Quality Abnormality Cause Analysis Support System

Abstract

In the present invention, a data storage unit stores a manufacturing condition for a product, a setting condition for facility equipment, and, as actual results data acquired from measurement equipment, time-series data for the time of manufacturing. A quality abnormality factor estimation calculation unit acquires a quality abnormality site of an abnormal product of which a quality evaluation value calculated from quality data exceeds an allowable range. The quality abnormality factor estimation calculation unit constructs a quality abnormality factor estimation model on the basis of the time series data as relates to a normal product group or on the basis of the manufacturing condition and setting condition, estimates a factor candidate for a quality abnormality on the basis of the constructed quality abnormality factor estimation model, and calculates the degree of relevance of the estimated factor candidate to the quality abnormality. A display information generation unit generates information for displaying, on a display unit, the quality data of the abnormal product, the quality abnormality site, the time series data or the manufacturing condition and setting condition, and the factor candidate and degree of relevance estimated and calculated by the quality abnormality factor estimation calculation unit.

Description

Technical Field

[0001] The present disclosure relates to a quality abnormality factor analysis support system. In particular, the present disclosure relates to a quality abnormality factor analysis support system that supports the analysis work of the factors of quality abnormality when a product such as a steel plate manufactured in a rolling plant is determined to have a quality abnormality.

Background Art

[0002] When a product manufactured in a rolling plant does not satisfy the quality control standard value determined by customer requirements or the like, it is determined to have a quality abnormality (including quality defects). Patent Document 1 below discloses a manufacturing process abnormality determination device.

[0003] A product determined to have a quality abnormality is sold as a lower-grade material or discarded, resulting in a decrease in production efficiency. When a quality abnormality occurs, it is necessary to determine its characteristics, analyze and investigate the causes, and determine and implement measures such as changes in operating conditions and control settings to prevent recurrence.

[0004] A series of operations from determining the characteristics of a quality abnormality to implementing measures requires comprehensive judgment from a large amount of data and information by experienced and skilled personnel, so a great deal of labor and time are required. Patent Document 2 below discloses a quality abnormality cause estimation system.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0006] In Patent Document 2, it is possible to find a group of products with similar conditions of various input information during manufacturing, and to detect quality anomalies in advance based on whether or not they match the trend of the output results of the product group. Further, based on the difference between the representative value of the output results of the product group with similar input information conditions and the prediction result calculated based on the input information of the target product, the accuracy of the prediction calculation for the target product can be improved.

[0007] However, Patent Document 2 focuses on detecting quality anomalies and analyzing their causes with respect to the manufacturing conditions and setting conditions of individual products. For this reason, Patent Document 2 was unable to capture even the position in the longitudinal direction of the product where a quality anomaly occurred. This is because the product is long in the longitudinal direction and the amount of information in the time-series data (actual data) is too large, so it is common to compress the time-series data to detect quality anomalies. For this reason, when a quality anomaly occurs, it is necessary to identify the location where the quality anomaly occurred and then analyze the cause of the occurrence at the identified location.

[0008] The present disclosure has been made to solve the above-described problems. An object of the present disclosure is to provide a quality anomaly cause analysis support system capable of displaying the position in the longitudinal direction of a product where a quality anomaly has occurred and the cause thereof when the product is determined to have a quality anomaly.

Means for Solving the Problems

[0009] A first aspect relates to a quality anomaly cause analysis support system. The quality anomaly cause analysis support system includes a data storage unit, a quality anomaly cause estimation calculation unit, and a display information generation unit. The data storage unit stores the manufacturing conditions of products manufactured in a rolling mill, the setting conditions of the equipment installed in the rolling mill, and the time-series data during manufacturing as actual data acquired from the measuring instruments installed in the rolling mill. The quality anomaly cause estimation calculation unit calculates a quality evaluation value from the quality data related to the product quality of the product selected from the time-series data during manufacturing Perform. The quality abnormality factor estimation calculation unit Identify abnormal products or product groups whose calculated quality evaluation values exceed the allowable range as abnormal products, and obtain quality abnormality locations that exceed the allowable range of the abnormal products. The quality abnormality factor estimation calculation unit uses time-series data related to normal product groups whose quality evaluation values are within the allowable range, Registers the manufacturing conditions and setting conditions as normal data. The quality abnormality factor estimation calculation unit uses the time series data registered as normal data, or Registered as normal data Based on manufacturing conditions and setting conditions , and utilizes machine learning or statistical methods Construct a quality abnormality factor estimation model Perform. The quality abnormality factor estimation calculation unit Based on the constructed quality abnormality factor estimation model, estimate candidate factors for quality abnormalities from time-series data, or from manufacturing conditions and setting conditions Perform. The quality abnormality factor estimation calculation unit Calculate the degree of relevance of the estimated candidate factors to the quality abnormality. The display information generation unit generates information for displaying the quality data of the abnormal product, the quality abnormality location, the time-series data, or the manufacturing conditions and setting conditions, and the candidate factors and degrees of relevance estimated and calculated by the quality abnormality factor estimation calculation unit on the display unit.

[0010] The second perspective further has the following features in addition to the first perspective. The display information generation unit generates information for displaying a list of the manufacturing conditions, setting conditions, quality evaluation values, and quality abnormality locations of the abnormal products.

