 1 column vector acquired from each tail end local image as the tail end local image acquired by the daily operation. Are arranged (for example, m = 600 to 800) to generate an 800 × m signal matrix X. Also, the mixing matrix A obtained as a result of ICA is an 800 × k matrix in which column vectors representing the base image are arranged by the basis number k. Here, the base image is an image showing a feature shape with strong independence at the tail end. The basis number k is a parameter for executing ICA.

  Here, FIG. 10A to FIG. 10E show examples of base images obtained as a result of ICA when the base number k is 5. FIG. 10A can be interpreted as a base image showing the tail end shape of the normal rolled material 12. The tail end of the normal rolled material 12 has a tongue-like shape with a high brightness value on the finish side. FIG. 10B can be interpreted as a base image showing the tail end shape of the rolled material 12 having a sharp point on the WS side. On the left side of FIG. 10B that is the WS side, a portion with a high luminance value is biased. On the other hand, FIG. 10C can be interpreted as a base image showing a tail end shape of the rolling material 12 having a sharp point on the DS side. In such a base image, a portion having a high luminance value is biased toward the DS side on the opposite side of WS. Further, FIG. 10D can be interpreted as a base image showing the tail end shape of the rolling material 12 when there is cooling water, and FIG. 11E shows the base showing the tail end shape of the rolling material 12 when there is no cooling water. It can be interpreted as an image. The mixing matrix A is obtained as a result of ICA, and is a set of column vectors representing base images as shown in FIGS. 10A to 10E.

  As a result of ICA, the signal matrix X is decomposed into a mixing matrix A and an independent component matrix s as expressed in Equation 1 above. The independent component matrix s is a k × m matrix that is linearly related to the signal matrix X. Each column constituting the independent component matrix s corresponds to each tail end local image and serves as an image feature amount indicating the feature of the tail end local image, and is hereinafter also referred to as an independent component vector.

Based on FIG. 9, the relationship between the signal matrix X, the mixing matrix A, and the independent component matrix s will be described with a simpler example. In FIG. 9, the mixing matrix A and the independent component vector Si, that is, the i-th column of the independent component matrix s when the basis number k = 3 is set for the i-th column of the signal matrix X representing one tail local image. Showing the relationship. In this case, the i-th column of the signal matrix X representing the tail end local image is an 800 × 1 column vector. The mixing matrix A is an 800 × 3 matrix composed of column vectors representing three base images based on the base number k = 3. At this time, the independent component vector Si = [s1, s2, s3] ^T represents a contribution rate for each row of the mixing matrix A. That is, the component s1 represents the weight of the base image (1) composed of the first column of the mixing matrix A in the tail end local image, and the component s2 is composed of the second column of the mixing matrix A in the tail end local image. The component s3 represents the weight of the base image (3) composed of the third column of the mixing matrix A in the tail end local image.

  In other words, the independent component matrix s represents the contribution rate to each base image, and is an image feature amount representing the degree to which each base image contributes to the tail end local image. From this, it can be seen that the independent component matrix s shown in FIG. 8 represents each image feature quantity of the tail end local image included in the signal matrix X.

  The image feature amount analysis unit 110 extracts the tail end local image from the moving image data of each rolling material 12 by performing the above-described series of processes, calculates the luminance matrix of each tail end local image, and then stores them. A signal matrix combined as a vector is created, and a mixing matrix A and an independent component matrix s are calculated based on ICA. Each column of the independent component matrix s obtained as a result represents an independent component vector of each tail end local image. The value of the independent component vector calculated for the tail-end local image shown in FIG. 11 has a high contribution ratio of the base image with sharp points on the WS side (FIG. 10B), while the base image with sharp points on the DS side (FIG. It can be seen that the contribution ratio of 10C) is low. Therefore, it is considered that the rolling material 12 corresponding to the tail end local image has a tail with a sharp point on the WS side. As described above, the image feature amount analysis unit 110 can analyze the tail end of the rolling material 12 generated in the rolling process 10 and quantify the tail end shape for each tail end image using the independent component vector.

  Each independent component vector acquired by the image feature quantity analysis unit 110 is stored in the modeling data storage unit 150, and the mixing matrix A is stored in the base image DB 140 with each column as a base image. The image feature amount analysis unit 110 also generates a separation matrix W that is a real number matrix for generating restored signal matrices that are statistically independent from the signal matrix X in the ICA. The details of the separation matrix W are based on Patent Documents 3 and 4 above.

