Process for predicting end-point oxygen content of molten steel and deoxidizing in refining process
By combining deep neural network and particle swarm algorithm, a prediction model of the end point of the steel molten steel is constructed, which solves the problem of difficult control of the oxygen content during the steel molten steel refining process, and realizes the precise control of the oxygen content and the optimization of the deoxygenation process, reducing production costs And improve product quality.
Patent Information
- Application Number
- CN202411490219.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2025-02-25
AI Technical Summary
It is difficult to accurately control the oxygen content in the molten steel during the existing molten steel refining process, resulting in increased consumption of alloys and raw materials and unqualified product quality.
Combining deep neural network technology and particle swarm algorithm, a prediction model of the end point of the steel liquid steel is constructed, and precise control of the oxygen content is achieved through the optimization of the deoxygenation process.
Accurate control of the oxygen content of the molten steel is achieved, the deoxygenation operation process is optimized, the consumption of raw and auxiliary materials is reduced, the production cost is reduced, and the product quality is improved.
Smart Images
Figure CN119517189A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of iron and steel metallurgy, and in particular to a prediction of the endpoint oxygen content of molten steel in a refining process and a deoxidation process based on machine learning technology. Background Art
[0002] The presence of oxygen in steel has an adverse effect on the performance of steel. During the refining process of molten steel, the oxygen content in the molten steel needs to be controlled at a low level. Different steel grades have different application conditions, and their requirements for oxygen content should also be different, so their deoxidation operation processes should also be different. Most of the existing deoxidation processes add sufficient deoxidizing alloys and refining slag agents to the molten steel to ensure a sufficiently low oxygen content, and it is difficult to accurately control the free oxygen content in the molten steel. On the one hand, it increases the consumption of raw and auxiliary materials such as alloys and lime, and on the other hand, it may also cause the oxygen content of some furnaces to fail to meet the requirements of the steel grade, reduce product quality, and even lead to unqualified quality.
[0003] Neural network technology can realize complex nonlinear mapping, has strong learning ability, and has strong adaptability. When the training samples change, the weights and thresholds of the connections between neural networks can be automatically adjusted to give the optimal expected output under the current input conditions. The particle swarm algorithm simulates the predation behavior of birds to optimize the calculation and finally obtain the optimal solution to converge the problem.
[0004] Based on the shortcomings of the deoxidation operation in the steel refining process, the particle swarm algorithm combined with the deep neural network technology is applied to the deoxidation operation in the refining process. It has very important value and significance for accurately controlling the oxygen content in steel, stably improving the deoxidation effect, reducing enterprise production costs, and improving product quality. Summary of the invention
[0005] The purpose of the present invention is to provide a prediction method for the oxygen content of molten steel at the end point of a refining process and a deoxidation process based on machine learning technology to solve the problems raised in the above-mentioned background technology.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] A process for predicting the oxygen content of liquid steel during a refining process and deoxidation, comprising the following steps:
[0008] S1, constructing a data set, obtaining a historical deoxygenation process data set and a corresponding endpoint oxygen content data set; preprocessing the data in the historical data set to obtain a preprocessed data set;
[0009] S2, dividing the preprocessing data set and the historical endpoint oxygen content data set into a training set and a test set;
[0010] S3, taking the deoxidation process as input and the endpoint oxygen content as output, a deep neural network prediction model for the endpoint oxygen content of molten steel during the refining process was constructed;
[0011] S4, using the training set to train the endpoint oxygen content prediction model, using the test set to predict the endpoint oxygen content prediction model, and calculating the error between the predicted value and the actual value of the endpoint oxygen content prediction model;
[0012] S5, judging the calculated prediction error value, if the prediction error value is less than the set threshold, obtaining the trained endpoint oxygen content prediction model; if the prediction error value is greater than or equal to the set threshold, adjusting the parameters such as the number of input layers, the number of intermediate layers, the number of nodes, and the learning rate of the endpoint oxygen content prediction model and returning to S4;
[0013] S6, obtaining an initial deoxidation process, and optimizing the initial deoxidation process based on a particle swarm algorithm; the number of iterations of the particle swarm algorithm is 50-200, and the population size is 10-200;
[0014] S7, inputting a plurality of optimized deoxygenation processes into the trained endpoint oxygen content prediction model to obtain a plurality of endpoint oxygen content prediction values;
[0015] S8, taking the optimized deoxidation process corresponding to the minimum value of the difference between the several endpoint oxygen content prediction values and the expected oxygen content as the optimal deoxidation process.