[0011] The third perspective further has the following features in addition to the first perspective. The display information generation unit generates information for displaying a trend chart of the quality data of the abnormal product, a trend chart of the time-series data corresponding to the candidate factor with the highest degree of relevance among the candidate factors estimated by the quality abnormality factor estimation calculation unit, and the quality abnormality locations in each trend chart.

[0012] The fourth perspective further has the following features in addition to the first perspective. The display information generation unit generates information for displaying a trend chart of the quality data of the abnormal product, a plurality of candidate factors estimated by the quality abnormality factor estimation calculation unit, the degree of relevance of each candidate factor, a trend chart of the time-series data corresponding to each candidate factor, and the quality abnormality locations in each trend chart.

[0013] In addition to the first aspect, the fifth aspect further has the following features. The display information generation unit generates information for displaying, for the same abnormal product with quality data and quality abnormal points, the setting conditions as factor candidates and the degree of relevance of each setting condition to the quality abnormality.

[0015] The 6 aspect further has the following features in addition to the first Any one from the first to the fifth aspect. The quality abnormality factor estimation calculation unit reconstructs the quality abnormality factor estimation model every time a product is manufactured, and updates the quality abnormality factor estimation model when the prediction error of the reconstructed quality abnormality factor estimation model is smaller than the prediction error of the quality abnormality factor estimation model before reconstruction.

Advantages of the Invention

[0016] According to the present disclosure, by obtaining the quality abnormal points where the quality evaluation value of the abnormal product exceeds the allowable range, it is possible to display at which position in the longitudinal direction of the product the quality abnormality has occurred. Furthermore, by constructing a quality abnormality factor estimation model using normal data, estimating factor candidates for the quality abnormality using the constructed quality abnormality factor estimation model, and calculating the degree of relevance of the estimated factor candidates to the quality abnormality, it is possible to display the time-series data with a high degree of relevance as the cause of the quality abnormality. Therefore, when a product is determined to have a quality abnormality, it is possible to provide a quality abnormality factor analysis support system that can display at which position in the longitudinal direction of the product the quality abnormality has occurred and display the cause thereof.

Brief Description of the Drawings

[0017]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Embodiments for Carrying Out the Invention

[0018] Hereinafter, with reference to the drawings, a quality abnormality factor analysis support system according to an embodiment of the present disclosure will be described. In addition, the same reference numerals are given to common elements in each figure, and redundant explanations are omitted.

[0019] Embodiment 1. FIG. 1 is a schematic diagram showing an example of a rolling mill 1 to which a quality abnormality factor analysis support system 20 according to Embodiment 1 is applied.

[0020] The rolling mill 1 is, for example, a hot rolling mill having a hot rolling line, but it may also be a cold rolling mill. The rolling mill 1 is installed with a heating furnace 2, a roughing mill 3, a crop shear 4, a finishing mill 5, a cooling device 6, and a coiler 7 as the main equipment constituting the hot rolling line. A conveying table (not shown) for conveying the steel plate Pr is installed between the rolling facilities in the rolling mill 1. These rolling facilities are driven by an electrical system of an electric motor or an actuator.

[0021] The heating furnace 2 is configured to heat the steel plate (slab) Pr before rolling to a predetermined temperature (for example, 1200 °C). The roughing mill 3 has at least one (usually one to three) rolling stands, and rolls the steel plate (slab) Pr heated in the heating furnace 2 in multiple passes in the forward direction (from the upstream to the downstream of the rolling line) and the reverse direction (from the downstream to the upstream of the rolling line). The crop shear 4 cuts off the defective shape portions existing at the tip or tail end of the steel plate Pr with the upper and lower blades based on the shape measured by a shape detector 81 described later. The finishing mill 5 is a tandem rolling mill having, for example, seven rolling stands F1 to F7 arranged in parallel in the rolling direction of the steel plate Pr. Each of the rolling stands F1 to F7 is provided with two upper and lower work rolls 51, two upper and lower backup rolls 52, and an electric motor 53 for roll rotation. A rolling reduction device 54 is provided on the backup roll 52, and the gap between the upper and lower work rolls 51 can be adjusted by the rolling reduction device 54. The rolling load of each of the rolling stands F1 to F7 is measured by a rolling load sensor 55. The cooling device 6 cools the steel plate Pr by injecting water into the steel plate Pr with a cooling bank. The cooled steel plate Pr is wound into a coil by the coiler 7.