(2) Modeling unit 120
Returning to the description of FIG. 3, the modeling unit 120 creates a model representing a correlation with the tail end shape of the rolled material 12 based on the process data. The modeling unit 120 uses data stored in a modeling data storage unit 150 (to be described later) that stores an independent component vector Si and process data Pi that are image feature amounts, and a control condition Po output from the control computer 60. Create a model.

  That is, the modeling unit 120 acquires, from the modeling data storage unit 150, the independent component vector Si and the process data Pi that are image feature amounts acquired by the image feature amount analysis unit 110. Further, the control condition (set value) calculated by the control computer 60 based on the manufacturing condition is input to the modeling unit 120 from the control computer 60. The control conditions are multidimensionalized and normalized by the modeling unit 120 to become process data Po.

(Acquisition of process data Po (data feature extraction processing))
First, acquisition of the process data Po will be described. The process data Po can be generated by selecting items closely related to the shape of the tail end among various conditions in the rolling process 10 based on operational knowledge. In the present embodiment, for example, leveling (L) indicating the inclination of the gap formed between the opposing rolling rolls 14, the differential load (G) at both ends of the rolling roll 14, and the conveyance direction of the rolling material 12 are adjacent. The tension (T) applied to the rolling material 12 between the rolling rolls 14 is used. In addition, as the target data for generating the process data Po, for example, a plate temperature difference at both ends of the rolling roll 14 can be used.

  The modeling unit 120 calculates a representative value for each rolling stand for each item as the data feature amount of the process data Po. As an example, a method of calculating a representative value for each rolling stand related to leveling will be described with reference to FIG. Leveling (L) in the i-th rolling stand (# i-std) has a large leveling value when the rolling material 12 passes between the rolling rolls 14 and rolling starts (ON) as shown in FIG. Thus, the leveling value is small when the rolling material 12 passes between the rolling rolls 14 and rolling is finished (OFF).

  At this time, the representative value of the leveling of the i-th rolling stand can be, for example, an average value of the leveling values during rolling. Here, since the fluctuation of the value is large in the vicinity of the rolling start time (ON) and the rolling end time (OFF), for example, data for a predetermined time after the start of rolling and a predetermined time before the end of rolling calculate a representative value. May be excluded from the data. The predetermined time can be, for example, about 5 seconds. Considering this, the average value calculated using the data from time ts to te in FIG. 12 becomes the representative value of the leveling of the i-th rolling stand. Thus, the modeling unit 120 calculates a representative value for each rolling stand for each item.

  Thereafter, the modeling unit 120 aggregates the calculated representative values for each rolling stand for each item, and generates a histogram. In FIG. 13, the total histogram of the representative value for every rolling stand regarding leveling is shown. In FIG. 13, F <b> 1 to F <b> 4 indicate the number of the rolling stand, and corresponds to the number i of the rolling stand in FIG. 13. Further, the horizontal axis of each histogram indicates a leveling representative value, and the horizontal axis indicates the number of rolling materials 12. Here, data on 640 sheets of rolled material 12 is tabulated. FIG. 13 shows the distribution of leveling representative values in each rolling stand. In each histogram, the hatched portion indicates the 3σ range, and the position indicated by the thick line indicates the average value of the leveling representative values in each rolling stand.

  As shown in FIG. 13, the average value of the leveling representative values in each rolling stand is not necessarily zero. For example, the leveling representative value is biased toward the WS side as in the second rolling stand (F2) and the gap between the WS-side rolling rolls 14 is large, or the DS as in the third rolling stand (F3). The leveling representative value is biased to the side, and the gap between the DS-side rolling rolls 14 may be large. Thus, the leveling characteristics of each coil, i.e., process data, is obtained from the distribution of the leveling representative values in each rolling stand by an i-dimensional vector (i = 7 in FIG. 13) normalized by N (0, 1) for each rolling stand. Po is represented.

  As described above, the process of extracting the data feature amount of the process data Po for leveling has been described, but the process data Po is similarly generated for items having a strong relationship with the shape of the tail end among various conditions in the other rolling processes 10. be able to. That is, the modeling unit 120 functions as a second feature amount extraction unit.

(Acquisition of independent component vector So)
The modeling unit 120 uses the acquired process data Po and the data (Pi, Si) stored in the modeling data storage unit 150 to acquire an independent component vector So that is a prediction result by the analysis processing of the analysis apparatus 100. To do. First, the process data Po and Pi acquired by the modeling unit 120 are expressed by the following mathematical formulas 2 and 3.