[0016] As a preferred technical solution, the above S1 also includes the following specific steps:
[0017] S11, collecting the deoxidation process in the refining process to obtain historical deoxidation process data, and obtaining the endpoint oxygen content to obtain historical endpoint oxygen content data;
[0018] S12, repeatedly executing S11 to obtain a historical deoxidation process data set and a historical endpoint oxygen content data set;
[0019] S13, deleting missing, repeated and erroneous historical deoxidation process data in the historical deoxidation process data set to obtain a deleted data set, and normalizing the data in the deleted data set to obtain a preprocessed data set;
[0020] S14, deleting the historical endpoint oxygen content data corresponding to the deleted historical deoxygenation process data in the initial historical endpoint oxygen content data set to obtain a historical endpoint oxygen content data set.
[0021] As a preferred technical solution, the normalization processing formula in the above S13 is as follows:
[0022]
[0023] Where: y min is the normalized minimum value, taking -1, y max is the normalized maximum value, taking 1, X min is the minimum value of X, X max is the maximum value of X, y is the normalized value, and X is the data before normalization.
[0024] As a preferred technical solution, in S3, the transfer function between the nodes of the endpoint oxygen content prediction model uses a Sigmoid function, and the Sigmoid function is as follows:
[0025]
[0026] In the formula, x j is the input of the endpoint oxygen content prediction model, w ij represents the weight of the connection between the i-th node and the j-th node, θ i Represents the threshold of the i-th node. j is the output value of the neuron node, and n is the number of input variables of the neural network node.
[0027] As a preferred technical solution, in S4, the error between the predicted value and the actual value of the endpoint oxygen content prediction model includes a root mean square error and a mean absolute error; the root mean square error and the mean absolute error are calculated as follows:
[0028]
[0029] Where: X Cal. is the endpoint oxygen content predicted by the endpoint oxygen content prediction model, X Exp. is the actual value of the endpoint oxygen content, n is the number of data in the test set, RMSE is the root mean square error, and MAE is the mean absolute error.
[0030] Compared with the prior art, the present invention has the following beneficial effects:
[0031] The present invention discloses a method for predicting the endpoint oxygen content of molten steel in the refining process and optimizing the deoxidation process based on machine learning technology, and applies a particle swarm algorithm combined with a deep neural network technology to the deoxidation operation in the refining process, thereby achieving precise control of the final oxygen content, optimizing the deoxidation operation process, reducing the amount of raw and auxiliary materials such as lime added, controlling the vacuum refining time, reducing production costs, and improving product quality. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 It is a flow chart of the prediction of key oxygen content in molten steel and the deoxidation process in the refining process of the present invention;
[0033] Figure 2It is a schematic diagram comparing the predicted value and the actual value based on the endpoint oxygen content prediction model of the present invention. DETAILED DESCRIPTION
[0034] like Figure 1 , 2 As shown, the present invention provides a method for predicting the key oxygen content of molten steel in a refining process and a deoxidation process based on machine learning technology, which includes:
[0035] S1, constructing a data set, obtaining a historical deoxygenation process data set and a corresponding endpoint oxygen content data set; preprocessing the data in the historical data set to obtain a preprocessed data set;
[0036] Specifically, S1 includes:
[0037] S11, collect the deoxidation process in the refining process, obtain the historical deoxidation process data, and obtain the endpoint oxygen content to obtain the historical endpoint oxygen content data. During the data collection process, ensure that the sampler is inserted at the same position and the same depth of the ladle to ensure the stability of the test results; when using a direct reading spectrometer to detect the composition of the molten steel, first polish the surface of the sample to make it bright, and detect the composition of at least two positions on the surface of the sample. If the composition results of the two positions are very different, continue to detect the composition of the third position, and select two similar components as the sample composition. When using an oxygen probe to measure the temperature and dissolved oxygen of the molten steel, ensure that the oxygen probe is inserted at the same position in the ladle.
