US20220406413A1 - Method for predicting yield of calcium in a calcium treatment process based on deep neural network - Google Patents
Method for predicting yield of calcium in a calcium treatment process based on deep neural network Download PDFInfo
- Publication number
- US20220406413A1 US20220406413A1 US17/844,364 US202217844364A US2022406413A1 US 20220406413 A1 US20220406413 A1 US 20220406413A1 US 202217844364 A US202217844364 A US 202217844364A US 2022406413 A1 US2022406413 A1 US 2022406413A1
- Authority
- US
- United States
- Prior art keywords
- calcium
- yield
- neural network
- deep neural
- molten steel
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/10—Analysis or design of chemical reactions, syntheses or processes
-
- C—CHEMISTRY; METALLURGY
- C21—METALLURGY OF IRON
- C21C—PROCESSING OF PIG-IRON, e.g. REFINING, MANUFACTURE OF WROUGHT-IRON OR STEEL; TREATMENT IN MOLTEN STATE OF FERROUS ALLOYS
- C21C7/00—Treating molten ferrous alloys, e.g. steel, not covered by groups C21C1/00 - C21C5/00
- C21C7/0056—Treating molten ferrous alloys, e.g. steel, not covered by groups C21C1/00 - C21C5/00 using cored wires
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/30—Prediction of properties of chemical compounds, compositions or mixtures
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/70—Machine learning, data mining or chemometrics
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/06—Multi-objective optimisation, e.g. Pareto optimisation using simulated annealing [SA], ant colony algorithms or genetic algorithms [GA]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0499—Feedforward networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
Definitions
- the invention relates to the field of molten steel refining in iron and steel metallurgy, in particularly to a method for predicting a yield of calcium in a calcium treatment process based on deep neural network.
- neural networks have been widely used in various industries.
- the neural networks are self-learning and self-adaptive.
- the neural network can automatically adjust weights of interconnections, so as to give a reasonable expected output under certain input conditions.
- the neural network has good fault tolerance, and local errors do not make the neural network produce more serious errors.
- the neural network has the ability to solve practical problems by constantly adjusting weights among respective neuron nodes.
- calcium Compared with alloys of aluminum, silicon and manganese, calcium has lower melting point and boiling point, a lower solubility in steel and a higher vapor pressure. Therefore, a calcium alloy is difficult to be added into the molten steel, and the calcium treatment process is difficult to control a calcium content in the molten steel within a certain range.
- the calcium treatment process is fed with calcium based on experience, and thus a yield of the calcium treatment process cannot be predicted, so that the control of calcium content is unstable. Therefore, how to predict a yield of calcium in the calcium treatment process based on deep neural network calculation is of great research value and significance for accurately controlling the calcium content in steel, stabilizing the calcium treatment operation, reducing production and operation costs of enterprises, and improving the product quality and alloy utilization efficiency.
- the invention provides a method for predicting a yield of calcium in a calcium treatment process based on deep neural network, for realizing the stable control of the calcium treatment operation consequently.
- an embodiment of the invention provides a method for predicting a yield of calcium in a calcium treatment process based on deep neural network, including:
- S 1 obtaining parameters affecting the yield of calcium in the calcium treatment process from production and operation data information, and thereby constructing a dataset; wherein the S 1 specifically includes:
- ⁇ 1 W ⁇ ( ⁇ [ Ca ] R - ⁇ [ Ca ] o ) ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ 10 3 ,
- ⁇ 2 W ⁇ ( ⁇ [ Ca ] T - ⁇ [ Ca ] o ) ⁇ ⁇ ⁇ ⁇ ⁇ 1 ⁇ 0 3 ,
- ⁇ 3 W ⁇ ( ⁇ [ Ca ] B - ⁇ [ Ca ] o ) ⁇ ⁇ ⁇ ⁇ ⁇ 1 ⁇ 0 3 ,
- W represents a molten steel weight, with a unit of ton (t); ⁇ [Ca] o represents a calcium content of molten steel before calcium treatment, with a unit of parts per million (ppm); ⁇ represents a calcium wire feeding length, with a unit of meter (m); ⁇ represents a calcium content of calcium wire, with a unit of %; ⁇ represents a meter weight of calcium wire, with a unit of kilogram per meter (kg/m); ⁇ [Ca] R represents a calcium content of molten steel after the refining process, with a unit of ppm; ⁇ [Ca] T represents a calcium content of molten steel in a tundish during continuous casting, with a unit of ppm; and ⁇ [Ca] B represents a calcium content in a continuous casting billet, with a unit of ppm;
- S 3 building a deep neural network, comprising: dividing the normalized dataset obtained in the S 2 into a training dataset and a testing dataset, using production and operation data information as input of the deep neural network and using an actual yield of calcium as output of the deep neural network, comparing a calculation result with an actual result, modifying weights and thresholds, and using data in the training dataset to train the deep neural network, thereby obtaining a trained deep neural network;
- the parameters affecting the yield of calcium in the calcium treatment process in the S 1 include: a carbon (C) content of molten steel, a silicon (Si) content of molten steel, a manganese (Mn) content of molten, a phosphorus (P) content of molten steel, a sulfur (S) content of molten steel, a free-oxygen content of molten steel, a molten steel temperature, a calcium wire type, a calcium wire feeding speed, an argon blowing flow rate, a molten steel weight, a calcium wire feeding length, type and quantity of raw and auxiliary materials added in the refining process, and a calcium content of calcium wire.