Po = [p1o, p2o,..., Pmo] (Expression 2)
Pi = [p1i, p2i,..., Pmi] (i = 1, 2,..., N)
... (Formula 3)

  In order to calculate the independent component vector So using the process data Po and Pi, the modeling unit 120 first calculates the Euclidean distance Li between the process data Po and each Pi. The Euclidean distance Li is expressed by Equation 4 below.

  Li = | Po−Pi | (i = 1, 2,..., N) (Formula 4)

  When the Euclidean distance Li between the process data Po and each Pi is calculated, the modeling unit 120 arranges them in ascending order of the distance, and selects one process data Pj that minimizes the Euclidean distance. Then, the independent component vector Sj associated with the process data Pj is acquired from the data M stored in the modeling data storage unit 150, and the independent component vector Sj is adopted as So. In this way, the modeling unit 120 can predict the independent component vector So.

  In this example, the independent component vector Sj associated with the process data Pj that minimizes the Euclidean distance Li between the process data Po and Pi is the independent component vector So. However, the independent component vector So is obtained by other methods. Can also get. For example, from the calculated Euclidean distance Li between the process data Po and each Pi, d pieces of process data are selected as similar process data in ascending order of distance, and are designated as Pr1 to Prd. At this time, the Euclidean distance between each of the similar process data Pr1 to Prd and the process data Po is Lr1 to Lrd. That is, the Euclidean distances Lr1 to Lrd are expressed by the following formula 5.

  Lrk = | Po−Prk | (k = 1, 2,..., D) (Formula 5)

  In addition, among the data M stored in the modeling data storage unit 150, independent component vectors associated with similar process data Pr1 to Prd are denoted as Sr1 to Srd. The independent component vectors Sr1 to Srd are expressed by the following mathematical formula 6.

  (Prk, Srk) εM (k = 1, 2,..., D) (Formula 6)

  Then, the modeling unit 120 takes the weighted average of the independent component vectors Sr1 to Srd and sets it as the independent component vector So. That is, the independent component vector So is expressed by the following formula 7. The weights w1 to wd are reciprocals of the Euclidean distances Lr1 to Lrd.

So = (w1 × Sr1 + w2 × Sr2 +... + Wd × Srd)
/(W1+w2+...+wd)
wk = 1 / Lrk (k = 1, 2,..., d) (Expression 7)

  Thus, the modeling unit 120 can predict the independent component vector So using the process data Po and Pi. That is, the modeling unit 120 also functions as a prediction unit that predicts the independent component vector So. The modeling unit 120 outputs the predicted independent component vector So to the model unit 130.

(3) Model unit 130
The model unit 130 has the shape of the tail end of the rolling material 12 generated in the rolling process 10 based on the independent component vector So predicted by the modeling unit 120 and a base image stored in the base image DB 140 described later. It is a prediction part which produces | generates a prediction estimated image. The model unit 130 calculates a signal vector Xo from the independent component vector So and a mixing matrix in which each column represents a base image based on the following Equation 8, and restores the signal vector Xo to generate a predicted estimated image. At that time, the generation of the prediction estimated image from the signal vector Xo is performed by the reverse process of the column vector conversion of FIG. That is, a 20 × 40 luminance matrix obtained by dividing the column vector Xo into 40 1 × 20 vectors and arranging the 40 column vectors in the column direction is a predicted estimated image.

  Xo = A · So (Formula 8)

  The predicted estimated image is notified to the operator, for example. The operator determines the system state from the shape of the tail end of the rolled material 12 appearing in the predicted estimated image, and corrects the set value as necessary. The control computer 60 calculates a new control condition (setting value) based on the corrected setting value, and outputs it to the modeling unit 120 of the analyzer 100. The analysis apparatus 100 performs the above-described processing, and predicts So using the independent component vector from the process data acquired in the operation with the new set value.

(4) Base image DB 140
The base image DB 140 is a storage unit that stores, as a base image, a singular shape with strong independence at the tail end of the rolled material 12 as shown in FIGS. 10A to 10E. Based on the base image stored in the base image DB 140, the model unit 130 generates a prediction estimated image. The base image stored in the base image DB 140 is composed of each column of the mixing matrix A obtained by the image feature amount analysis unit 110 using ICA.