[0038] S12, repeatedly executing S11 to obtain a historical deoxidation process data set and a historical endpoint oxygen content data set.
[0039] S13, deleting missing, repeated and erroneous historical deoxidation process data in the historical deoxidation process data set to obtain a deleted data set, and normalizing the data in the deleted data set to obtain a preprocessed data set.
[0040] The normalization process is to normalize each data in the deleted data set separately. The formula is as follows:
[0041]
[0042] Where: y min is the normalized minimum value, taking -1, y max is the normalized maximum value, taking 1, X min is the minimum value of X, X max is the maximum value of X, y is the normalized value, and X is the data before normalization.
[0043] S14, deleting the historical endpoint oxygen content data corresponding to the deleted historical deoxygenation process data in the initial historical endpoint oxygen content data set to obtain a historical endpoint oxygen content data set.
[0044] Optionally, the deoxidation process includes the composition and temperature of the molten steel, the vacuum refining pure degassing time and degassing time, the circulation and bottom blowing argon flow, the weight of the molten steel, and the types and quantities of raw and auxiliary materials added during the refining process. The added raw and auxiliary materials include modifiers, deoxidizers, recarburizers, and lime quality.
[0045] S2, dividing the preprocessing data set and the historical endpoint oxygen content data set into a training set and a test set;
[0046] S3, taking the deoxidation process as input and the endpoint oxygen content as output, a deep neural network prediction model for the endpoint oxygen content of molten steel during the refining process was constructed;
[0047] The transfer function between nodes of the endpoint oxygen content prediction model uses the Sigmoid function, and the Sigmoid function is as follows:
[0048]
[0049] In the formula, x j is the input of the endpoint oxygen content prediction model, w ij represents the weight of the connection between the i-th node and the j-th node, θ i Represents the threshold of the i-th node. j is the output value of the neuron node, and n is the number of input variables of the neural network node.
[0050] S4, using the training set to train the endpoint oxygen content prediction model, using the test set to predict the endpoint oxygen content prediction model, and calculating the error between the predicted value and the actual value of the endpoint oxygen content prediction model;
[0051] In this embodiment, the prediction error value includes the root mean square error and the mean absolute error. The root mean square error and the mean absolute error are calculated as follows:
[0052]
[0053] Where: X Cal. is the endpoint oxygen content predicted by the endpoint oxygen content prediction model, X Exp. is the actual value of the endpoint oxygen content, n is the number of data in the test set, RMSE is the root mean square error, and MAE is the mean absolute error.
[0054] S5, judging the calculated prediction error value, if the prediction error value is less than the set threshold, obtaining the trained endpoint oxygen content prediction model; if the prediction error value is greater than or equal to the set threshold, adjusting the parameters such as the number of input layers, the number of intermediate layers, the number of nodes, and the learning rate of the endpoint oxygen content prediction model and returning to S4;
[0055] Preferably, the input layer of the endpoint oxygen content prediction model is 10-30 nodes; the number of intermediate layers is 1-10, each intermediate layer is 4-20 nodes; the output layer is 1 node, the maximum number of iterations is 1000-2000, and the learning rate is 0.1-0.5. Specifically, the input layer of the endpoint oxygen content prediction model is 25 nodes; the number of intermediate layers is 3, each intermediate layer is 8 nodes; the output layer is 1 node, the maximum number of iterations is 1500, the learning target is 0.0004, and the learning rate is 0.2.
[0056] The threshold is set to 5 ppm. When the root mean square error and the mean absolute error are both less than 5 ppm, the trained endpoint oxygen content prediction model is obtained.
[0057] The trained endpoint oxygen content prediction model was tested, and the results were as follows: Figure 2 As shown, from Figure 2 It can be seen that the error value between the predicted endpoint oxygen content and the actual endpoint oxygen content is less than 5ppm.
[0058] S6, obtaining an initial deoxidation process, and optimizing the initial deoxidation process based on a particle swarm algorithm. The number of iterations of the particle swarm algorithm is 50-200, and the population size is 10-200.
