EP3831511A1 - Method and computer system for predicting a shrinkage of a cast metal product - Google Patents

Abstract

The invention relates to a method for predicting shrinkage of a metal product which has been produced from liquid metal by a casting process. The object of the invention is to create a method which can predict shrinkage very precisely. This object is achieved by a neural network (10) which consists of a multi-layer feedforward network with a large number of input parameters. The input parameters are characteristic parameters of the casting process and the metal product. The shrinkage of the cast metal product is output at the output of the neural network. The task is also achieved by a computer system with a memory which contains a neural network (10). This neural network (10) consists of a multi-layered feedforward network, which has a large number of input parameters and the shrinkage of a cast metal product is output as the output. The computer system (3) has inputs (5) for data which are used as input parameters and the shrinkage is transmitted to a display device (8), a control device and / or a regulating device (9).

Description

Field of technology

The present invention relates to the field of metal casting processes.

On the one hand, the invention relates to a method for predicting a shrinkage, preferably a shrinkage width and / or a shrinkage height of a strand cross-section, of a metal product which was produced by a casting process from liquid metal, preferably a slab cast by a continuous casting plant.

On the other hand, the invention relates to a computer system with a memory.

State of the art

Within a casting plant for metals, for example a continuous casting plant, the problem arises that a specified dimension of the cast product - in the cold state - has to be set during and / or before the casting process. As a result of the solidification processes and cooling of a metal melt, the cast product undergoes shrinkage, which results in a reduction in the dimensions of the cast product, which are influenced by the casting and quality parameters of the metal melt.
During casting in a continuous casting plant, ferrostatic pressure also acts on a strand shell, which leads to so-called creep. As a result, a slab cast by the continuous caster experiences an increase in width. Furthermore, a change in the structure of the metal during cooling also influences a change in the dimensions.

The consequence of this is that the dimensions of the cooled slab deviate from the dimensions that were set on a mold. In order to be able to produce slabs with defined dimensions, the deviation must be compensated in advance. The setting of the mold - in particular the mold side plates - must therefore be made in accordance with the desired slab dimensions. Due to a large number of influencing factors on the shrinkage of the cast product, a mathematical representation is difficult or imprecise. This makes the prediction and thus the correct setting of the casting process difficult. Mathematical models or tables based on empirical values are currently used. In the EP2279052 B1 a method for continuous casting is shown in which a mathematical simulation model is used to calculate the shrinkage. These are simplified and can only take into account changes in the melt, production parameters or other influencing parameters with a certain error range.

Summary of the invention

The object of the invention is to provide a reliable method for predicting the shrinkage as precisely as possible.

The object is achieved by a neural network which consists of a multi-layer feedforward network. The neural network has a large number of input parameters, which are characteristic parameters of the casting process and the metal product. The output is the shrinkage of the cast metal product.

The multi-layered feedforward network has shown very good results. The input parameters depend very much on the casting system and the cast metal melt. The molten metal can be made from a variety of Alloy elements exist, each of which can have a different influence on the shrinkage. Furthermore, any cooling devices and other system-specific features also have an influence on the shrinkage.
In addition to an output layer, a multilayer feedforward network also has at least one hidden layer. The output of these hidden layers is not visible from the outside. The neural network is trained for each casting plant, for example by recording measurement data during commissioning and feeding it to the neural network accordingly.

• eine aktuelle Gießbreite
• eine Temperatur der Schmelze
• eine Gießgeschwindigkeit
• eine Temperatur des Erstarrungspunktes
• Zusammensetzung der Schmelze
• eine aktuelle Position von Seitenwänden einer Kokille
• Angaben über Art und Menge von Legierungselementen der Schmelze,
• ein Gießpulvertyp,
• Parameter der Kühlstrecke sind
• Betriebsparameter einer Gießanlage, insbesondere einer Stranggussanlage.

In a preferred embodiment, a selection from the following input parameters is used for the feedforward network:

• a current casting width
• a temperature of the melt
• a casting speed
• a temperature of the freezing point
• Composition of the melt
• a current position of sidewalls of a mold
• Information on the type and quantity of alloying elements in the melt,
• a casting powder type,
• Parameters of the cooling section are
• Operating parameters of a casting plant, in particular a continuous casting plant.

This list is not exhaustive. Of course, other input parameters can also be used. The selection of which input parameters are used to predict the shrinkage depends on the molten metal and its composition. The casting plant is also of crucial importance. The molten metal usually has several alloy elements such as carbon, silicon, manganese, titanium, chromium, nickel, bromine, arsenic and / or other alloy elements. To as much as possible To obtain exact results, the respective proportion of the alloying elements should be available to the neural network as an input parameter.

A particularly preferred embodiment provides that the multilayer feedforeard network has at least two hidden layers, particularly preferably at least three hidden layers. The use of at least two hidden layers enables a very good prediction of the strand shrinkage. By using three hidden layers, noise and additional non-linearities are also taken into account. The number of layers used depends very much on the particular casting plant. In most cases, having two layers hidden will give the best results. However, in some cases - especially with more complex systems - it can be advantageous to use more than two shifts. If too many hidden layers are used, the problem arises that the too high model order makes the results worse again - this is known as the overfitting tendency.