- a carbon (C) content of molten steel a silicon (Si) content of molten steel, a manganese (Mn) content of molten, a phosphorus (P) content of molten steel, a sulfur (S) content of molten steel, a free-oxygen content of mol
- the composition of molten steel before calcium treatment is detected through spark direct reading spectrometer method, including: firstly polishing a sample, then exciting at least two points on a surface of polished sample, observing a composition of each of excited points and deleting the point with larger composition deviation, until a stable detection result is obtained; and
- an oxygen probe is used to measure the molten steel temperature and a content of dissolved oxygen when molten steel enters a station after calcium treatment, and the oxygen probe is inserted into the same position of molten steel.
- training of the deep neural network in the S 3 includes following sub-steps:
- X represents a normalized X variable
- w represents a weight of connection between the neurons
- ⁇ represents a threshold of neuron node
- the deep neural network built in the S 3 has an input layer containing 21 nodes, has 3 middle layers each of which contains 6 nodes, has an output layer containing one node, has a maximum number of iterations of 1500, and has a learning rate of 0.2. In some embodiments, the deep neural network built in the S 3 has an input layer containing 15-25 nodes, has 1-6 middle layers each of which contains 1-10 nodes, has an output layer containing one node, has a maximum number of iterations of 1500, and has a learning rate of 0.1-0.5.
- the minimum threshold of the error is 5%.
- the invention may achieve the following beneficial effects:
- a dataset is constructed by obtaining operation parameter information of each production charge in advance, a deep neural network model is built based on constructed dataset as a prediction model, production operation information of each charge are obtained, and prediction is carried out based on deep neural network to obtain a reasonable yield of calcium in current charge as a reference basis for calcium treatment operation.
- this model to predict and calculate the yield of calcium, the yield of calcium can be stably controlled and improved.
- the yield of calcium can be increased from 10%-20% to 30% or more, which can reduce the feeding amount of calcium and thereby reduce the production cost.
- the reasonable feeding amount of calcium can also reduce the occurrence of nozzle clogging, thereby ensuring the smooth running of continuous casting process, and improving the production efficiency and the product quality.
- FIG. 1 illustrates a flowchart of a method for predicting a yield of calcium in a calcium treatment process of molten steel based on deep neural according to an embodiment of the invention.
- FIG. 2 illustrates a schematic diagram of showing comparison of yields calculated based on a deep neural network model with actual results.
- a method for predicting a yield of calcium in a calcium treatment process of molten steel based on deep neural network is provided.
- the method for predicting a yield of calcium in a calcium treatment process of molten steel may include step 1 through step 5 as follows.
- Step 1 obtaining production and operation data information of a steel grade employing a calcium treatment process of a domestic factory in recent one year in advance, including a carbon (C) content of molten steel, a silicon (Si) content of molten steel, a manganese (Mn) content of molten, a phosphorus (P) content of molten steel, a sulfur (S) content of molten steel, a free-oxygen content of molten steel, a molten steel temperature, a calcium wire type, a calcium wire feeding speed, an argon blowing flow rate, a molten steel weight, a calcium wire feeding length, type and quantity of raw and auxiliary materials added in a refining process, and a calcium content of calcium wire, totaling 561 groups of data; calculating yields of calcium in the calcium treatment process and constructing a dataset.
- C carbon
- Si silicon
- Mn manganese
- P phosphorus
- S sulfur
- S free-oxygen content of molten steel
- the step 1 may specifically include the sub-step S 11 as follows:
- a calculation formula of the yield of calcium in the refining process is that:
- ⁇ 1 W ⁇ ( ⁇ [ Ca ] R - ⁇ [ Ca ] o ) ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ 10 3 .
- a calculation formula of a yield of calcium of a tundish is that:
- ⁇ 2 W ⁇ ( ⁇ [ Ca ] T - ⁇ [ Ca ] o ) ⁇ ⁇ ⁇ ⁇ ⁇ 1 ⁇ 0 3 .
- a calculation formula of a yield of calcium of a continuous casting billet is that:
- ⁇ 3 W ⁇ ( ⁇ [ Ca ] B - ⁇ [ Ca ] o ) ⁇ ⁇ ⁇ ⁇ ⁇ 1 ⁇ 0 3 .
- W represents the molten steel weight, with a unit of t (ton); ⁇ [Ca] o represents the calcium content of molten steel before calcium treatment, with a unit of ppm (abbreviation for parts per million); ⁇ represents the calcium wire feeding length, with a unit of m (meter); ⁇ represents the calcium content of calcium wire, with a unit of % (percentage); ⁇ represents a meter weight of calcium wire, with a unit of kg/m (kilogram per meter); ⁇ [Ca] R represents the calcium content of molten steel after refining, with a unit of ppm; ⁇ [Ca] T represents the calcium content of molten steel in a tundish during continuous casting, with a unit of ppm; and, ⁇ [Ca] B represents the calcium content in a continuous casting billet, with a unit of ppm.