(5) Modeling data storage unit 150
The modeling data storage unit 150 stores the independent component vector Si calculated by the image feature quantity analysis unit 110 and the process data Pi stored in the process data DB 22 of the data collection server 20. The modeling data storage unit 150 associates the independent component vector Si related to the same rolled material 12 with the process data Pi corresponding thereto, and data M = {(Pi, Si)} (i = 1, 2,..., n).

  Heretofore, the configuration of the analysis apparatus 100 according to the present embodiment and the analysis processing performed thereby have been described. The analysis apparatus 100 according to the present embodiment acquires an image feature amount from the tail end local image of the rolling material 12 and acquires a data feature amount from the process data. The analysis apparatus 100 estimates the shape of the tail end of the rolling material 12 generated in the rolling process 10 based on the correlation between the image feature amount and the data feature amount, and thereby the quality of the generated rolling material 12 and It is possible to estimate the equipment state of the rolling process 10 and appropriately set various conditions in operation.

  When the analysis processing by the analysis apparatus 100 according to the present embodiment is compared with Patent Document 1, the Patent Document 1 uses a neural network that inputs a past image and a past operation amount and outputs a predicted image. Modeling techniques are described. In the method described in Patent Document 1, it is estimated that the number of input / output points of the neural network becomes enormous and the time required for learning and prediction becomes longer. In order to solve this, for example, as described in Patent Document 2, quantification is performed by preprocessing such as the ratio of the area of the specific part to the area of the entire image, the length of the outline of the specific part, etc. It is conceivable that the obtained image feature amount is used as an input / output of the neural network model. However, the effectiveness of the preprocessing algorithm is limited because it depends on human knowledge and judgment. Moreover, it is difficult to confirm the system of the neural network model acquired as a result of learning.

  On the other hand, the method using independent component analysis performed by the analysis apparatus 100 according to the present embodiment is a matrix operation algorithm, so that high-speed calculation is possible. In addition, since modeling uses a method that has been proven in online process control such as a Just-In-Time model in which process data and contribution rate (that is, independent component vectors) are independent variables and dependent variables, processing speed is a problem. It will never be. Furthermore, by the analysis of the analysis apparatus 100, it is possible to acquire a plurality of base images that can be said to be basic image components common to the sample images, and the contribution rate to the base images that each sample image has. Therefore, it is an image feature amount by confirming that the acquired base image represents a unique process state or operation state, such as “normal state image” or “abnormal state image” in actual operation. The contribution rate can be given a clear meaning, and the reliability of the model can be secured.

(Application of multiple regression modeling)
In the above description, the modeling unit 120 identifies a model using a so-called Just-In-Time model, but the present invention is not limited to this example. For example, the modeling unit 120 may identify the multiple regression model using the least square method or the like based on the image feature amount and each result data of the process data. Hereinafter, a modeling method using a multiple regression model will be described with reference to FIGS. 14 and 15. FIG. 14 is a block diagram illustrating a configuration of an analysis apparatus 300 that performs modeling using a multiple regression model. FIG. 15 is an explanatory diagram illustrating a configuration example of the modeling data storage unit 350. Detailed description of components having the same configuration and function as those of the analyzer 100 shown in FIG. 3 is omitted.

  As shown in FIG. 14, the analysis apparatus 300 that performs modeling using a multiple regression model includes an image feature amount analysis unit 310, a modeling unit 320, a model unit 330, a base image DB 340, and a modeling data storage unit 350. Consists of.

(1) Image feature amount analysis unit 310
The image feature amount analysis unit 310 acquires the tail end image of the rolling material 12 from the moving image data stored in the moving image DB 42 of the moving image collection server 40, and extracts the image feature amount of the tail end image. The image feature quantity analysis unit 310 has the same configuration and function as the image feature quantity analysis unit 110 shown in FIG. The image feature amount Si extracted by the image feature amount analysis unit 310 is stored in the modeling data storage unit 350.

(2) Modeling unit 320 and modeling data storage unit 350
The modeling unit 320 identifies a multiple regression model based on the actual data of the image feature amount Si and the process data Pi stored in the modeling data storage unit 350. Furthermore, the independent component vector So is calculated from the control condition (process data Po) input from the control computer 60 using the identified multiple regression model. In the modeling data storage unit 350, the image feature amount Si input from the image feature amount analysis unit 310 and each result data of the process data are stored as shown in FIG. 15, for example.