An advantageous embodiment provides that the respective hidden layers each have up to 250 neurons. A grid search with a simplification (grid search with pruning) and a cross-validation result in an optimized number of neurons in the hidden layers. For the prediction of the shrinkage, the number depends on the number of input parameters. It has been found that with a number of up to 250 neurons, very good predictions for the shrinkage can be achieved.

An expedient embodiment provides that a rectified linear unit, a rectangular function, a Tanh function or a Gaussian function is used as the activation function of the individual neurons. The Retified Linear Unit (RELI) function has proven to be particularly advantageous and leads to very precise results. But it is also a rectangle, a Tanh or a Gaussian function conceivable to get the good results of the shrinkage. It is possible for several different activation functions to be used in the neural network. It can therefore be advantageous that not all neurons have the same activation function.

A particularly preferred embodiment provides that the shrinkage, preferably the shrinkage width and / or shrinkage height, is fed to a control and / or regulating device of the continuous casting plant. The prediction of the shrinkage - i.e. the dimensions of the cast slab - can be used directly for the control and / or regulating device in order to make the desired settings on the continuous casting plant. By directly including this forecast, it is always possible to react to changed conditions - such as changed composition or temperature of the molten metal. By predicting a change in shrinkage, one or more parameters of the continuous casting plant - for example the casting width - can be adjusted accordingly.

In a particularly advantageous embodiment, the control and / or regulating device controls and / or regulates the position of the side walls of a mold.
The described method allows the position of the side walls of a mold to be controlled or regulated in a particularly simple manner so that the cast slab has the desired dimensions in the cooled state.

In a particularly preferred embodiment, the actual width and / or actual height of the metal product is measured. If there is a discrepancy between the shrinkage width and actual width and / or between the shrinkage height and the actual height, the output is used to calculate the input parameters back to the input parameters and the cause of the respective deviation is determined. The actual width and actual height are measured in the state in which the prediction of the shrinkage is made - for example, in the cooled state.

By calculating back from the output layer along the individual layers to the input, the input parameters can be calculated and thus possible deviations can be determined. Due to the incorrect input parameter, the cause, such as incorrect data transmission or incorrect measurement, can be determined. This advantageous embodiment enables faulty measuring equipment to be identified quickly, for example, or measuring errors to be recognized.

The object is also achieved by a computer system of the type mentioned above. The computer system has a memory which contains a neural network which consists of a multilayered feedforward network. This has a large number of input parameters. The output is the shrinkage of a cast metal product. The computer system has inputs for measurement data and / or other data which are used as input parameters. The computer system is connected to a display device, a control device and / or a regulating device in order to transmit the shrinkage to the latter. With this computer system the shrinkage of a metal product can be predicted very well.

A preferred embodiment provides that the multilayer feedforward network has at least two hidden layers, particularly preferably at least three hidden layers. The same advantages result as already explained under the method.

Another preferred embodiment provides that the computer system is connected to a continuous casting plant for casting slabs and operating data are used as input parameters.

In an expedient embodiment, the open-loop and / or closed-loop control of a position of side walls of a mold is carried out by the control and / or regulating device.

• Fig. 1 eine schematische Darstellung einer Stranggußanlage
• Fig. 2 ein neuronales Netz zur Vorhersage einer Schrumpfung
• Fig. 3 Vergleich von Messergebnissen und Vorhersage der Schrumpfung

Brief description of the drawings

• Fig. 1 a schematic representation of a continuous casting plant
• Fig. 2 a neural network for predicting shrinkage
• Fig. 3 Comparison of measurement results and prediction of shrinkage

Description of the embodiments

In the Fig. 1 a continuous casting plant 1 is shown schematically. Liquid metal 6 is poured into a mold 2 and a cast strand 7 is then withdrawn from the mold. A computer system 3, which is connected to a memory 4, calculates the shrinkage of the cast strand 7 with the aid of a neural network. The shrinkage is then fed to a regulating, control device 9 and / or a display unit 8. The regulating and / or control device 9 can regulate and / or control the continuous casting plant 1 due to the shrinkage. This is done, for example, by adjusting the side walls of the mold 2. The computer system 3 receives input parameters via inputs 5. These input parameters can be transferred from measuring instruments 5a, 5b, from memory 4 via memory line 5b and / or from a higher-level control system of the industrial plant. The parameters recorded by measuring instruments are, for example, the measured strand dimensions, temperature of the melt, casting speed and / or parameters of the cooling section. A composition of the melt can either be stored in the memory or transmitted through the higher-level control system of the industrial plant. In addition to specific data from the continuous casting plant 1, measurement data can also be stored in the memory 4, which data can be used to learn the neural network.