- ⁇ [Ca] o represents the calcium content of molten steel before calcium treatment, with a unit of ppm (
- Step 2 performing normalization processing on the dataset obtained in the step 1 to obtain a normalized dataset.
- a normalization method uses the following formula:
- y min and y max respectively represent normalized minimum value and maximum value, and respectively may be ⁇ 1 and 1 in the illustrated embodiment
- X min and X max respectively represent a minimum value and a maximum value of a variable X.
- Step 3 dividing the dataset obtained by the step 2 into a training dataset and a testing dataset, randomly selecting 511 groups of data as the training dataset and 50 groups of data as the testing dataset, building a deep neural network by using data such as composition data of molten steel and operation conditions of calcium feeding as input and actual yields of calcium as output, and training built deep neural network by the data in the training dataset.
- Step 4 testing trained deep neural network obtained in the step 3 by the testing dataset.
- input data in the testing dataset are used as input data of the deep neural network to obtain calculation results of yield of calcium
- an error generated by actual results of yield of calcium and the calculation results of yield of calcium of the trained deep neural network reaching a minimum threshold is taken as an optimization objective, and the minimum threshold of the optimization objective of the error is preferably set as 5%.
- a calculation formula for the error generated by the actual results of yield of calcium and the calculation results of yield of calcium of the trained deep neural network is that:
- Y Cal represents the yield of calcium calculated by the deep neural network
- Y Exp represents the actual yield of calcium
- n represents a quantity of selected testing data.
- Step 5 determining the trained deep neural network whether meets a requirement or not based on the error calculated in the step 4; if yes, using current trained deep neural network as a prediction model of the yield of calcium; whereas if not, modifying structural parameters of the trained deep neural network, and repeating the step 4 until the yield of calcium calculated by the trained depth neural network meets a predetermined requirement.
- FIG. 2 illustrated a comparison result of yields of calcium predicted by the trained deep neural network with actual yields of calcium, and an error between current prediction result of yield of calcium and the actual yield of calcium is less than 5%.
- the structural parameters of the deep neural network include a hidden layer quantity, a node quantity in each hidden layer, and a learning rate.
- Step 6 predicting a yield of calcium in an actual calcium treatment process according to the trained deep neural network model obtained by the step 5.
- the obtaining of the data in advance and the constructing of the dataset in the step 1 may include the following step:
- the information of various parameters of each the production charge may include: contents of C, Si, Mn, P, S, Al, O, and Ca of molten steel before calcium treatment, a molten steel temperature, a molten steel weight, weights of a modifier, a desulfurizer, a wax, and lime added during refining, a calcium wire feeding speed, and an argon blowing flow rate, etc.
- Samplings may be carried out before and after calcium treatment to test the composition of molten steel.
- a sampler shall be placed at a same position of a steel ladle during the samplings, to ensure the samplings at the same position.
- the composition of molten steel before the calcium treatment may be detected by spark direct-reading spectrometer method, which may include: firstly, polishing a sample to be bright, and then exciting at least two points on a surface of the sample and observing compositions of the excited points and deleting the point with larger composition deviation, until a stable detection result is obtained.
- spark direct-reading spectrometer method may include: firstly, polishing a sample to be bright, and then exciting at least two points on a surface of the sample and observing compositions of the excited points and deleting the point with larger composition deviation, until a stable detection result is obtained.
- an oxygen probe is used to measure the molten steel temperature and the content of dissolved oxygen, and the oxygen probe is ensured to be inserted into the same position of the molten steel.
- X represents the normalized variable X
- w represents a weight of connection between neurons
- ⁇ represents a threshold of neuron node
- the input layer contains 21 nodes, there is three middle layers and each the middle layer contains 6 nodes, and the output layer contains one node.
- the maximum number of iterations may be 1500, and the learning rate may be 0.2.
- the deep neural network model is trained based on the constructed dataset until the error of the deep neural network model stabilizes at a small value, indicating that the deep neural network has converged.
- a yield of calcium under current production conditions can be predicted after obtaining relevant operation parameter information of each charge before calcium treatment.
- a yield of calcium in a calcium treatment process of molten steel can be predicted by collecting operation parameter information before calcium treatment of molten steel, which can more stably control a calcium content in the calcium treatment process, and thereby improve the calcium treatment effect, reduce the production cost, and improve the production efficiency and product quality.
Abstract
Description
- The invention relates to the field of molten steel refining in iron and steel metallurgy, in particularly to a method for predicting a yield of calcium in a calcium treatment process based on deep neural network.