  In FIG. 15, a coil number 351 for specifying the rolling material 12, an image feature amount 352, and process data 353 are stored in association with each rolling material 12. The image feature quantity 352 stores each contribution rate to the u number of base images, and the process data 353 stores m pieces of record data such as the load and leveling of each rolling material 12. It is assumed that the modeling data storage unit 350 stores data for n rolling materials 12.

  The modeling unit 320 uses the multiple regression model identified based on the actual data of the image feature amount Si and the process data Pi stored in the modeling data storage unit 350 to control conditions (processes) input from the control computer 60. The contribution ratio of the image feature quantity, that is, the independent component vector So is calculated from the data Po). At this time, the independent component vector So can be expressed by Equation 9 below. Further, the process data Po is expressed by Equation 2 above. The modeling unit 320 outputs the calculated independent component vector So to the model unit 330.

So1 = A01 + A11 × p1o + A21 × p2o +... + Am1 × pmo
So2 = A02 + A12 × p1o + A22 × p2o +... + Am2 × pmo
:
Sou = A0u + A1u × p1o + A2u × p2o +... + Amu × pmo
... (Formula 9)

(3) Model unit 330 and base image DB 340
Based on the independent component vector So predicted by the modeling unit 220 and the base image stored in the base image DB 440, the model unit 330 predicts and estimates the shape of the tail end of the rolling material 12 generated in the rolling process 10. Is generated. In the base image DB 440, for example, a base image representing a singular shape having a strong independence of the tail end portion of the rolled material 12 as illustrated in FIGS. 10A to 10E is stored.

  The model unit 330 uses the independent component vector So and the base image to generate a predicted estimated image by the same process as the model unit 130 described above. The predicted estimated image is used, for example, as a material for notifying the operator and determining the system state from the shape of the tail end of the rolled material 12 appearing in the predicted estimated image. The operator corrects the setting value as necessary, and the control computer 60 calculates a new control condition (setting value) based on the corrected setting value, and outputs it to the model unit 330 of the analyzer 300. The analysis apparatus 300 performs the above-described processing, and predicts So using the independent component vector from the process data acquired in the operation with the new set value.

  In this way, a predicted estimated image of the shape of the tail end of the rolled material 12 can also be obtained by a modeling method using a multiple regression model.

<3. Example>
Hereinafter, the correlation between the image feature amount of the tail end local image of the rolled material 12 acquired by the analysis apparatus 100 and the data feature amount of the process data will be described according to examples.

  In Example 1, the correlation between the image feature amount of the tail end local image and the data feature amount of the process data will be described for the shapes of the tail ends of the two rolled materials 12 shown in FIG. Here, the case where the shape of the tail end is sharp on the WS side (Analysis 1) and the case where the tail end is sharp on the DS side (Analysis 2) will be considered.

(Analysis 1: WS sharp)
First, a case where the shape of the tail end has a sharp point on the WS side will be considered based on FIGS. When the shape of the tail end is sharp on the WS side, in the tail end local image, the WS side (left side on the paper surface) is brighter than the DS side (right side on the paper surface) as shown in the upper diagram of FIG. Also, it can be seen that many locations with high luminance values are distributed on the WS side (left side of the drawing). Thus, ten rolling materials having leveling characteristics (see paragraph 0054) that are close to the representative image and the Euclidean distance are extracted with the representative image as an image in which the shape of the tail with a sharp point on the WS side appears. . The extracted rolled material is referred to as a similar leveling material. As shown in FIG. 17, it is understood that the leveling values of the similar leveling materials have values close to each other in each rolling stand.

  FIG. 18 shows a part of the luminance matrix generated from the tail end local image of the similar leveling material. When the six luminance matrices shown in FIG. 18 are visually compared with the luminance matrix of the representative image shown in the upper diagram of FIG. 16, the tail end tends to be sharp on the WS side similar to the luminance matrix of the representative image. I understand that. In addition, as shown in FIG. 19, the independent component vector indicating the image feature amount of the similar leveling material has a partly different tendency, but the contribution ratio of the WS-pointed base image is high, and the base of the DS-pointed base is high. It can be seen that the contribution ratio to the image tends to be low.