[0022] Various sensors as measuring devices are installed at key points of the rolling plant 1. The key points of the rolling plant 1 are, for example, the outlet side of the heating furnace 2, the outlet side of the rough rolling mill 3, the outlet side of the finishing rolling mill 5, and the inlet side of the coiler 7, etc. The various sensors may also be provided between the rolling stands F1 to F7 of the finishing rolling mill 5. The various sensors include a shape detector 81 capable of measuring the shape of the steel plate Pr on the outlet side of the rough rolling mill 3, a thermometer 82 for measuring the surface temperature of the steel plate Pr on the inlet side of the finishing rolling mill 5, a speed detector 83 for measuring the speed Va of the steel plate Pr on the outlet side of the finishing rolling mill 5, a thickness-width gauge 84 for measuring the plate thickness and width of the steel plate Pr on the outlet side of the finishing rolling mill 5, and a thermometer 85 for measuring the surface temperature of the steel plate Pr on the inlet side of the coiler 7, and include the above-mentioned rolling load sensor 55. The various sensors sequentially measure the state of the steel plate Pr and each equipment. The performance data measured by the various sensors is transmitted to the control computer 11 at all times. Therefore, the performance data is time-series data. Width

[0023] ​The rolling plant 1 is operated (operated) by a control system using a computer having a hierarchical structure. The computer includes a process control computer (hereinafter referred to as "control computer") 11 and a host computer 12 that are connected to each other via a network. The control computer 11 is connected to a quality abnormality factor analysis support system 20 described later via a network. The control computer 11 has a controller for control such as a PLC (Programmable Logic Controller). An HMI (Human Machine Interface) device 13 is connected to the control computer 11 via a network. The HMI 13 presents data of the monitoring target (rolling equipment) so that the user (including the administrator) can monitor or operate (control) the monitoring target device. When hot rolling command information, which is a rolling plan, is input to the host computer 12 by the designer of the rolling plan, the hot rolling command information is sent from the host computer 12 to the control computer 11. The hot rolling command information includes a target sheet thickness, a target sheet width, a target temperature, and the like. The target temperature includes the target temperature on the outlet side of the finishing rolling mill 5 (hereinafter referred to as "finishing outlet temperature") and the target temperature on the inlet side of the coiler 7 (hereinafter referred to as "coiling temperature"). The process control computer 11 receives the input of the hot rolling command information (rolling plan), which is the manufacturing condition from the host computer 12, and calculates the set Fixed data, and the calculated set Fixed data is transmitted to the rolling plant 1 to execute the control of various rolling equipment.

[0024] The quality abnormality factor analysis support system 20 collects data (rolling data) exchanged between each rolling equipment of the rolling plant line 1 and the control computer 12, calculates information for supporting the quality abnormality factor analysis using the collected rolling data, and provides it to the user (including the administrator).

[0025] The quality abnormality factor analysis support system 20 includes a data storage unit 21, a quality abnormality factor estimation calculation unit 22, and a display information generation unit 23. The functions of each of the units 21 to 23 of the quality abnormality factor analysis support system 20 can be realized, for example, by the processor 20b shown in FIG. 12 described later reading and executing a program stored in the memory 20c.

[0026] The data storage unit 21 has a function of collecting rolling data and storing it in the database DB. The rolling data includes the above-mentioned hot rolling instruction information, setting data, and performance data. The quality abnormality factor estimation calculation unit 22 performs calculation (computation) of information for supporting the evaluation of product quality and the analysis of quality abnormality factors based on the evaluation results, using the rolling data stored in the data storage unit 21. The display information generation unit 23 causes the calculation results of the quality abnormality factor estimation calculation unit 22 to be displayed on the HMI device 13 as a display unit. The operator confirms and operates on the information (calculation results) displayed on the HMI device 13, and performs analysis work.

[0027] FIG. 2 is a diagram showing the configuration of data collected and registered by the data storage unit 21. As shown in FIG. 2, the information collected by the data storage unit 21 includes information assigned to each of the steel plates Pr, such as hot rolling instruction information and setting data, and time-series data obtained every moment for the rolling of the steel plate Pr, such as performance data. These information and data are linked to the manufacturing number of the steel plate Pr (hereinafter referred to as "coil ID") and the extraction time of the steel plate Pr from the heating furnace 2, etc., and are stored as the database DB. In the example shown in FIG. 1, the database DB is provided inside the quality abnormality factor analysis support system 20, but the database DB may be provided outside the quality abnormality factor analysis support system 20, and information and data may be exchanged through network connection.

[0028] Next, the procedure for calculating information for supporting the evaluation of product quality and the analysis of quality abnormality factors based on the evaluation results in the quality abnormality factor estimation calculation unit 22 will be described with reference to FIG. 3. FIG. 3 is a diagram showing the processing flow of the quality abnormality factor analysis support system 20 according to the first embodiment.

[0029] The quality abnormality factor estimation calculation unit 22 first obtains, from among the performance data which is the time-series data described above, quality data xi used for evaluating the quality of the steel plate Pr manufactured in the rolling plant 1, such as the finished side plate thickness, the finished side plate width, the shape, the finished side temperature, and the coil inlet temperature (step S1), and calculates a quality evaluation value (step S2). Allowable ranges, which are reference values, are respectively defined for these quality data xi, and the quality of the steel plate Pr is evaluated based on this allowable range. The quality evaluation value is calculated, for example, by the following formulas (1) and (2). [Number] [Number]

[0030] Here, performance A indicates the quality evaluation value of any one of the finished side plate thickness, the finished side plate width, the shape, the finished side temperature, the coil inlet temperature, etc. of the steel plate Pr. I s indicates the start point of the object for calculating the quality evaluation value with respect to the longitudinal position of the steel plate. I L indicates the end point of the object for calculating the quality evaluation value with respect to the longitudinal position of the steel plate. The longitudinal direction of the steel plate Pr corresponds to the rolling direction or the conveying direction. Also, xi indicates the value of the quality data at point i from I s to I L inclusive. Tolerance indicates any one of the quality reference values such as the finished side plate thickness, the finished side plate width, the shape, the finished side temperature, and the coil inlet temperature. The start point and the end point of the object for calculating the quality evaluation value in the longitudinal direction of the steel plate Pr are often taken as the length excluding the non-steady deformation parts at the leading and trailing ends in the longitudinal direction of the steel plate Pr (so-called leading and trailing end cuts). However, for example, the length excluding the non-steady deformation parts at the leading and trailing ends in the longitudinal direction of the steel plate Pr may be further divided into several parts, and the quality evaluation value may be calculated for each interval.