In the Fig. 2 a structure of a neural network 10 is shown. This neural network 10 allows the shrinkage of a cast metal product to be determined very precisely. The neural network 10 consists of an input layer 11. The input layer 11 transfers important parameters of the casting process and of the liquid metal to the neural network 11. These important parameters include the current temperature of the melt, the current casting speed, a current casting width, a current angular position of the side walls of the mold, a temperature above the solidification point of the melt, alloying elements, casting powder type and / or parameters of the cooling section. These parameters are also used to train the neural network.
The neural network 10 also consists of a first hidden layer 12 and a second hidden layer 13, each of which has a large number of neurons 15. The number of neurons 15 of each hidden layer depends on the input parameters. If the input layer consists of fourteen input parameters, the first hidden layer 12 and the second hidden layer 13 each have around 250 neurons 15, for example. The neural network 10 is completed by the output layer 14. The output layer 14 outputs the shrinkage. The shrinkage can be indicated by the ratio of the cast width to that in the cooled state. Of course, the shrinkage in height, length or other dimensions can also be determined.

In the Fig. 3 shows the shrinkage of slabs that were produced by a continuous caster. A first curve shows the shrinkage, which was determined with measurement data (16). A second curve represents a prediction (17) of the shrinkage, which was determined by the neural network. The shrinkage is shown as the ratio of the cast width - the width set on the mold - to the width in the cooled state at 25 ° C.

As can be seen from the curves, the prediction (17) by the neural network agrees very well with the actually measured data.

Although the invention has been illustrated and described in more detail by the preferred exemplary embodiments, the invention is not restricted by the disclosed examples and other variations can be derived from them by the person skilled in the art without departing from the scope of protection according to the claims.

List of reference symbols

1

Continuous casting plant

2

Mold

3

Computer system

4th

Storage

5

Entrances

5a

Measuring instruments

5b

Storage line

6th

Liquid metal

7th

strand

8th

Display unit

9

Regulation and / or control device

10

Neural network

11

Entry layer

12th

First hidden layer

13th

Second hidden layer

14th

Output layer

15th

Neurons

16

Measurement data

17th

forecast

Claims (12)

A method for predicting a shrinkage, preferably a shrinkage width and / or a shrinkage height of a strand cross-section, of a metal product which was produced by a casting process from liquid metal, preferably a slab cast by a continuous casting plant (1), characterized in that a neural network (10 ), which consists of a multi-layer feedforward network with a large number of input parameters, which are characteristic parameters of the casting process and the metal product and the shrinkage of the cast metal product is output as the output. Method for predicting shrinkage of a metal product cast from liquid metal according to Claim 1, characterized in that input parameters are at least a selection from the following parameters: - a current casting width - a temperature of the melt - a casting speed - a temperature of the freezing point - composition of the melt - a current position of side walls of a mold - information on the type and quantity of alloying elements in the melt, - a type of casting powder, - parameters of the cooling section are, - Operating parameters of a casting plant, in particular a continuous casting plant. Method for predicting a shrinkage of a metal product cast from liquid metal according to claim 1-2, characterized in that the multilayer feedforeard network has at least two has hidden layers (12, 13) particularly preferably at least three hidden layers. Method for predicting a shrinkage of a metal product cast from liquid metal according to Claims 1-3, characterized in that the respective hidden layers (12, 13) each have up to 250 neurons. Method for predicting a shrinkage of a metal product cast from liquid metal according to claims 1-4, characterized in that a retified linear unit function, a rectangle function, a tanh function or a Gaussian function is used as the activation function of the individual neurons. Method for predicting shrinkage of a metal product cast from liquid metal according to claims 1-5, characterized in that the shrinkage, preferably the shrinkage width and / or shrinkage height, is fed to a control and / or regulating device (9) of the continuous casting plant (1). Method for predicting shrinkage of a metal product cast from liquid metal according to Claim 6, characterized in that the control and / or regulating device (9) controls or regulates the position of side walls of a mold (2). Method for predicting a shrinkage of a metal product cast from liquid metal according to claim 1-7, characterized in that the actual width and / or actual height of the metal product is measured and in the event of a deviation from the shrinkage width and actual width and / or from the shrinkage height and Actual height using the neural Network (10) is calculated back from the output to the input parameters and the cause of the respective deviation is determined. Computer system (3) with a memory (4), characterized in that the memory (4) contains a neural network (10) which consists of a multilayer feedforward network which has a large number of input parameters and the shrinkage of a cast metal product as the output is output, the computer system (3) having inputs (5) for data, in particular measurement data, which are used as input parameters and the computer system (3) is equipped with a display device (8), a control device and / or a regulating device (9) connected to which the shrinkage is transmitted. Computer system with a memory according to Claim 9, characterized in that the multilayer feedforeard network has at least two hidden layers (12, 13), particularly preferably at least three hidden layers. Computer system with a memory according to Claims 9-10, characterized in that the computer system is connected to a continuous casting plant (1) for casting slabs and operating data are used as input parameters. Computer system with a memory according to Claim 11, characterized in that the control and / or regulating device (9) controls or regulates the position of side walls of a mold (2).

Images