- In a smelting process of molten steel, in order to effectively reduce an oxygen content in molten steel to a low level, aluminum as a strong deoxidizer is widely used in a steelmaking process. However, an addition of aluminum would produce a large number of alumina inclusions, which would easily lead to nozzle clogging, affect a smooth running of continuous casting process and lead to the degradation of product performance. Therefore, calcium is added into the molten steel to modify the alumina inclusions in the molten steel into liquid calcium aluminate, thereby reduce the nozzle clogging, ensure smooth running of continuous casting and improve the product quality. Moreover, the presence of calcium in steel can also control the morphology and quantity of MnS inclusions in steel. However, there is a reasonable range of calcium requirement in steel. If an addition amount of calcium is low, which cannot achieve the effect of calcium treatment; whereas if the addition amount of calcium is too high, it is easy to produce high melting point CaS inclusions, which also easily lead to nozzle clogging.
- In recent years, benefiting from the rapid development of computer and big data technology, neural networks have been widely used in various industries. The neural networks are self-learning and self-adaptive. When the environment changes, that is, when a new training sample is given to the neural network, the neural network can automatically adjust weights of interconnections, so as to give a reasonable expected output under certain input conditions. Moreover, the neural network has good fault tolerance, and local errors do not make the neural network produce more serious errors. The neural network has the ability to solve practical problems by constantly adjusting weights among respective neuron nodes.
- Compared with alloys of aluminum, silicon and manganese, calcium has lower melting point and boiling point, a lower solubility in steel and a higher vapor pressure. Therefore, a calcium alloy is difficult to be added into the molten steel, and the calcium treatment process is difficult to control a calcium content in the molten steel within a certain range. For most iron and steel enterprises, the calcium treatment process is fed with calcium based on experience, and thus a yield of the calcium treatment process cannot be predicted, so that the control of calcium content is unstable. Therefore, how to predict a yield of calcium in the calcium treatment process based on deep neural network calculation is of great research value and significance for accurately controlling the calcium content in steel, stabilizing the calcium treatment operation, reducing production and operation costs of enterprises, and improving the product quality and alloy utilization efficiency.
- In order to solve the shortcomings of the related art, the invention provides a method for predicting a yield of calcium in a calcium treatment process based on deep neural network, for realizing the stable control of the calcium treatment operation consequently.
- Specifically, an embodiment of the invention provides a method for predicting a yield of calcium in a calcium treatment process based on deep neural network, including:
- S1, obtaining parameters affecting the yield of calcium in the calcium treatment process from production and operation data information, and thereby constructing a dataset; wherein the S1 specifically includes:
- S11, collecting production and operation data information in a refining process of each production charge and calculating the yield of calcium of each the production charge, as a piece of record; wherein the yield of calcium of each the production charge includes a yield of calcium in the refining process, a yield of calcium η2 of tundish, and a yield of calcium η3 of continuous casting billet;
- a calculation formula for the yield of calcium in the refining process is that:
-
- a calculation formula for the yield of calcium of tundish is that:
-
- a calculation formula for the yield of calcium of continuous casting billet is that:
-
- where, W represents a molten steel weight, with a unit of ton (t); ω[Ca]o represents a calcium content of molten steel before calcium treatment, with a unit of parts per million (ppm); χ represents a calcium wire feeding length, with a unit of meter (m); β represents a calcium content of calcium wire, with a unit of %; μ represents a meter weight of calcium wire, with a unit of kilogram per meter (kg/m); ω[Ca]R represents a calcium content of molten steel after the refining process, with a unit of ppm; ω[Ca]T represents a calcium content of molten steel in a tundish during continuous casting, with a unit of ppm; and ω[Ca]B represents a calcium content in a continuous casting billet, with a unit of ppm;
- S12, repeatedly collecting the production and operation data information in the refining process of each production charge, and thereby establishing a dataset; and
- S13, preprocessing the dataset established in the S12 to remove incomplete data and unreasonable data;
- S2, performing normalization processing on the dataset constructed in the S1 to obtain a normalized dataset;
- S3, building a deep neural network, comprising: dividing the normalized dataset obtained in the S2 into a training dataset and a testing dataset, using production and operation data information as input of the deep neural network and using an actual yield of calcium as output of the deep neural network, comparing a calculation result with an actual result, modifying weights and thresholds, and using data in the training dataset to train the deep neural network, thereby obtaining a trained deep neural network;
- S4, using the testing dataset to test the trained deep neural network obtained in the S3, using input data in the testing dataset as input data of the trained deep neural network to obtain calculation results of yield of calcium, and taking an error generated by actual results of yield of calcium and the calculation results of yield of calcium of the trained deep neural network reaches a minimum threshold as an optimization objective to optimize the trained deep neural network;
- S5, determining, based on the error in the S4, the trained deep neural network whether meets a requirement or not; in response to the error reaches the minimum threshold, taking current trained deep neural network as a resultant prediction model of yield of calcium; in response to the error does not reach the minimum threshold, modifying a hidden layer quantity, a node quantity, and a learning rate of the trained deep neural network, and repeating the S4 until the error reaches the minimum threshold which indicates a yield of calcium calculated by the trained deep neural network meets a predetermined requirement; and
- S6, predicting a yield of calcium in an actual calcium treatment process based on the resultant prediction model of yield of calcium obtained in the S5.