[0031] The quality evaluation value calculated in the above step S2 is stored by the data storage unit 21 in association with the coil ID of the steel plate Pr, etc. (step S3). Next, it is determined whether the calculated quality evaluation value satisfies the allowable range as the reference value (step S4). If the quality evaluation value satisfies the allowable range, it is registered as normal data of that quality, and if it does not, it is registered as abnormal data of that quality, each in association with the coil ID of the steel plate Pr, etc. (step S5), and stored by the data storage unit 21. The label of the abnormal data of the quality does not necessarily have to be assigned as a distinction from the normal data, and it may be registered as a label according to the characteristics of the abnormality. For example, when the finish side plate thickness exceeds the allowable range, it may be registered as information such as whether it is thicker than the upper limit value, or thinner than the lower limit value, and further, how much it deviates. Also, by registering the measurement points that exceed the allowable range, that is, the position and range in the longitudinal direction of the steel plate Pr, it is possible to specify which part (position) in the longitudinal direction of the steel plate Pr exceeded the allowable range. At this time, for example, for each of the degree of deviation from the upper limit value or the lower limit value of the allowable range, the position and range of the measurement point that exceeds the allowable range, a label may be registered as an index according to a predetermined numerical range.

[0032] Whether the number of steel plates Pr labeled as normal data for the target quality has reached an arbitrary number N L has been collected, that is, whether a certain number N L has been collected in the normal data group is determined (step S6). The certain number N LWhen it exceeds, a quality abnormality factor estimation model for the target quality (hereinafter also referred to as "model") is constructed (step S7), the constructed model is stored by the data storage unit 21 (step S8), and this routine ends once. After the model is constructed, if the above-mentioned quality evaluation value exceeds the allowable range, it is determined as NO in step S4 above, determined as YES in step S9, and the process proceeds to step S10. In step S10, the abnormal factors for the target quality are estimated using the quality abnormality factor estimation model. That is, the quality abnormality factor estimation model estimates the factor candidates for quality abnormality, and calculates the degree of association between each estimated factor candidate and the quality abnormality. The factor candidates estimated to be related to the target quality abnormality and the actual data are quantified as the "degree of association", and each item of the actual data and the calculated degree of association are associated with the coil ID of the steel plate Pr and stored by the data storage unit 21.

[0033] Note that in this embodiment, the construction of the quality abnormality factor estimation model is only at the timing that meets the above-mentioned conditions. After that, as described in Embodiment 2 to be described later, it can be carried out at an arbitrary timing determined by the user (including the administrator) of the quality abnormality factor analysis support system 20.

[0034] The above-mentioned quality abnormality factor estimation model may use, for example, an auto encoder (AE), a random forest (RF), a support vector regression (SVR), etc. that fall within the category of machine learning, or a method that utilizes statistical methods. Here, AE , the algorithms of RF and SVR are widely known in general, so they will be briefly described below.

[0035] AE is a type of Neural Network model (hereinafter referred to as "NN model"). It is a method for detecting anomalies from the divergence between input variables and output variables by using an NN model that is trained only with normal data and a group of data with registered labels (hereinafter referred to as "normal data") so as to output variables similar to the input variables. In the simplest configuration, the NN model has a three-layer structure consisting of an input layer, an intermediate layer, and an output layer, and the number of intermediate layers can also be increased. By having a large number of intermediate layers, deep learning can also be achieved. Each layer is composed of one or more neurons. The neurons in each layer have weight coefficients and bias values, and they are all connected. Usually, an activation function is defined for each neuron, and it is designed such that the output state changes according to the level of the input value.

[0036] RF is a method in which weak learners with a tree structure whose output is determined by the conditions of input variables called a plurality of decision trees are connected in parallel, and the majority vote or average of each output is taken to predict the final output. SVR is a method in which the support vector machine algorithm for determining a boundary line, a hyperplane, that divides data into two groups is applied to a regression problem.

[0037] For any of these methods, it is necessary to construct a model for predicting the output variable from the input variable using the group of data prepared for learning. In the present embodiment, a quality anomaly factor estimation model is constructed using the time-series data of steel plates with registered labels as the above-mentioned normal data, or the setting data of steel plates with registered labels as the normal data.

[0038] Figure 4 shows a configuration example of input variables and output variables for a model that predicts and calculates output variables from input variables. As input variables, there are cases where setting data related to target quality data and quality evaluation values are used (cases C and D), and cases where performance data related to target quality data and quality data are used (cases A and B). A model that predicts and calculates output variables from either one of them may be constructed, or both may be constructed. When quality data and performance data related to the quality data are used (case A), the output variable is the same as the input variable. Here, in the case of RF or SVR, only the performance data related to the quality data is used as the input variable (case B), and the quality data may be used as the output variable. Also, when abnormal labels are assigned to each section obtained by dividing the length excluding the unsteady deformation part at the head and tail ends in the longitudinal direction of the steel plate Pr into several sections, the target section can be cut out and used as the input variable.