(Analysis 2: DS sharpness)
Next, based on FIGS. 16 and 20 to 22, a case where the shape of the tail end is sharp on the DS side will be considered. When the shape of the tail edge is sharp on the DS side, in the tail edge local image, the DS side (right side on the paper surface) is brighter than the WS side (left side on the paper surface) as shown in the lower diagram of FIG. It can be seen that many locations with high luminance values are distributed on the DS side (right side of the drawing). In this way, a representative material is an image in which the shape of the tail with a sharp point on the DS side appears, and a rolling material having a leveling characteristic close to the representative image and the Euclidean distance (similar leveling material) as in Analysis 1. 10 are extracted. As shown in FIG. 20, it is understood that the leveling characteristics of the similar leveling materials have close values in the respective rolling stands.

  A part of the luminance matrix generated from the tail end local image of the similar leveling material is shown in FIG. When the six luminance matrices shown in FIG. 21 are visually compared with the luminance matrix of the representative image shown in the lower diagram of FIG. 16, the tail end tends to be sharp on the DS side similar to the luminance matrix of the representative image. I understand. In addition, as shown in FIG. 22, the independent component vector indicating the image feature amount of the similar leveling material has some different tendencies, but the contribution rate of the DS sharp base to the base image is high, and the WS sharp base is high. It can be seen that the contribution ratio to the image tends to be low.

  From analyzes 1 and 2, it can be said that the rolled material having similar leveling characteristics has similar characteristics of the independent component vectors, and there is a correlation between the independent component vectors and the leveling characteristics.

  Next, as a second embodiment, based on FIGS. 23 to 26, image feature amounts, that is, tail local images having a close Euclidean distance of independent component vectors are classified, and process data of the classified tail local images are classified. Verify relevance. In Example 2, based on the Euclidean distance of the independent component vector, the tail end local image is an image having a WS sharp tendency (Category 1: FIG. 23), an image having a DS sharp tendency (Category 2: FIG. 24), and a fishtail. The images were classified into four types: images having a shape tendency (Category 3: FIG. 25) and images having a normal shape (Category 4: FIG. 26). And the relationship was verified about the value of a differential load and leveling among the process data of the classified rolling material 12. FIG. In the present embodiment, the relevance of process data is verified using three tail end local images for each classification.

(Category 1: WS sharp)
First, the relevance of the process data of the tail end local image classified as an image having a WS sharp tendency is verified. As shown in FIG. 23, the shape of the tail end appearing in the local image of the tail end classified as an image having a WS sharpness tendency has a sharpness on the WS side (left side of the paper) and is visually similar. ing. On the other hand, looking at the process data of the rolling material 12 corresponding to these tail end local images, it can be seen that the values at the respective rolling stands are similar for both the differential load and the leveling.

(Category 2: DS sharpness)
Then, the relevance of the process data of the tail-tip local image classified as an image having a DS sharpness tendency is verified. As shown in FIG. 24, the shape of the tail end that appears in the local image of the tail end classified as an image having a DS sharpness tendency has a sharpness on the DS side (right side of the paper) and is visually similar. ing. On the other hand, looking at the process data of the rolling material 12 corresponding to these local images at the tail end, it can be seen that the values at the respective rolling stands are similar for both the differential load and the leveling as in the case of the classification 1.

(Category 3: Fishtail)
Furthermore, the relevance of the process data of the tail end local image classified as an image having a fishtail shape tendency is verified. As shown in FIG. 25, the shape of the tail end that appears in the local image of the tail end classified as an image having a fishtail shape tendency has a sharp point on both the WS side and the DS side, and is visually similar. doing. On the other hand, regarding the process data of the rolling material 12 corresponding to these tail end local images, the value at each rolling stand is similar for the differential load, but the value of each process data is different depending on the rolling stand for leveling. It can be seen that there is a difference.

(Category 4: Normal)
Finally, the relevance of the process data of the tail end local image classified as an image having a normal shape is verified. As shown in FIG. 26, the shape of the tail end appearing in the tail end local image classified as an image having a normal shape is both tongue-like and visually similar. On the other hand, when looking at the process data of the rolling material 12 corresponding to these tail end local images, it can be seen that the values at the respective rolling stands are similar for both the differential load and the leveling as in the case of classifications 1 and 2.

  From the verification of the above classifications 1 to 4, it can be seen that the process data of the rolled material having similar image feature amounts tend to be generally similar. However, even if the image feature amounts are similar as in Category 3, some of the process data may have weak data feature amounts. As this cause, there are a plurality of factors in the case where the shape of the tail end of the rolled material is sharp, and it is considered that the tendency does not easily appear in one process data. Moreover, the influence by the dispersion | variation in the quality of the tail end local image which extracts image feature-value is also considered.