- In a preferred embodiment, the parameters affecting the yield of calcium in the calcium treatment process in the S1 include: a carbon (C) content of molten steel, a silicon (Si) content of molten steel, a manganese (Mn) content of molten, a phosphorus (P) content of molten steel, a sulfur (S) content of molten steel, a free-oxygen content of molten steel, a molten steel temperature, a calcium wire type, a calcium wire feeding speed, an argon blowing flow rate, a molten steel weight, a calcium wire feeding length, type and quantity of raw and auxiliary materials added in the refining process, and a calcium content of calcium wire.
- In a preferred embodiment, during the collecting in the S11, performing samplings before and after calcium treatment to test a composition of molten steel, while placing a sampler into a same position of a steel ladle during the samplings;
- during detecting a composition of molten steel in the tundish, performing samplings at a same position of the tundish at half of pouring; and
- during detecting a composition of continuous casting billet, taking a continuous casting billet of stabilized pouring, and sampling at a position of ¼ of a plate width and ¼ distance from an inner arc side for analysis.
- In a preferred embodiment, the composition of molten steel before calcium treatment is detected through spark direct reading spectrometer method, including: firstly polishing a sample, then exciting at least two points on a surface of polished sample, observing a composition of each of excited points and deleting the point with larger composition deviation, until a stable detection result is obtained; and
- an oxygen probe is used to measure the molten steel temperature and a content of dissolved oxygen when molten steel enters a station after calcium treatment, and the oxygen probe is inserted into the same position of molten steel.
- In a preferred embodiment, training of the deep neural network in the S3 includes following sub-steps:
- S31, initializing weights and thresholds of each layer by using a small random number; calculating, backwards from a first layer of network, outputs of neurons according to set network structure and weights and thresholds of a preceding iteration; then modifying the weights and the thresholds, calculating, forwards from a last layer, an influence of the thresholds and the weights to a total error, and adjusting the weights and the thresholds based on an error; wherein the two processes are alternately performed until convergence is reached; and
- S32, a transfer function between the neurons using Sigmoid function and employing the following formulas:
-
- where, X represents a normalized X variable, w represents a weight of connection between the neurons, and θ represents a threshold of neuron node.
- In a preferred embodiment, the deep neural network built in the S3 has an input layer containing 21 nodes, has 3 middle layers each of which contains 6 nodes, has an output layer containing one node, has a maximum number of iterations of 1500, and has a learning rate of 0.2. In some embodiments, the deep neural network built in the S3 has an input layer containing 15-25 nodes, has 1-6 middle layers each of which contains 1-10 nodes, has an output layer containing one node, has a maximum number of iterations of 1500, and has a learning rate of 0.1-0.5.
- In a preferred embodiment, the minimum threshold of the error is 5%.
- Compared with the related art, the invention may achieve the following beneficial effects:
- In the method for predicting the yield of calcium in the calcium treatment process based on deep neural network, a dataset is constructed by obtaining operation parameter information of each production charge in advance, a deep neural network model is built based on constructed dataset as a prediction model, production operation information of each charge are obtained, and prediction is carried out based on deep neural network to obtain a reasonable yield of calcium in current charge as a reference basis for calcium treatment operation. By using this model to predict and calculate the yield of calcium, the yield of calcium can be stably controlled and improved. For some steel grades, by adjusting operation parameters of calcium feeding, the yield of calcium can be increased from 10%-20% to 30% or more, which can reduce the feeding amount of calcium and thereby reduce the production cost. The reasonable feeding amount of calcium can also reduce the occurrence of nozzle clogging, thereby ensuring the smooth running of continuous casting process, and improving the production efficiency and the product quality.
-
FIG. 1 illustrates a flowchart of a method for predicting a yield of calcium in a calcium treatment process of molten steel based on deep neural according to an embodiment of the invention. -
FIG. 2 illustrates a schematic diagram of showing comparison of yields calculated based on a deep neural network model with actual results. - Illustrative embodiments, features, and aspects of the invention will be described in detail below with reference to the accompanying drawings. Like reference numerals in the drawings indicate functionally identical or similar elements. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise indicated.
- According to an embodiment of the invention, a method for predicting a yield of calcium in a calcium treatment process of molten steel based on deep neural network is provided.
- As shown in
FIG. 1 , the method for predicting a yield of calcium in a calcium treatment process of molten steel according to the illustrated embodiment may include step 1 through step 5 as follows. - Step 1, obtaining production and operation data information of a steel grade employing a calcium treatment process of a domestic factory in recent one year in advance, including a carbon (C) content of molten steel, a silicon (Si) content of molten steel, a manganese (Mn) content of molten, a phosphorus (P) content of molten steel, a sulfur (S) content of molten steel, a free-oxygen content of molten steel, a molten steel temperature, a calcium wire type, a calcium wire feeding speed, an argon blowing flow rate, a molten steel weight, a calcium wire feeding length, type and quantity of raw and auxiliary materials added in a refining process, and a calcium content of calcium wire, totaling 561 groups of data; calculating yields of calcium in the calcium treatment process and constructing a dataset.
- The step 1 may specifically include the sub-step S11 as follows:
- S11, collecting production and operation data information in a refining process of each production charge and calculating a yield of calcium of each production charge, as a piece of record.