[0039] The model that predicts and calculates output variables from input variables is constructed so that the input variable and the output variable match, that is, so that the prediction error of the model that predicts and calculates output variables from input variables is minimized. When only the performance data related to the target quality data is used and the target quality data is used as the output variable, it is constructed so that the actual target quality data and the predicted output variable match. In any case, the error is defined as a mathematical formula (error function), and it is constructed (learned) so that it is minimized. This error function may be defined as follows, for example, in the case of AE.

Number

Number

[0040] Here, L is the loss function, h is the intermediate layer of the NN, g is the activation function, and W 1 is the weight coefficient of the first layer of the intermediate layer of the NN, RMSE is the root mean square error, and λ is the regularization parameter. W 1It is defined by a matrix of the number of input variables \(j = \{1, 2, \ldots, j\}\) and the number of neurons in the first hidden layer \(k=\{1, 2, \ldots, k\}\). As described above, in the AE, a model is constructed to predict the output variable from the input variables so as to well predict the normal data. That is, for the data to be detected later, if there is a deviation between the input variable and the output variable, it can be determined as abnormal. Using the above-described loss function, an abnormality determination criterion is defined using the score indicating the relevance defined by the following formulas (5) and (6), and it is possible to detect an abnormality for the input variables of the data to be detected later.

Number

Number

[0041] Here, \(\gamma_j\) is a score indicating the relevance for the input variable \(j\), and \(\gamma_j'\) is a score indicating the standardized relevance. Using the score \(\gamma_j\) indicating the relevance calculated from the normal data used for learning, an abnormality determination criterion can be defined. The abnormality determination criterion can be any numerical value. For example, when the scores indicating the relevance calculated from the normal data used for learning are arranged in descending order, the top 8 Number items can be used, etc., but it is not limited to this. Data related to quality abnormality is detected based on the deviation from this abnormality determination criterion. In the case of AE, this degree of deviation \(Abnormality_j\) is taken as the "degree of relevance" in the estimation of quality abnormality factors as shown in the following formula (7).

Number

[0042] In the case of RF or SVR, for example, the root mean square error of the input variable and the output variable, or the root mean square error of the actual target quality data and the predicted output variable is defined as the loss function, and a model with improved prediction performance in normal data is constructed. Then, the output of the input data is evaluated to detect data related to quality abnormalities. As a method for detecting data related to quality abnormalities in this case, for example, SHAP (Shapley Additive Explanation), a method that utilizes the Shapley value, may be used. Since SHAP is a well-known method, it will not be described in detail here and will be briefly explained. SHAP can represent how much each input variable contributed to the prediction result using the average marginal contribution (Shapley value) of players in a cooperative game in game theory. When abnormal data is input into a model constructed using only normal data as in this case, the contribution in abnormal data items increases, so it can be estimated as a quality abnormality factor.

[0043] When a statistical method is used for the abnormality factor estimation model, for example, for each data item of the performance data and quality data related to the target quality data in the aforementioned normal data, a representative waveform is constructed, and by looking at the similarity with the representative waveform for the subsequent data to be detected, the quality abnormality factor can be estimated. The representative waveform may be, for example, the average value of each point as shown in the following formula (8). [Number]

[0044] Here, rep j,i is the data of the i-th point of the representative waveform in data item j. The similarity with the representative waveform may be, for example, the root mean square error. That is, the larger the value, the lower the similarity can be evaluated. By treating the similarity in the same way as the score indicating the relevance in AE, the relevance can be calculated and the quality abnormality factor can be estimated.

[0045] In any method, it is desirable to standardize the input variables for each data item. The standardization may be calculated based on the maximum and minimum values of the normal data, or may be calculated from the average value and the standard deviation. Also, even for a single steel plate Pr, the number of measurement points for each data item of the performance data related to the quality data and the quality data, that is, the length of the steel plate at the measurement location, often differs. Therefore, interpolation of the number of measurement points by approximation or standardization of the length of the steel plate may be performed.

[0046] Using the information with the above-described degree of relevance registered and the performance data associated with the steel plate Pr, etc., the display information generation unit 23 supports the analysis of the quality abnormality factor. Various screens generated by the display information generation unit 23 are shown in FIG. 5. FIG. 5 is a diagram showing various screens displayed on the HMI device 13 based on the information generated by the display information generation unit 23. The display information generation unit 23 configures information for displaying a steel plate information list screen 231, a quality abnormality factor estimation result display screen 232, a similar quality abnormality factor estimation analysis screen 233, and a quality abnormality factor / setting analysis display screen 234, and displays these screens 231 to 234 on the HMI device 13. The steps related to the support for the quality abnormality factor analysis will be described in sequence.