  Such a cause may be improved by, for example, a method for calculating a representative value of process data. For example, a representative value is calculated using a value at a rolling stand near the end of the rolling stands, or a representative value is calculated using a difference between adjacent stands in the rolling direction of the rolling material. Etc. are considered.

  From Examples 1 and 2, the correlation between the image feature amount extracted from the tail end local image and the data feature amount extracted from the process data can be confirmed. Therefore, according to the analysis apparatus 100 according to the present embodiment, it is possible to analyze whether the rolling process 10 is operating normally, whether there is an abnormality, or the like from the independent component vector predicted based on such correlation. . Moreover, the tail end local image (target image) of the target coil is set, and the operator corrects the control condition from the difference between the predicted estimated image estimated by the model unit 130 and the target image, and the rolling process 10 It is possible to maintain the quality of the produced rolling material by reflecting it in the control conditions.

<4. Hardware configuration example>
The analysis apparatus 100 according to the present embodiment described above can be realized by an information processing apparatus such as a computer. The information processing apparatus includes, for example, a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), a bus, an interface, an input device, an output device, a storage device, a communication device, and the like. And a device.

  The CPU functions as an arithmetic processing device and a control device, and controls the overall operation in the analysis device 100 according to various programs. The CPU may be a microprocessor. The ROM stores programs used by the CPU, calculation parameters, and the like. The RAM temporarily stores programs used in the execution of the CPU, parameters that change as appropriate during the execution, and the like. These are connected to each other by a bus. The bus is connected to an interface that interconnects the input device, the output device, the storage device, and the communication device.

  The input device includes input means for a user to input information such as a mouse, keyboard, touch panel, button, microphone, switch, and lever, and an input control circuit that generates an input signal based on the input by the user and outputs the input signal to the CPU. It is composed of

  The output device includes display devices such as a CRT (Cathode Ray Tube) display device, a liquid crystal display (LCD) device, an OLED (Organic Light Emitting Diode) device, and a lamp. Furthermore, the output device includes a sound output device such as a speaker and headphones.

  The storage device is a data storage device configured as an example of a storage unit of the analysis device 100, and includes, for example, an HDD (Hard Disk Drive). The storage device drives a hard disk, for example, and stores programs executed by the CPU and various data. The communication device is a communication interface configured by a communication device or the like for connecting to a communication network, for example. As the communication device, for example, a wireless LAN (Local Area Network) compatible communication device, a wireless USB compatible communication device, a communication device that performs wired communication, or the like can be used.

  In addition, this information processing apparatus can also be provided with the reader / writer for storage media incorporated in the information processing apparatus or attached externally, for example. The storage medium reader / writer reads information recorded on a mounted removable recording medium such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, and outputs the information to the RAM. The information processing apparatus can also be provided with an interface connected to an external device, and can be used as a connection port with an external device capable of transmitting data by, for example, USB (Universal Serial Bus).

  Further, the analysis method by the analysis apparatus 100 according to the present embodiment described above may be executed by dedicated hardware, but may be executed by software. When a series of processing is performed by software, for example, the above-described series of processing can be realized by causing a general-purpose or dedicated computer to execute the program.

  The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to such examples. It is obvious that a person having ordinary knowledge in the technical field to which the present invention pertains can come up with various changes or modifications within the scope of the technical idea described in the claims. Of course, it is understood that these also belong to the technical scope of the present invention.

  For example, in the above embodiment, the analysis apparatus 100 is applied to the analysis of the rolling process 10, but the present invention is not limited to such an example. For example, the present invention may be applied to a process control field other than the rolling process 10.

  Moreover, in the said embodiment, although the image feature-value was calculated by independent component analysis (ICA), this invention is not limited to this example. For example, the image feature amount may be calculated by principal component analysis (PCA) or the like.

