JP2024000012A - Teacher data creation method - Google Patents

Abstract

To provide a teacher data creation method which can automatically create correct and large amounts of teacher data for predicting the breakout of a continuous casting apparatus with Supervised Learning.SOLUTION: Time sequence molten metal face level data detected by molten metal level detection mean provided at a tundish 2 are differentiated with respect to time sequence temperature data detected by temperature detection means 4 installed in a mold 3 of a continuous casting apparatus to determine the data type of the time sequence temperature data, further, time sequence casting speed data detected by casting speed detection means are differentiated to determine the starting point and/or the terminal point of a constant region of the data type, and further, a region in which the casting speed data is reduced to a prescribed value is determined as a break out part by an existing break out detection device and the time sequence temperature data are labeled to create teacher data.SELECTED DRAWING: Figure 3

Description

The present invention relates to a method for creating training data for predicting breakouts in continuous casting equipment using supervised learning.

If a breakout occurs in a continuous casting machine, casting becomes impossible and recovery from the breakout requires a lot of money. Therefore, methods for predicting breakouts have been devised in the past, such as the method described in Patent Document 1 (conventional method). Proposed.

In the conventional method described above, a large number of thermocouples are embedded in the mold of a continuous casting machine to collect temperature transition data when breakout occurs and when breakout does not occur. A temperature change pattern detection network and a breakout detection network are trained, and the output of the temperature change pattern detection network and the peak signal obtained by processing the thermocouple output signal are input to the breakout detection network to detect a breakout detection network. We are predicting outs.

JP 7-178524

However, since the conventional method described above still does not have sufficient accuracy in predicting breakout, there is a problem of deterioration of production efficiency due to unnecessary reduction of casting speed.

Therefore, in order to predict breakouts with high accuracy and improve production efficiency, it is possible to use supervised learning of AI (artificial intelligence) to predict breakouts more accurately. In order to function well, it is necessary to prepare a large amount of accurate training data for learning, and there is a problem in that preparing the training data requires a large amount of man-hours.

The present invention solves these problems and provides a method for creating training data that can automatically generate accurate and large amounts of training data for predicting breakouts in continuous casting equipment using supervised learning. With the goal.

In order to achieve the above object, the first invention provides a method for creating training data for predicting a breakout in a continuous casting device by supervised learning, the method comprising: The time-series temperature data detected by the temperature detection means (4) is differentiated from the time-series temperature data detected by the hot water level detection means provided in the tundish (2), and the time-series temperature is determined by differentiating the time-series temperature data detected by the temperature detection means (4). In addition to determining the data type of the data, the time-series casting speed data detected by the casting speed detection means is differentiated to determine the starting point and/or end point of the "steady part" of the data type, and the existing breakout detection device The area where the casting speed data is reduced to a predetermined value is determined to be a breakout portion, and the time-series temperature data is labeled, thereby creating the teacher data.

In the second invention, there are four types of labeling of the teacher data during normal operation: "steady part", "rising part + steady part", "steady part + falling part", "rising part + steady part + falling part". The labeling of the teacher data at the time of breakout prediction is of two types: "steady part + breakout part" and "rising part + steady part + breakout part".

The above reference numerals in parentheses indicate for reference the correspondence with specific means described in the embodiments to be described later.

According to the teaching data creation method of the present invention, it is possible to automatically generate a large amount of accurate teaching data for predicting a breakout of a continuous casting machine by supervised learning.

FIG. 1 is a schematic configuration diagram of a continuous casting apparatus. It is a schematic perspective view of a mold. FIG. 3 is a developed view of a mold side wall on which a thermocouple is installed. FIG. 3 is a diagram showing an example of changes over time in mold temperature, molten metal level, and casting speed during normal operation. FIG. 3 is a diagram showing an example of changes over time in mold temperature, molten metal level, and casting speed when predicting BO. 3 is a flowchart showing a procedure for generating labeled teacher data in a computer. FIG. 7 is a diagram showing an example of labeled teacher data during normal operation. It is a figure which shows an example of labeled teacher data at the time of BO prediction.

Note that the embodiments described below are merely examples, and various design improvements made by those skilled in the art without departing from the gist of the present invention are also included within the scope of the present invention.

FIG. 1 schematically shows a continuous casting apparatus that is the target of breakout (BO) prediction. In Fig. 1, molten steel Sm is supplied from a ladle 1 to a tundish 2, and from the tundish 2 passes through a mold 3, a slab St in which unsolidified molten steel Sm is wrapped in a thin solidified shell is formed on a large number of supports. It is pulled out between rolls 31.

In such a continuous casting apparatus, as described above, the generation of BO is prevented by predicting BO due to restraint or breakage of the solidified shell in the mold 3 and reducing the casting speed. If the accuracy is low, the casting speed will be lowered frequently and production efficiency will deteriorate.

For this reason, there is a need to improve the accuracy of BO prediction, and it is possible to improve the prediction accuracy through supervised learning of AI. In order to successfully perform supervised learning, accurate and large amounts of teaching data are required.Hereinafter, a method of the present invention that can automatically generate such teaching data will be specifically described.

FIG. 2 shows an overview of the mold 3. In FIG. 2, a large number of thermocouples 4 serving as temperature detection means are installed on the outer periphery of the side wall of a square container-shaped mold 3, and temperature changes in the casting direction (arrow in the figure) are measured over the entire circumference of the mold 3. ing. The temperature signals from these thermocouples 4 are input to a prediction system using, for example, the above conventional method, so that when BO is predicted, the casting speed is reduced.