[0059] S7, inputting a plurality of optimized deoxygenation processes into the trained endpoint oxygen content prediction model to obtain a plurality of endpoint oxygen content prediction values.
[0060] S8, taking the optimized deoxidation process corresponding to the minimum value of the difference between the several endpoint oxygen content prediction values and the expected oxygen content as the optimal deoxidation process.
[0061] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A process for predicting the oxygen content of liquid steel during a refining process and for deoxidation, characterized in that: The steps include: S1, constructing a data set, obtaining a historical deoxygenation process data set and a corresponding endpoint oxygen content data set; preprocessing the data in the historical data set to obtain a preprocessed data set; S2, dividing the preprocessing data set and the historical endpoint oxygen content data set into a training set and a test set; S3, taking the deoxidation process as input and the endpoint oxygen content as output, a deep neural network prediction model for the endpoint oxygen content of molten steel during the refining process was constructed; S4, using the training set to train the endpoint oxygen content prediction model, using the test set to predict the endpoint oxygen content prediction model, and calculating the error between the predicted value and the actual value of the endpoint oxygen content prediction model; S5, judging the calculated prediction error value, if the prediction error value is less than the set threshold, obtaining the trained endpoint oxygen content prediction model; if the prediction error value is greater than or equal to the set threshold, adjusting the number of input layers, the number of intermediate layers, the number of nodes and the learning rate of the endpoint oxygen content prediction model and returning to S4; S6, obtaining an initial deoxidation process, and optimizing the initial deoxidation process based on a particle swarm algorithm; The number of iterations of the particle swarm algorithm is 50-200, and the population size is 10-200; S7, inputting a plurality of optimized deoxygenation processes into the trained endpoint oxygen content prediction model to obtain a plurality of endpoint oxygen content prediction values; S8, taking the optimized deoxidation process corresponding to the minimum value of the difference between the several endpoint oxygen content prediction values and the expected oxygen content as the optimal deoxidation process.
2. The process for predicting the oxygen content of liquid steel during refining and deoxidation according to claim 1, characterized in that: S1 includes the following steps: S11, collecting the deoxidation process in the refining process to obtain historical deoxidation process data, and obtaining the endpoint oxygen content to obtain historical endpoint oxygen content data; S12, repeatedly executing S11 to obtain a historical deoxidation process data set and a historical endpoint oxygen content data set; S13, deleting missing, repeated and erroneous historical deoxidation process data in the historical deoxidation process data set to obtain a deleted data set, and normalizing the data in the deleted data set to obtain a preprocessed data set; S14, deleting the historical endpoint oxygen content data corresponding to the deleted historical deoxygenation process data in the initial historical endpoint oxygen content data set to obtain a historical endpoint oxygen content data set.
3. The process for predicting the oxygen content of liquid steel during refining and deoxidation according to claim 2, characterized in that: In S13, the normalization processing formula is as follows: Where: y min is the normalized minimum value, taking -1, y max is the normalized maximum value, taking 1, X min is the minimum value of X, X max is the maximum value of X, y is the normalized value, and X is the data before normalization.
4. The process for predicting the key oxygen content of molten steel and deoxidation in the refining process according to claim 1, characterized in that: In S3, the transfer function between nodes of the endpoint oxygen content prediction model uses the Sigmoid function, which is as follows: In the formula, x j is the input of the endpoint oxygen content prediction model, w ij represents the weight of the connection between the i-th node and the j-th node, θ i Represents the threshold of the i-th node. j is the output value of the neuron node, and n is the number of input variables of the neural network node.
5. The process for predicting the key oxygen content of molten steel and deoxidation in the refining process according to claim 1, characterized in that: In S4, the error between the predicted value and the actual value of the endpoint oxygen content prediction model includes the root mean square error and the mean absolute error; the root mean square error and the mean absolute error are calculated as follows: Where: X Cal. is the endpoint oxygen content predicted by the endpoint oxygen content prediction model, X Exp. is the actual value of the endpoint oxygen content, n is the number of data in the test set, RMSE is the root mean square error, and MAE is the mean absolute error.