Evaluation method for filling state of crystallizer casting powder liquid slag

Abstract

The invention relates to a method for evaluating the filling state of crystallizer casting powder liquid slag, and belongs to the technical field of continuous casting methods in the metallurgical industry. According to the technical scheme, the method comprises the steps that thermocouple temperatures at all positions of a crystallizer copper plate, vibration strokes on the two sides of a crystallizer vibration unit and hydraulic cylinder pressure are collected on line, and data are marked; and respectively establishing and training an MASKRCNN neural network and an LSTM neural network, and carrying out online evaluation on whether filling of the casting powder liquid slag is normal or not and an abnormal position by combining prediction results of the MASKRCNN neural network and the LSTM neural network. According to the method, heat and force related data of the crystallizer are collected online, an artificial intelligence model method is adopted, online evaluation of the filling state of the casting powder liquid slag is achieved in real time, dynamic tracking evaluation can be conducted on the whole using process of the casting powder, and the result better conforms to actual production; the position of a liquid slag filling abnormal state area can be accurately provided, and operators can conveniently check and confirm abnormal conditions.

Description

Technical Field

[0001] The present invention relates to a method for evaluating the liquid slag filling state of a mold powder, belonging to the technical field of continuous casting methods in the metallurgical industry. Background Art

[0002] During the continuous casting production process, the mold powder plays an important role. It generates a liquid slag film between the mold wall and the solidified shell to play a lubricating role, reducing the drawing resistance and preventing the bonding of the solidified shell and the copper plate; filling the air gap between the billet shell and the mold to improve the heat transfer of the mold. During the use of the mold powder, it is necessary to maintain a good state to achieve the purpose of improving the surface quality of the cast billet and ensuring the smooth progress of continuous casting.

[0003] Regarding the evaluation of the mold powder, current research mainly focuses on detecting and analyzing the performance indicators of the mold powder offline with the help of experimental equipment. For example, Patent 201710035568.0 discloses a method for evaluating the friction and lubrication performance of a continuous casting mold powder, which requires preparing a mold powder briquette and measuring the tensile fracture strength or compressive fracture strength of the mold powder through a tensile testing machine. Patent 202310903368.8 discloses a method for evaluating the use effect of a continuous casting mold powder, which dissolves and heats the raw materials in a vacuum melting furnace to a certain temperature, then adds the mold powder, and measures the thickness of the liquid slag layer formed by the mold powder at different time points within the billet discharging time range. Patent 201510466447.2 discloses a method for detecting the fluidity of a liquid mold powder. After melting the mold powder at the required temperature, it is poured onto a V-shaped chute with a certain inclination angle, and the flow length of the liquid mold powder is used to evaluate the flow characteristics of the mold powder.

[0004] Based on the offline mold powder detection method, it is a spot check and overall evaluation under ideal experimental conditions. However, the melting and lubrication behaviors of the mold powder during the continuous casting production process are dynamic processes, which are dynamically affected by various process conditions such as molten steel temperature, molten steel cleanliness, mold cross-section, and drawing speed. Moreover, the mold powder is a mixed material, and it is difficult to ensure that the compositions and performances of different batches are exactly the same. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for evaluating the liquid slag filling state of a mold powder. By online collecting heat and force related data of the mold, and using the artificial intelligence model method, the online evaluation of the liquid slag filling state of the mold powder can be realized in real time, and the whole process of using the mold powder can be dynamically tracked and evaluated, and the result is more in line with the production reality; it can accurately provide the location of the abnormal state area of the liquid slag filling, which is convenient for the operator to check and confirm the abnormal situation, and effectively solves the above problems existing in the background art.

[0006] The technical solution of the present invention is: a method for evaluating the liquid slag filling state of a mold powder, comprising the following steps:

[0007] (1) Collect the thermocouple temperatures at various positions of the mold copper plate, the vibration strokes on both sides of the mold vibration unit, and the hydraulic cylinder pressures online.

[0008] (2) Mark the historical data collected.