The arrangement of the thermocouples 4 on the four sides of the outer periphery of the side wall of the mold 3 is actually as shown in FIG. 3, and in this embodiment, a total of 28 thermocouples 4 are provided. The temperature signal of the thermocouple 4 is input to a prediction system using the conventional method as described above, and is also input to a computer 5 for implementing the method of the present invention described below. In addition to inputting mold temperature signals 4a from each of the 28 thermocouples 4, the computer 5 also receives a molten metal level signal 5a from a molten metal level meter that detects the height of molten steel in the mold 3, and a molten metal level signal 5a from a pinch roll. A casting speed signal 5b calculated from the rotational speed is input, and each of these signals 4a, 5a, 5b is converted into digital data and processed in the computer 5 as described later.

Figures 4 (1) to (4) show four examples of changes over time in mold temperature (actually color-coded according to the temperature detected by each thermocouple), molten metal level, and pouring speed during normal operation. Figures 5 (1) to (4) show the mold temperature (actually color-coded according to the temperature detected by each thermocouple), the hot water level, and the temperature when BO is predicted by the existing prediction system. Four examples of changes in pouring speed over time are shown.

As is clear from Figs. 4 and 5, the hot water level data shows large values corresponding to the rising and falling parts of the mold temperature data (arrows in Figs. 4 and 5). Four types of data types can be determined: ``steady part,'' ``rising part + steady part,'' ``steady part + falling part,'' and ``rising part + steady part + falling part.'' The boundaries (arrows in FIGS. 4 and 5) where the change mode of the casting speed data suddenly changes correspond to the start and/or end points of the "steady part" in the data type of the mold temperature data. Further, when BO is predicted by the existing prediction system, the casting speed data is suddenly dropped to a predetermined value or 0 (arrow in FIG. 5), and this BO prediction point becomes the end point of the "steady part".

Based on such knowledge, labeled teacher data for supervised learning is created in the computer 5 according to the following procedure. The procedure for creating teacher data within the computer 5 is shown in FIG. In step 101 of FIG. 6, the oil level data is differentiated to determine the above-mentioned data type of the mold temperature data. In step 102, smoothing processing is applied to the casting speed data to remove noise, and in step 103, the casting speed data is differentiated to find the boundary where the slope suddenly changes, which is the "steady part" in the data type of mold temperature data. Determine the start point and/or end point. In step 104, the process branches to the following steps depending on whether the mold temperature data is from normal operation or from BO prediction.

If the mold temperature data is during normal operation, labeled teacher data is generated from the data type and the start point and/or end point position determined in step 101 and step 103, respectively (step 105). In this case, the teacher data is labeled in four types: "steady part," "rising part + steady part," "steady part + falling part," and "rising part + steady part + falling part."

If the mold temperature data is at the time of BO prediction, in step 106, the period after the point when the casting speed data suddenly drops is set as the BO part, and in the following step 107, the data type, starting point, and/or data determined in step 101 and step 103 are determined, respectively. Alternatively, teacher data labeled with the end point position and, in addition to these, the BO portion is generated (step 107). In this step 107, the "falling part" of the data type is replaced with the "BO part" in which the casting speed after BO prediction is suddenly reduced. Therefore, the labeling of the teacher data in this case is of two types: "steady part + BO part" and "rising part + steady part + BO part".

FIG. 7 shows an example of labeled teacher data during normal operation. "Rising part + steady part + falling part" is labeled according to the change in mold temperature over time, and the labeling is represented by a string of numbers, with each number "0", "1", and "2" representing the "rising part". , "steady part", and "falling part". The boundary between the rising part and the steady part is the starting point, and the boundary between the steady part and the falling part is the ending point.

FIG. 8 shows an example of labeled teacher data at the time of BO prediction. "Rising part + Steady part + BO part" is labeled according to the change in mold temperature over time, and the labeling is represented by a string of numbers, with each number "0", "1", and "3" indicating "Rising part" and "Steady part". "falling part" and "falling part". The boundary between the rising part and the steady part is the starting point, and the boundary between the steady part and the BO part is the ending point.

As described above, for a large number of mold temperature data during normal operation and during BO prediction, the same number of mold temperature data, which are appropriately labeled as "steady part," "rising part," "falling part," and "BO part," are prepared. Training data is automatically generated. The training data generated in this way is accurate because it does not involve human intervention, and by using such training data to perform supervised learning on AI, it is possible to accurately predict the occurrence of BO. Therefore, production efficiency can be improved.

1... Ladle, 2... Tundish, 3... Mold, 4... Thermocouple (temperature detection means), 5... Computer.

Claims (2)

A method for creating training data for predicting a breakout in a continuous casting device using supervised learning, the method comprises The data type of the time-series temperature data is determined by differentiating the time-series hot water level data detected by the hot water level detection means provided in the dish, and the time-series pouring speed data detected by the pouring speed detecting means is determined. Differentiate to determine the starting point and/or end point of the "steady part" of the data type, and further determine the area where the casting speed data is reduced to a predetermined value by the existing breakout detection device as the breakout part, and A teacher data creation method for creating the teacher data by labeling time-series temperature data. There are four types of labeling for the training data during normal operation: "steady part," "rising part + steady part," "steady part + falling part," and "rising part + steady part + falling part." Breakout prediction 2. The teacher data creation method according to claim 1, wherein the teacher data is labeled in two types: "steady part + breakout part" and "rising part + steady part + breakout part".