[0047] First, on the steel plate information list screen 231, main product information such as the coil ID and target plate thickness of the steel plate, as well as quality evaluation values, are displayed. FIG. 6 is a diagram schematically showing the display content of the steel plate information list screen 231. The main product information such as the coil ID and target plate thickness of the aforementioned steel plate Pr, as well as quality evaluation values, are displayed in a table format as information on one steel plate Pr (one coil ID) per row. For example, for a steel plate Pr whose quality evaluation value exceeds the allowable range, such as coil ID "xxx2", the row may be highlighted, for example, by changing the background color. Also, highlighting may be performed by changing the color of the cell or the color of the text of a quality item (for example, quality A "A2") whose quality evaluation value exceeds the allowable range. On this screen 231, it is possible to check whether there is a steel plate Pr whose quality evaluation value exceeds the allowable range and its quality data. By pressing the cell of the target quality (quality A "A2") in the row of the steel plate Pr with the target coil ID "xxx2", a screen 231a as shown in FIG. 7 pops up, and an overview of the quality abnormality cause estimation result of the steel plate Pr can be known. FIG. 7 is a diagram schematically showing the display content of the screen 231a that pops up on the steel plate information list screen 231. Here, the target quality data "quality A" and the performance data "data b" related to the target quality, which has the highest relevance among the quality abnormality cause estimations, are displayed as a trend chart. Here, the portion where the allowable range of the target quality is exceeded is highlighted as shaded in the figure, and it is possible to check where (which position) of the steel plate Pr the allowable range was exceeded. On this pop-up screen 231a, only the data b with the highest relevance is displayed as the target quality abnormality cause. However, by pressing the detailed display button at the bottom of this pop-up screen 231a, a transition is made to the quality abnormality cause estimation result display screen 232 shown in FIG. 8, and the details of the quality abnormality cause estimation result for the target quality of the steel plate Pr can be confirmed.

[0048] FIG. 8 is a diagram schematically showing the display content of the quality abnormality factor estimation result display screen 232. On the quality abnormality factor estimation result display screen 232, the details of the quality abnormality factor estimation result for the target quality A of the steel plate Pr (coil ID: xxx2) selected on the steel plate information list screen 231 described above are displayed. The main product information such as the coil ID and target plate thickness of the steel plate Pr, the list of quality evaluation values, etc., and the trend chart of the target quality data are displayed. Among the quality abnormality factor estimation results of the target quality data, the performance data related to the target quality A with the highest relevance is displayed as a trend chart. The performance data related to the target quality A with the highest relevance is displayed in ascending order of relevance, and data that can be considered as a quality abnormality factor at an early stage can be confirmed. Also, by scrolling the screen, data with low relevance can be confirmed in order. Similar to the pop-up screen 231a of the steel plate information list screen 231 described above, the portions that exceed the allowable range of the target quality A are highlighted with hatching in the figure, and it is possible to confirm where on the steel plate the allowable range has been exceeded. Since the performance data with high relevance and the portions that exceed the allowable range can be confirmed at a glance, if the user is a skilled person with long experience in the hot rolling line, it is possible to easily reach the countermeasures for quality improvement. Even if the user is not a skilled person, since the data with high relevance is displayed at the top, it is possible to quickly connect to the next action such as reporting problems and exchanging countermeasures.

[0049] Furthermore, by pressing the similarity analysis button 232a on the quality abnormality factor estimation result display screen 232, it is possible to transition to the similar quality abnormality factor estimation analysis screen 233 shown in FIG. 9. FIG. 9 is a diagram schematically showing the display content of the similar quality abnormality factor estimation analysis screen 233. On the similar quality abnormality factor estimation analysis screen 233, the tab button of "actual data" is pressed. On the similar quality abnormality factor estimation analysis screen 233, it is possible to compare and analyze the quality abnormality factor estimation results between other steel plates Pr in which the characteristics of quality abnormalities similar to the steel plate Pr selected on the steel plate information list screen 231 described above are registered. The quality data of the steel plates in which the characteristics of the same quality abnormality are registered are collectively displayed as a trend chart in the same graph. The coil ID of each steel plate (eight coil IDs in the figure) is also shown at the same time, and by pressing this, functions more specialized for visualization may be provided, such as only the trend chart of the steel plate Pr being highlighted. At the same time, the actual data related to the abnormality of the target quality is listed and displayed as a trend chart. Similar to the quality abnormality factor estimation result screen 232 described above, the portions (quality abnormality occurrence locations) of the trend charts of the target quality data and related actual data that exceed the allowable range of the target quality are highlighted as shaded in the figure, but the range of the corresponding portions may be color-coded, for example, by the minimum value, maximum value, average value, etc. Also, the distribution of the relevance for each of the data b, d, a is similarly listed and displayed. In the figure, the distribution of the relevance for each data is displayed in a table format as the numerical value of the number of occurrences of the steel plate, but those with high numerical values may be highlighted by the shade of the background color or the like. Alternatively, it may be displayed as a histogram chart.

[0050] By pressing the tab button of "setting data" on the similar quality abnormality factor estimation analysis screen 233, it is possible to transition to the quality abnormality factor - setting analysis display screen 234 shown in FIG. 10. FIG. 10 is a diagram schematically showing the display content of the quality abnormality factor - setting analysis display screen 234. On the quality abnormality factor - setting analysis display screen 234, not only the actual data related to the target quality abnormality but also the analysis regarding the setting data is possible. For example, as a factor candidate, the setting FixedIt is possible to support the analysis of the target quality abnormality and the setting data by, for example, chart - displaying the degree of relevance to the setting data such as values 4, 7, 2, 8, 5 as a box - and - whisker plot.