- A calculation formula of the yield of calcium in the refining process is that:
-
- A calculation formula of a yield of calcium of a tundish is that:
-
- A calculation formula of a yield of calcium of a continuous casting billet is that:
-
- Where, W represents the molten steel weight, with a unit of t (ton); ω[Ca]o represents the calcium content of molten steel before calcium treatment, with a unit of ppm (abbreviation for parts per million); χ represents the calcium wire feeding length, with a unit of m (meter); β represents the calcium content of calcium wire, with a unit of % (percentage); μ represents a meter weight of calcium wire, with a unit of kg/m (kilogram per meter); ω[Ca]R represents the calcium content of molten steel after refining, with a unit of ppm; ω[Ca]T represents the calcium content of molten steel in a tundish during continuous casting, with a unit of ppm; and, ω[Ca]B represents the calcium content in a continuous casting billet, with a unit of ppm.
- Step 2, performing normalization processing on the dataset obtained in the step 1 to obtain a normalized dataset.
- For the normalization processing of the dataset in the step 2, a normalization method uses the following formula:
-
- where, ymin and ymax respectively represent normalized minimum value and maximum value, and respectively may be −1 and 1 in the illustrated embodiment; Xmin and Xmax respectively represent a minimum value and a maximum value of a variable X.
- Step 3, dividing the dataset obtained by the step 2 into a training dataset and a testing dataset, randomly selecting 511 groups of data as the training dataset and 50 groups of data as the testing dataset, building a deep neural network by using data such as composition data of molten steel and operation conditions of calcium feeding as input and actual yields of calcium as output, and training built deep neural network by the data in the training dataset.
- Step 4, testing trained deep neural network obtained in the step 3 by the testing dataset. In particular, input data in the testing dataset are used as input data of the deep neural network to obtain calculation results of yield of calcium, an error generated by actual results of yield of calcium and the calculation results of yield of calcium of the trained deep neural network reaching a minimum threshold is taken as an optimization objective, and the minimum threshold of the optimization objective of the error is preferably set as 5%.
- A calculation formula for the error generated by the actual results of yield of calcium and the calculation results of yield of calcium of the trained deep neural network is that:
-
- where, YCal represents the yield of calcium calculated by the deep neural network, YExp represents the actual yield of calcium, and n represents a quantity of selected testing data.
- Step 5, determining the trained deep neural network whether meets a requirement or not based on the error calculated in the step 4; if yes, using current trained deep neural network as a prediction model of the yield of calcium; whereas if not, modifying structural parameters of the trained deep neural network, and repeating the step 4 until the yield of calcium calculated by the trained depth neural network meets a predetermined requirement.
FIG. 2 illustrated a comparison result of yields of calcium predicted by the trained deep neural network with actual yields of calcium, and an error between current prediction result of yield of calcium and the actual yield of calcium is less than 5%. - In an embodiment, the structural parameters of the deep neural network include a hidden layer quantity, a node quantity in each hidden layer, and a learning rate.
- Step 6, predicting a yield of calcium in an actual calcium treatment process according to the trained deep neural network model obtained by the step 5.
- The obtaining of the data in advance and the constructing of the dataset in the step 1 may include the following step:
- collecting information of various parameters in the refining process of each production charge, as a piece of record;
- wherein the information of various parameters of each the production charge may include: contents of C, Si, Mn, P, S, Al, O, and Ca of molten steel before calcium treatment, a molten steel temperature, a molten steel weight, weights of a modifier, a desulfurizer, a carburant, and lime added during refining, a calcium wire feeding speed, and an argon blowing flow rate, etc.
- Samplings may be carried out before and after calcium treatment to test the composition of molten steel. In order to ensure the stability of sampling results, a sampler shall be placed at a same position of a steel ladle during the samplings, to ensure the samplings at the same position.
- The composition of molten steel before the calcium treatment may be detected by spark direct-reading spectrometer method, which may include: firstly, polishing a sample to be bright, and then exciting at least two points on a surface of the sample and observing compositions of the excited points and deleting the point with larger composition deviation, until a stable detection result is obtained. When the molten steel enters a station, an oxygen probe is used to measure the molten steel temperature and the content of dissolved oxygen, and the oxygen probe is ensured to be inserted into the same position of the molten steel.
- Specific steps of training the deep neural network in the step 3 may be as follows:
- initializing weights and thresholds of each layer by using a small random number, calculating, backwards from the first layer of the network, outputs of respective neurons through set network structure and weights and thresholds of the preceding iteration, then modifying the weights and the thresholds, and calculating, forwards from the last layer, an influence of the thresholds and the weights to a total error, and adjusting the weights and the threshold according to an error. The two processes are alternately carried out until convergence is reached; and
- a transfer function between neurons using the Sigmoid function and employing the formulas as follows:
-
- where, X represents the normalized variable X, w represents a weight of connection between neurons, and θ represents a threshold of neuron node.
- For the built deep neural network, the input layer contains 21 nodes, there is three middle layers and each the middle layer contains 6 nodes, and the output layer contains one node. The maximum number of iterations may be 1500, and the learning rate may be 0.2.