DESCRIPTION OF SYMBOLS 10 Rolling process 12 Coil 14 Rolling roll 20 Data collection server 22 Process data DB
31, 32 Camera 33, 34 Monitor 40 Video collection server 42 Video DB
DESCRIPTION OF SYMBOLS 100 Analysis apparatus 110 Analysis part 120 Modeling part 122 Operation analysis DB
124 condition input unit 126 model generation unit 130 estimation unit 132 addition unit 140 base image DB

Claims (8)

In the process of generating an analysis object, an analysis device for analyzing the analysis object generated in the process,
A first feature amount extraction unit that extracts an image feature amount representing a shape feature appearing in the image from the analysis target image of the analysis target;
A second feature amount extraction unit for extracting a data feature amount representing a characteristic on data appearing in the process data from the process data in the process of generating the analysis object;
Analytical object generated by the process based on the correlation between the image feature quantity extracted by the first feature quantity extraction unit and the data feature quantity extracted by the second feature quantity extraction unit A prediction unit that calculates a predicted image feature amount predicted for
Equipped with a,
The image feature amount is a contribution rate to the base image representing a characteristic shape with respect to the analysis target image of the analysis target,
The first feature amount extraction unit obtains a contribution rate of the analysis object to the base image with respect to the analysis image using independent component analysis,
The prediction unit performs the process based on the correlation between the contribution rate of the base image acquired by the first feature amount extraction unit and the data feature amount extracted by the second feature amount extraction unit. An analysis device that calculates a contribution rate to a base image predicted for an analysis object generated in this manner . The prediction unit
Calculating the Euclidean distance between each of the data feature and each process data in the process of generating the analysis object,
The analysis apparatus according to claim 1, wherein the image feature amount associated with the process data that minimizes each Euclidean distance is a predicted image feature amount of the analysis target. The prediction unit
Calculating the Euclidean distance between each of the data feature and each process data in the process of generating the analysis object,
The plurality of process data in which each Euclidean distance is equal to or less than a predetermined value is extracted as similar process data, and a weighted average of the image feature amounts associated with the similar process data is used as a predicted image feature amount of the analysis object. The analyzer according to claim 1, wherein: The second feature amount extraction unit normalizes a value of at least one item data among the item data included in the process data in the process of generating the analysis target object to obtain the data feature amount. The analyzer according to any one of claims 1 to 3 , wherein the analyzer is characterized. The predicted image feature amount calculated by the prediction unit, based on said base image, further comprising an estimation unit that estimates a shape of the analyte and wherein any one of claims 1 to 4 The analyzer described in 1. The analysis object is a rolling material in a rolling process,
Characterized by a tail end image representing the shape of the tail end of the rolled material and the analysis target image, the analyzer according to any one of claims 1-5. In the process of generating an analysis object, an analysis method for analyzing the analysis object generated in the process,
A first feature amount extraction step of extracting an image feature amount representing a shape feature appearing in the image from the analysis target image of the analysis target;
A second feature amount extraction step of extracting a data feature amount representing a characteristic on data appearing in the process data from the process data in the process of generating the analysis object;
A prediction step of calculating a predicted image feature amount predicted for the analysis target generated in the process based on a correlation between the image feature amount and the data feature amount;
Only including,
The image feature amount is a contribution rate to the base image representing a characteristic shape with respect to the analysis target image of the analysis target,
The first feature amount extraction step obtains a contribution rate of the analysis object to the base image with respect to the analysis image using independent component analysis,
The prediction step performs the process based on the correlation between the contribution ratio of the base image acquired by the first feature amount extraction step and the data feature amount extracted by the second feature amount extraction step. An analysis method characterized by calculating a contribution rate to a base image predicted for an analysis object generated in this way. Computer
A first feature amount extraction unit that extracts an image feature amount representing a shape feature appearing in the image from the analysis target image of the analysis target;
A second feature amount extraction unit for extracting a data feature amount representing a characteristic on data appearing in the process data from the process data in the process of generating the analysis object;
Generated by the process of generating the analysis object based on the correlation between the image feature quantity extracted by the first feature quantity extraction unit and the data feature quantity extracted by the second feature quantity extraction unit A prediction unit that calculates a predicted image feature amount predicted for the analysis target to be analyzed;
Equipped with a,
The image feature amount is a contribution rate to the base image representing a characteristic shape with respect to the analysis target image of the analysis target,
The first feature amount extraction unit obtains a contribution rate of the analysis object to the base image with respect to the analysis image using independent component analysis,
The prediction unit performs the process based on the correlation between the contribution rate of the base image acquired by the first feature amount extraction unit and the data feature amount extracted by the second feature amount extraction unit. In order to function as an analysis device for analyzing an analysis object generated in the process of generating the analysis object, which calculates a contribution rate to the base image predicted for the analysis object generated Computer program.