[0051] By checking such similar quality abnormalities and the results of their cause estimation for a large number of steel plates Pr, it is possible to support the analysis of whether the quality abnormality occurs constantly, whether it always occurs due to the same cause, whether it occurs suddenly under specific conditions such as steel type and target plate thickness, as well as related setting data. Although all are shown as trend charts across the entire length of the steel plate Pr, when calculating the quality evaluation value for each section divided by the length excluding the unsteady deformation part at the head and tail ends in the longitudinal direction of the steel plate Pr, it may also be a chart corresponding to the target section. Also, since the set value is set before rolling, by analyzing the set value estimated as a candidate cause before rolling based on the degree of relevance, the occurrence of quality abnormalities can be prevented. By combining this with the analysis of quality abnormalities based on the time - series data (actual data) after rolling, the usability of the user can be improved.

[0052] As described above, according to this embodiment, by acquiring the quality - abnormal location where the quality evaluation value of the abnormal product exceeds the allowable range, it is possible to display at which position in the longitudinal direction of the product the quality abnormality occurs. Furthermore, by constructing a quality - abnormality cause - estimation model using normal data, estimating the candidate causes of the quality abnormality using the constructed quality - abnormality cause - estimation model, and calculating the degree of relevance of the estimated candidate causes to the quality abnormality, it is possible to display the time - series data with a high degree of relevance as the cause of the quality abnormality. Therefore, when a product is determined to have a quality abnormality, it is possible to provide a quality - abnormality cause - analysis support system that can display both at which position in the longitudinal direction of the product the quality abnormality occurs and its cause.

[0053] Embodiment 2. Next, regarding Embodiment 2 of the present disclosure, the differences from the above-described Embodiment 1 will be mainly described. In the above-described Embodiment 1, the construction of the quality abnormality factor estimation model was only at the timing that met the above-described conditions or at any timing determined by the user (including the administrator) of the quality abnormality factor analysis support system 20. Just constructing the quality abnormality factor estimation model at such a timing, that is, only with the dataset at that timing, it is difficult to cover all the quality abnormality patterns of the rolling mill 1.

[0054] This embodiment is characterized in that, after the above timing, the above-described quality abnormality factor estimation model is constructed for each timing when the rolling of the steel plate Pr is completed. FIG. 11 is a diagram showing the processing flow of the quality abnormality factor analysis support system according to Embodiment 2.

[0055] The routine shown in FIG. 11 is started every time rolling is completed. In this routine, similar to the routine shown in FIG. 3, quality data xi is acquired (step S1), and a quality evaluation value is calculated (step S2). Then, it is determined whether the quality evaluation value is out of the allowable range that is the reference value (step S4). The processing up to this point is the same as that in Embodiment 1.

[0056] In this embodiment, when it is determined in step S5 that the quality evaluation value of the steel plate Pr is within the allowable range, it is registered as normal data with a label (step S5), and the data of the steel plate Pr is newly added to the normal data used when the model was constructed last time, and the quality abnormality factor estimation model is reconstructed (step S12). Next, the error calculated from the loss function of the quality abnormality factor estimation model reconstructed in step S12 and the error calculated from the loss function of the model constructed last time are compared (step S13). If the error of the reconstructed model is small, the reconstructed quality abnormality factor estimation model is saved (step S14).

[0057] According to this embodiment, after constructing the quality abnormality factor estimation model in the above-described Embodiment 1, the quality abnormality factor estimation model can be updated by reconstructing the model every time rolling is completed. As a result, the reproducibility of the input variables or the prediction performance of the target quality data is improved, and it becomes possible to estimate the quality abnormality factors using the quality abnormality factor estimation model that always matches the state of the latest rolling plant 1. Further, when the prediction error of the reconstructed quality abnormality factor estimation model is smaller than the prediction error of the currently used quality abnormality factor estimation model before reconstruction, the quality abnormality factor estimation model is updated (saved in the data storage unit 21), thereby preventing a decrease in accuracy due to model reconstruction.

[0058] Note that, in this embodiment, the model is reconstructed every time rolling is completed, but the present invention is not limited to this, and the timing of reconstruction can be appropriately set by the user. For example, when the user feels something strange when checking the quality abnormality factor estimation result display screen 232, the model can also be reconstructed. Thereby, the accuracy can be improved.