[0009] (3) Establish a MASKRCNN neural network and an LSTM neural network.

[0010] (4) Use the MASKRCNN neural network to train the thermocouple temperature data, and use the LSTM neural network to train the vibration stroke and hydraulic cylinder pressure data.

[0011] (5) Online deploy the MASKRCNN neural network and the LSTM neural network models to predict the liquid slag filling state of the mold powder respectively.

[0012] (6) Combine the prediction results of the MASKRCNN neural network and the LSTM neural network to evaluate whether the liquid slag filling of the mold powder is normal and the abnormal positions.

[0013] In the step (1), the acquisition frequency of the thermocouple temperature value is 1 second, and the acquisition frequencies of the vibration stroke of the mold vibration unit and the hydraulic cylinder pressure value are 2 milliseconds.

[0014] In the step (2), the types of marks are divided into two states: normal liquid slag filling and abnormal liquid slag filling. Normal liquid slag filling is marked as 0, and abnormal liquid slag filling is marked as 1.

[0015] In the step (2), the data format of the mark is that the thermocouple temperature values are recorded as a two-dimensional array in the clockwise order of the spatial sequence of the thermocouples on the mold copper plate where n is the thermocouple column number; m is the thermocouple row number. At the same time, the numbers of the thermocouples in the abnormal liquid slag filling area on the mold are recorded for the data of abnormal liquid slag filling. The mold vibration data records 1000 groups of data as a two-dimensional array in the order from far to near in time where S is the vibration stroke, mm; F is the hydraulic cylinder pressure, N; L is the data of the left hydraulic cylinder; R is the data of the left hydraulic cylinder.

[0016] In the step (3), the input pixels of the MASKRCNN neural network are n*m, the backbone network uses resnet101, the anchor points of the network RPN candidate boxes use [1×2, 2×2, 4×2, 8×2, 16×2], the network learning rate uses 0.001, and the detection threshold of the network for the abnormal liquid slag filling area is 0.3.

[0017] In step (3), the LSTM neural network adopts a seven-layer structure, including an input layer with 4 nodes, five LSTM layers with 8, 16, 16, 4, and 2 nodes in each layer respectively, and a Dense layer with 1 node; the optimizer of the network adopts the adam optimizer, the activation function of the first six layers of the network adopts the Relu function, the activation function of the last layer adopts the softmax function, the loss function of the network is the MSE function, and the learning rate of the network is 0.001.

[0018] In step (6), the evaluation method for whether the liquid slag filling of the mold powder is normal is as follows: if the following conditions are met simultaneously, it is evaluated that the liquid slag filling of the mold powder is abnormal: Among them, P T is the prediction probability value result of the MASKRCNN neural network, and P V is the prediction probability value result of the LSTM neural network.

[0019] In step (6), the determination method for the abnormal position of the liquid slag filling of the mold powder is as follows: after evaluating that the liquid slag filling of the mold powder is abnormal, the positioning result predicted by the MASKRCNN neural network is used as the determination result of the abnormal position of the liquid slag filling of the mold powder.

[0020] The beneficial effects of the present invention are as follows: By online collecting the heat and force related data of the mold, and adopting the artificial intelligence model method, the online evaluation of the liquid slag filling state of the mold powder is realized in real time, the whole process of using the mold powder can be dynamically tracked and evaluated, and the result is more in line with the production reality; it can accurately provide the location of the abnormal state area of the liquid slag filling, which is convenient for the operator to check and confirm the abnormal condition. Specific Embodiments

[0021] In order to make the purpose, technical solutions and advantages of the invention implementation cases clearer, the technical solutions in the invention implementation cases will be clearly and completely described below. Obviously, the described implementation cases are a small part of the invention implementation cases, rather than all of them. Based on the implementation cases in the present invention, all other implementation cases obtained by those of ordinary skill in the art without creative work belong to the protection scope of the present invention.

[0022] An evaluation method for the liquid slag filling state of a mold powder for a continuous casting mold includes the following steps:

[0023] (1) Online collect the thermocouple temperatures at various positions of the mold copper plate, the vibration strokes on both sides of the mold vibration unit, and the hydraulic cylinder pressure;

[0024] (2) Mark the collected historical data;

[0025] (3) Establish a MASKRCNN neural network and an LSTM neural network;

[0026] (4) Use the MASKRCNN neural network to train the thermocouple temperature data, and use the LSTM neural network to train the vibration stroke and hydraulic cylinder pressure data;

[0027] (5) Online deploy the MASKRCNN neural network and the LSTM neural network models to predict the liquid slag filling state of the powder flux respectively;

[0028] (6) Combine the prediction results of the MASKRCNN neural network and the LSTM neural network to evaluate whether the liquid slag filling of the powder flux is normal and the location of the abnormality.

[0029] In the above step (1), the acquisition frequency of the thermocouple temperature value is 1 second, and the acquisition frequencies of the vibration stroke and the hydraulic cylinder pressure value of the mold vibration unit are 2 milliseconds.

[0030] In the above step (2), the types of labels are divided into two states: normal liquid slag filling and abnormal liquid slag filling. The normal liquid slag filling is labeled as 0, and the abnormal liquid slag filling is labeled as 1.

[0031] In the above step (2), the data format of the label is that the thermocouple temperature values are recorded as a two-dimensional array in the clockwise order according to the spatial order of the thermocouples on the mold copper plate where n is the thermocouple column number; m is the thermocouple row number. At the same time, the numbers of the thermocouples in the abnormal liquid slag filling area on the mold are recorded for the data of the abnormal liquid slag filling. The mold vibration data records 1000 groups of data as a two-dimensional array in the order from far to near in time where S is the vibration stroke, mm; F is the hydraulic cylinder pressure, N; L is the data of the left hydraulic cylinder; R is the data of the left hydraulic cylinder.

[0032] In the above step (3), the input pixels of the MASKRCNN neural network are n*m, the backbone network uses resnet101, the anchor points of the network RPN candidate boxes use [1×2, 2×2, 4×2, 8×2, 16×2], the network learning rate uses 0.001, and the detection threshold of the network for the abnormal liquid slag filling area is 0.3.

[0033] In the above step (3), the LSTM neural network adopts a seven-layer structure, including an input layer with 4 nodes, including five LSTM layers, and the number of nodes in each layer is 8, 16, 16, 4, and 2 in turn, including a Dense layer with 1 node; the optimizer of the network uses the adam optimizer, the activation functions of the first six layers of the network use the Relu function, the activation function of the last layer uses the softmax function, the loss function of the network is the MSE function, and the network learning rate uses 0.001.

[0034] In the step (6), the evaluation method for whether the liquid slag filling of the mold powder is normal is that if the following conditions are met simultaneously, it is evaluated that the liquid slag filling of the mold powder is abnormal: Among them, P T is the prediction probability value result of the MASKRCNN neural network, and P V is the prediction probability value result of the LSTM neural network.

[0035] In the step (6), the determination method for the abnormal position of the liquid slag filling of the mold powder is that after evaluating that the liquid slag filling of the mold powder is abnormal, the positioning result predicted by the MASKRCNN neural network is used as the determination result of the abnormal position of the liquid slag filling of the mold powder.

[0036] Example:

[0037] In practical applications, the operation steps of the present invention are as follows:

[0038] First step, online collect the thermocouple temperatures at various positions of the mold copper plate, the vibration strokes on both sides of the mold vibration unit, and the hydraulic cylinder pressures. The acquisition frequency of the thermocouple temperature values is 1 second, and the acquisition frequencies of the vibration strokes and hydraulic cylinder pressure values of the mold vibration unit are 2 milliseconds;

[0039] Second step, mark the collected historical data. The types of marks are divided into two states: normal liquid slag filling and abnormal liquid slag filling. Normal liquid slag filling is marked as 0, and abnormal liquid slag filling is marked as 1. The data format of the mark is that the thermocouple temperature values are recorded as a two-dimensional array in the clockwise order of the spatial order of the thermocouples on the mold copper plate Among them, T is the thermocouple temperature, °C; n is the thermocouple column number; m is the thermocouple row number. In this embodiment, n = 28 and m = 7. At the same time, the data of abnormal liquid slag filling records the numbers of the thermocouples in the abnormal liquid slag filling area on the mold. For example, if there is a poor mold powder filling phenomenon in the thermocouple areas of column 5 row 1, column 5 row 2, column 6 row 1, and column 6 row 2, the pixel points ([5,1], [6,2]) are recorded. The mold vibration data records 1000 groups of data as a two-dimensional array in the order from far to near in time Among them, S is the vibration stroke, mm; F is the hydraulic cylinder pressure, N; L is the data of the left hydraulic cylinder; R is the data of the left hydraulic cylinder. In this embodiment, 1000 groups of normal liquid slag filling data and 800 groups of abnormal liquid slag filling data are collected;

[0040] Step 3: Establish a MASKRCNN neural network and an LSTM neural network. The input pixels of the MASKRCNN neural network are 28*7. The backbone network uses resnet101. The anchors of the network RPN candidate boxes are [1×2, 2×2, 4×2, 8×2, 16×2]. The network learning rate is 0.001. The detection threshold of the network for the abnormal area of slag filling is 0.3. The LSTM neural network adopts a seven-layer structure, including an input layer with 4 nodes, 5 LSTM layers, and the number of nodes in each layer is 8, 16, 16, 4, and 2 in sequence, and a Dense layer with 1 node; the optimizer of the network adopts the adam optimizer, the activation function of the first six layers of the network adopts the Relu function, the activation function of the last layer adopts the softmax function, the loss function of the network is the MSE function, and the network learning rate is 0.001;

[0041] Step 4: Use the MASKRCNN neural network to train the thermocouple temperature data, and use the LSTM neural network to train the vibration stroke and hydraulic cylinder pressure data;

[0042] Step 5: Online arrange the trained MASKRCNN neural network and LSTM neural network models to predict the powder slag filling state respectively;

[0043] Step 6: Combine the prediction results of the MASKRCNN neural network and the LSTM neural network to evaluate whether the powder slag filling is normal and the abnormal position. The evaluation method for whether the powder slag filling is normal is that the powder slag filling is evaluated as abnormal if the following conditions are met simultaneously: Among them, P T is the prediction probability value result of the MASKRCNN neural network, and P V is the prediction probability value result of the LSTM neural network. The determination method for the abnormal position of the powder slag filling is that after evaluating that the powder slag filling is abnormal, the positioning result predicted by the MASKRCNN neural network is used as the determination result of the abnormal position of the powder slag filling. In this embodiment, if the prediction probability value result P T of the MASKRCNN neural network is 0.2 at a certain moment, and the prediction probability value result P V of the LSTM neural network is 0.9, because P T <0.3 does not meet condition 2, the model evaluates that the current powder slag filling is normal. If the prediction probability value result P T of the MASKRCNN neural network is 0.6 at a certain moment, and the prediction probability value result P Vis 0.5. Since 0.3 * 0.6 + 0.7 * 0.5 = 0.53, which does not meet Condition 1, the model evaluates that the liquid slag filling of the current mold powder is normal. For example, at a certain moment, the predicted probability value result P of the MASKRCNN neural network T is 0.4, and the predicted probability value result P of the LSTM neural network V is 0.7. Since 0.3 * 0.4 + 0.7 * 0.7 = 0.61, which meets all three conditions simultaneously, the model evaluates that the liquid slag filling of the current mold powder is abnormal. At the same time, the abnormal location result of the mold powder predicted by the MASKRCNN neural network is pixel points ([8,1], [10,3]). Then, it is considered that there is a phenomenon of poor liquid slag filling in the thermocouple position areas of thermocouples numbered 8 column 1 row, 8 column 2 row, 8 column 3 row, 9 column 1 row, 9 column 2 row, 9 column 3 row, 10 column 1 row, 10 column 2 row, and 10 column 3 row, reminding the operator to confirm the mold powder status in this area.

Claims

1. An evaluation method for the molten slag filling state of a mold powder, characterized in that It includes the following steps: (1) Online collect the thermocouple temperatures at various positions of the mold copper plate, the vibration strokes on both sides of the mold vibration unit, and the hydraulic cylinder pressure; (2) Mark the historical data collected; (3) Establish a MASKRCNN neural network and an LSTM neural network; (4) Use the MASKRCNN neural network to train the thermocouple temperature data, and use the LSTM neural network to train the vibration stroke and hydraulic cylinder pressure data; (5) Online deploy the MASKRCNN neural network and the LSTM neural network models to predict the molten slag filling state of the mold powder respectively; (6) Combine the prediction results of the MASKRCNN neural network and the LSTM neural network to evaluate whether the molten slag filling of the mold powder is normal and the abnormal position; 2. The evaluation method for the molten slag filling state of the mold powder according to claim 1, characterized in that: In the step (1), the acquisition frequency of the thermocouple temperature value is 1 second, and the acquisition frequencies of the vibration stroke of the mold vibration unit and the hydraulic cylinder pressure value are 2 milliseconds.

3. The evaluation method for the molten slag filling state of a mold powder according to claim 1, wherein: In the step (2), the types of marks are divided into two states: normal molten slag filling and abnormal molten slag filling. Normal molten slag filling is marked as 0, and abnormal molten slag filling is marked as 1.

4. The evaluation method for the liquid slag filling state of a mold powder according to claim 1, wherein: In the step (2), the marked data format is that the thermocouple temperature values are recorded as a two-dimensional array in the clockwise spatial order of the thermocouples on the mold copper plate. Among them, n is the thermocouple column number; m is the thermocouple row number, and at the same time, the numbers of the thermocouples in the abnormal liquid slag filling area on the mold are recorded for the data of abnormal liquid slag filling. The mold vibration data are recorded as a two-dimensional array of 1000 groups of data in the order from far to near in time. Among them, S is the vibration stroke, in mm; F is the hydraulic cylinder pressure, in N; L is the data of the left hydraulic cylinder; R is the data of the left hydraulic cylinder.

5. The evaluation method for the liquid slag filling state of a mold powder according to claim 1, characterized in that: In the step (3), the input pixels of the MASKRCNN neural network are n*m, the backbone network uses resnet101, the anchor points of the network RPN candidate boxes use [1×2, 2×2, 4×2, 8×2, 16×2], the network learning rate uses 0.001, and the detection threshold of the network for the abnormal molten slag filling area is 0.

3.

6. The evaluation method for the molten slag filling state of the mold powder according to claim 1, wherein: In the step (3), the LSTM neural network adopts a seven-layer structure, including an input layer with 4 nodes, including five LSTM layers, and the number of nodes in each layer is 8, 16, 16, 4, and 2 in turn, including a Dense layer with 1 node; the optimizer of the network uses the adam optimizer, the activation function of the first six layers of the network uses the Relu function, the activation function of the last layer uses the softmax function, the loss function of the network is the MSE function, and the network learning rate uses 0.

001.

7. The evaluation method for the molten slag filling state of the mold powder according to claim 1, characterized in that: In the step (6), the evaluation method for whether the liquid slag filling of the mold powder is normal is that if the following conditions are met simultaneously, it is evaluated that the liquid slag filling of the mold powder is abnormal: Among them, P T is the prediction probability value result of the MASKRCNN neural network, and P V is the prediction probability value result of the LSTM neural network.

8. The evaluation method for the molten slag filling state of the mold powder according to claim 1, characterized in that: In the step (6), the determination method for the abnormal position of the molten slag filling of the mold powder is that after evaluating the abnormal molten slag filling of the mold powder, the positioning result predicted by the MASKRCNN neural network is used as the determination result of the abnormal position of the molten slag filling of the mold powder.

Images