- The deep neural network model is trained based on the constructed dataset until the error of the deep neural network model stabilizes at a small value, indicating that the deep neural network has converged. By using the trained deep neural network, a yield of calcium under current production conditions can be predicted after obtaining relevant operation parameter information of each charge before calcium treatment.
- According to the invention, a yield of calcium in a calcium treatment process of molten steel can be predicted by collecting operation parameter information before calcium treatment of molten steel, which can more stably control a calcium content in the calcium treatment process, and thereby improve the calcium treatment effect, reduce the production cost, and improve the production efficiency and product quality.
- Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the invention, rather than limiting the invention. Although the invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions described in the foregoing embodiments can still be modified, or some or all of the technical features thereof can be equivalently substituted; and these modifications or substitutions do not cause essences of the corresponding technical solutions to deviate from scopes of the technical solutions of the embodiments of the invention.
Claims (7)
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110691320.6A CN113343576B (en) | 2021-06-22 | 2021-06-22 | Prediction method of calcium yield in calcium treatment process based on deep neural network |
CN202110691320.6 | 2021-06-22 |
Publications (2)
Publication Number | Publication Date |
---|---|
US20220406413A1 true US20220406413A1 (en) | 2022-12-22 |
US11545239B1 US11545239B1 (en) | 2023-01-03 |
Family
ID=77478703
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US17/844,364 Active US11545239B1 (en) | 2021-06-22 | 2022-06-20 | Method for predicting yield of calcium in a calcium treatment process based on deep neural network |
Country Status (3)
Country | Link |
---|---|
US (1) | US11545239B1 (en) |
JP (1) | JP7093090B1 (en) |
CN (1) | CN113343576B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116011149A (en) * | 2023-01-18 | 2023-04-25 | 鞍钢股份有限公司 | Method for determining target moving-out temperature of RH refining furnace |
CN116772662A (en) * | 2023-07-17 | 2023-09-19 | 东方空间技术(山东)有限公司 | Rocket recovery sub-level landing leg control method, computing equipment and storage medium |
CN116822753A (en) * | 2023-08-30 | 2023-09-29 | 北京建筑大学 | Electric leaching repair sediment prediction optimization method and system based on neural network |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115436239B (en) * | 2022-11-07 | 2023-02-07 | 四川亿欣新材料有限公司 | Calcium carbonate granularity detection method |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5327357A (en) * | 1991-12-03 | 1994-07-05 | Praxair Technology, Inc. | Method of decarburizing molten metal in the refining of steel using neural networks |
KR20030047589A (en) * | 2001-12-11 | 2003-06-18 | 주식회사 포스코 | The method of increasing yield of calcium when it is inputted into deoxygenized melting steel |
US20120145353A1 (en) * | 2009-07-15 | 2012-06-14 | Kabushiki Kaisha Kobe Seiko Sho (Kobe Steel, Ltd.) | Method for producing alloy ingot |
CN110777230A (en) * | 2019-11-13 | 2020-02-11 | 北京科技大学 | Steel continuous casting billet solidification structure refining method based on target isometric crystal size and ratio |
US20200257933A1 (en) * | 2019-02-05 | 2020-08-13 | Imagars Llc | Machine Learning to Accelerate Alloy Design |
US20220072593A1 (en) * | 2019-01-17 | 2022-03-10 | Jfe Steel Corporation | Production specification determination method, production method, and production specification determination apparatus for metal material |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP6037014B2 (en) * | 2013-07-10 | 2016-11-30 | Jfeスチール株式会社 | Steel manufacturing method |
US20210027860A1 (en) * | 2019-07-23 | 2021-01-28 | The Florida State University Research Foundation, Inc. | Methods of Designing and Predicting Proteins |
CN110866659B (en) * | 2019-12-09 | 2023-08-15 | 东北大学 | On-line prediction method for caustic ratio in high-pressure pipeline digestion process of Bayer process alumina production |
CN111549286B (en) * | 2020-06-22 | 2021-10-15 | 中南大学 | Easily-welded ultrahigh-strength high-toughness steel and preparation method thereof |
CN111950427B (en) * | 2020-08-06 | 2022-02-15 | 中国地质科学院矿产资源研究所 | Garnet subclass identification method based on thermal infrared spectrum characteristics and BP neural network model |
-
2021
- 2021-06-22 CN CN202110691320.6A patent/CN113343576B/en active Active
-
2022
- 2022-04-28 JP JP2022075169A patent/JP7093090B1/en active Active
- 2022-06-20 US US17/844,364 patent/US11545239B1/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5327357A (en) * | 1991-12-03 | 1994-07-05 | Praxair Technology, Inc. | Method of decarburizing molten metal in the refining of steel using neural networks |
KR20030047589A (en) * | 2001-12-11 | 2003-06-18 | 주식회사 포스코 | The method of increasing yield of calcium when it is inputted into deoxygenized melting steel |
US20120145353A1 (en) * | 2009-07-15 | 2012-06-14 | Kabushiki Kaisha Kobe Seiko Sho (Kobe Steel, Ltd.) | Method for producing alloy ingot |