[0059] FIG. 12 is a diagram showing an example of the hardware configuration of the quality abnormality factor analysis support system 20. Each of the above-described functions of the quality abnormality factor analysis support system 20 can be realized by the processing circuit shown in FIG. 12. This processing circuit 20 may be dedicated hardware 20a. This processing circuit may include a processor 20b and a memory 20c. This processing circuit may be partly formed as dedicated hardware 20a and further include a processor 20b and a memory 20c. In the example of FIG. 12, part of the processing circuit 20 is formed as dedicated hardware 20a, and the processing circuit 20 also includes a processor 20b and a memory 20c. The processing circuit 20 may be at least one dedicated hardware 20a. In this case, the processing circuit 20 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel-programmed processor, an ASIC, an FPGA, or a combination thereof. The processing circuit 20 may include at least one processor 20b and at least one memory 20c. In this case, each function of the quality abnormality factor analysis support system 20 is realized by software, firmware, or a combination of software and firmware. The software and firmware are described as programs and stored in the memory 20c. The processor 20b realizes each function of the quality abnormality factor analysis support system 20 by reading and executing the programs stored in the memory 20c. The processor 20b is also called a CPU (Central Processing Unit), a central processing unit, a processing unit, an arithmetic unit, a microprocessor, a microcomputer, or a DSP. The memory 20c corresponds to a storage device such as a non-volatile or volatile semiconductor memory such as a RAM, a ROM, a flash memory, an EPROM, or an EEPROM. The memory 20c can also serve as a database DB. Thus, the processing circuit 20 can realize each function of the quality abnormality factor analysis support system 20 by hardware, software, firmware, or a combination thereof.

[0060] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and can be implemented with various modifications without departing from the gist of the present invention. In the above embodiment, the information generated by the display information generation unit 23 is transmitted to the HMI device 13, and the HMI device 13 is configured to display various screens 231, 232, 233, 234. However, a display unit may be provided inside the quality abnormality factor analysis support system 20, and various screens may be configured to be displayed on this display unit.

[0061] Also, when referring to numbers such as the number, quantity, amount, range, etc. of each element in the above-described embodiments, the present invention is not limited to the mentioned number, except when specifically stated or clearly specified by principle. Further, the structures and the like described in the above-described embodiments are not necessarily essential to the present invention, except when specifically stated or clearly specified by principle.

Explanation of Reference Numerals

[0062] 1... Rolling plant, 13... HMI device (display unit), 20... Quality abnormality factor analysis support system, processing circuit, 20a... Dedicated hardware, 20b... Processor, 20c... Memory, 21... Data storage unit, 22... Quality abnormality factor estimation calculation unit, 23... Display information generation unit, 231... Steel plate information list screen, 232... Quality abnormality factor estimation result display screen, 233... Similar quality abnormality factor estimation analysis screen, 234... Quality abnormality factor / setting analysis display screen

Claims

**Claim 1**: A quality abnormality factor analysis support system comprising a data storage unit, a quality abnormality factor estimation calculation unit, and a display information generation unit. The data storage unit is configured to store time-series data during manufacturing as manufacturing conditions of products manufactured by a rolling plant, setting conditions of facility equipment installed in the rolling plant, and performance data obtained from measuring equipment installed in the rolling plant. The quality abnormality factor estimation calculation unit calculates a quality evaluation value from quality data related to the product quality of the product selected from the time-series data during manufacturing, regards an abnormal product or product group whose calculated quality evaluation value exceeds the allowable range as an abnormal product, and acquires a quality abnormality location where the abnormal product exceeds the allowable range, registers the time-series data, the manufacturing conditions, and the setting conditions related to a normal product group whose quality evaluation value is within the allowable range as normal data, constructs a quality abnormality factor estimation model that uses machine learning or statistical methods based on the time-series data registered as the normal data, or the manufacturing conditions and the setting conditions registered as the normal data, is configured to estimate a factor candidate for the quality abnormality from the time-series data, or the manufacturing conditions and the setting conditions, based on the constructed quality abnormality factor estimation model, and calculate the degree of relevance of the estimated factor candidate to the quality abnormality. The display information generation unit is configured to generate information for displaying, on a display unit, the quality data of the abnormal product, the quality abnormality location, the time-series data, or the manufacturing conditions and the setting conditions, and the factor candidate and the degree of relevance estimated and calculated by the quality abnormality factor estimation calculation unit. **Claim 2** The quality abnormality factor analysis support system according to claim 1, wherein the display information generation unit generates information for displaying a list of the manufacturing conditions, the setting conditions, the quality evaluation value, and the quality abnormality location of the abnormal product. **Claim 3** The quality abnormality factor analysis support system according to claim 1, wherein the display information generation unit generates information for displaying a trend chart of the quality data of the abnormal product, a trend chart of the time-series data corresponding to the factor candidate having the highest degree of relevance among the factor candidates estimated by the quality abnormality factor estimation calculation unit, and the quality abnormality location in each trend chart. **Claim 4** The display information generation unit generates information for displaying a trend chart of the quality data of the abnormal product, a plurality of the factor candidates estimated by the quality abnormality factor estimation calculation unit, the relevance of each factor candidate, a trend chart of the time-series data corresponding to each factor candidate, and the quality abnormality location in each trend chart. The quality abnormality factor analysis support system according to claim 1.

5. The display information generation unit generates information for displaying, for the abnormal product with the same quality data and quality abnormality location, the set conditions as the factor candidates and the relevance of each set condition to the quality abnormality. The quality abnormality factor analysis support system according to claim 1.

6. The quality abnormality factor estimation calculation unit reconstructs the quality abnormality factor estimation model every time the product is manufactured, and updates the quality abnormality factor estimation model when the prediction error of the reconstructed quality abnormality factor estimation model is smaller than the prediction error of the quality abnormality factor estimation model before reconstruction. The quality abnormality factor analysis support system according to any one of claims 1 to 5.

Images