US20220072593A1 (en) * | 2019-01-17 | 2022-03-10 | Jfe Steel Corporation | Production specification determination method, production method, and production specification determination apparatus for metal material |
US20200257933A1 (en) * | 2019-02-05 | 2020-08-13 | Imagars Llc | Machine Learning to Accelerate Alloy Design |
CN110777230A (en) * | 2019-11-13 | 2020-02-11 | 北京科技大学 | Steel continuous casting billet solidification structure refining method based on target isometric crystal size and ratio |
Non-Patent Citations (1)
Title |
---|
Zhao, Liping., "Intelligent Prediction Method of Quality for Continuous Casting Process"., IEEE Xplore., 978-1-4673-9613-4/16/$31.00 ©2016 IEEE (Year: 2016) * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116011149A (en) * | 2023-01-18 | 2023-04-25 | 鞍钢股份有限公司 | Method for determining target moving-out temperature of RH refining furnace |
CN116772662A (en) * | 2023-07-17 | 2023-09-19 | 东方空间技术(山东)有限公司 | Rocket recovery sub-level landing leg control method, computing equipment and storage medium |
CN116822753A (en) * | 2023-08-30 | 2023-09-29 | 北京建筑大学 | Electric leaching repair sediment prediction optimization method and system based on neural network |
Also Published As
Publication number | Publication date |
---|---|
JP2023002457A (en) | 2023-01-10 |
CN113343576B (en) | 2022-03-11 |
CN113343576A (en) | 2021-09-03 |
US11545239B1 (en) | 2023-01-03 |
JP7093090B1 (en) | 2022-06-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11545239B1 (en) | Method for predicting yield of calcium in a calcium treatment process based on deep neural network | |
CN104630410B (en) | A kind of pneumatic steelmaking quality real-time dynamic forecast method based on data parsing | |
CN109447346B (en) | Converter oxygen consumption prediction method based on gray prediction and neural network combined model | |
WO2018012257A1 (en) | Method for estimating phosphorus concentration in molten steel and converter blowing control device | |
CN114611844B (en) | Method and system for determining alloy addition amount in converter tapping process | |
JP5822053B2 (en) | Control device and control method for converter blowing equipment | |
CN113919206A (en) | Narrow composition control method for steelmaking alloying | |
CN113435114B (en) | Method for accurately controlling nitrogen content of steel grade of CV-RH-CC process route | |
CN113192568B (en) | Method and system for forecasting desulfurization end point of refining furnace | |
JP6579136B2 (en) | Refining process state estimation device, refining process state estimation method, and molten metal manufacturing method | |
CN107630122B (en) | A kind of RH dynamic decarburization optimization method based on flue gas analysis | |
CN104419799A (en) | Method for online prediction of carbon content of high-carbon steel during converter smelting process | |
JP5854171B2 (en) | Correction device, correction method, and steel refining method | |
JP6376200B2 (en) | Molten state estimation device, molten state estimation method, and molten metal manufacturing method | |
CN116469481B (en) | LF refined molten steel composition forecasting method based on XGBoost algorithm | |
CN111518980A (en) | Correction method and system for converter end point carbon content prediction model | |
CN114875196B (en) | Method and system for determining converter tapping quantity | |
CN114841277A (en) | Method and system for predicting yield of converter tapping alloy | |
CN112989687B (en) | Method for forecasting transverse crack defect of continuous casting billet angle based on metallurgical principle | |
Xu et al. | Element yield rate prediction in ladle furnace based on improved GA-ANFIS | |
CN117113836A (en) | Steelmaking carbon dioxide blowing method and system based on double-pressure coupling | |
CN117077997A (en) | Method and system for optimizing molten steel desulfurization process in refining process | |
Wang | Hybrid Models for BOF Steelmaking and Continuous Casting Applications | |
CN114561509A (en) | Method for predicting carbon content of electric arc furnace end point | |
Huang et al. | Prediction of alloy yield based on optimized BP neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: UNIVERSITY OF SCIENCE AND TECHNOLOGY BEIJING, CHINA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:ZHANG, LIFENG;WANG, WEIJIAN;REN, QIANG;AND OTHERS;REEL/FRAME:060251/0534 Effective date: 20220620 Owner name: YANSHAN UNIVERSITY, CHINA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:ZHANG, LIFENG;WANG, WEIJIAN;REN, QIANG;AND OTHERS;REEL/FRAME:060251/0534 Effective date: 20220620 Owner name: NORTH CHINA UNIVERSITY OF TECHNOLOGY, CHINA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:ZHANG, LIFENG;WANG, WEIJIAN;REN, QIANG;AND OTHERS;REEL/FRAME:060251/0534 Effective date: 20220620 |
|
FEPP | Fee payment procedure |
Free format text: ENTITY STATUS SET TO UNDISCOUNTED (ORIGINAL EVENT CODE: BIG.); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY |
|
FEPP | Fee payment procedure |
Free format text: ENTITY STATUS SET TO SMALL (ORIGINAL EVENT CODE: SMAL); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY |
|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |