A deep neural network-based copper converter blowing endpoint prediction system and method

Abstract

This invention proposes a copper converter blowing endpoint prediction system and method based on deep neural networks, belonging to the field of metallurgical technology. The method includes: dividing the entire copper converter blowing process into five sub-stages; extracting key features for each stage and building a dedicated deep neural network model; integrating these models after training and optimization; identifying stages in real time and calling the corresponding models; and completing model switching and accurate prediction of the blowing endpoint. This invention, by finely dividing the copper converter blowing process into five sub-stages and building and optimizing dedicated prediction models for each sub-stage, abandons the single full-process model, making the models adaptable to the reaction characteristics of each stage, significantly improving the accuracy of reaction progress and blowing endpoint prediction, and achieving refined and highly accurate prediction of the blowing process.

Description

Technical Field

[0001] This invention relates to the field of metallurgical technology, and more specifically, to a copper converter blowing endpoint prediction system and method based on deep neural networks. Background Technology

[0002] As the copper smelting industry transforms and upgrades towards intelligence and precision, copper converter blowing, as the core link in crude copper preparation, directly affects the quality of crude copper, energy consumption, and production safety through accurate endpoint determination. This process is characterized by complex operating conditions, fast reaction rates, and strong coupling of process parameters, leading to increasingly higher requirements for the real-time performance and accuracy of endpoint prediction. Although AI technologies such as deep neural networks are gradually being applied to metallurgical process prediction, existing technologies do not align with the staged process characteristics of the entire blowing process and are difficult to adapt to the actual needs of high-stability and high-precision endpoint prediction in industrial settings. The industry urgently needs targeted intelligent prediction solutions.

[0003] Traditional copper converter blowing endpoint prediction generally adopts a single model covering the entire blowing process, without finely breaking down the blowing process into stages. This makes it impossible to match the specific reaction patterns and process requirements of each stage, including slag formation stage I, slag formation stage II, copper formation stage I, copper formation stage II, and copper formation stage III. The model input does not screen stage characteristic parameters, making it susceptible to interference from redundant data. Furthermore, there is no mechanism for independent training, tuning, and optimization for each stage, resulting in large prediction biases, poor stability, and frequent endpoint misjudgments, which seriously affects the quality of crude copper production and operational efficiency.

[0004] Therefore, it is necessary to design a copper converter blowing endpoint prediction system and method based on deep neural networks to solve the problem that the existing technology uses a single model to adapt to the entire blowing process, which cannot match the process characteristics of each sub-stage, resulting in insufficient accuracy in blowing endpoint prediction. Summary of the Invention

[0005] In view of this, the present invention proposes a copper converter blowing endpoint prediction system and method based on deep neural networks, aiming to solve the problem that the existing technology uses a single model to adapt to the entire blowing process, which cannot match the process characteristics of each sub-stage, resulting in insufficient accuracy in blowing endpoint prediction.

[0006] In one aspect, the present invention proposes a copper converter blowing endpoint prediction system based on deep neural networks, comprising: The stage division module is used to acquire historical multi-source process data and divide the entire copper converter blowing process into five sub-stages based on the historical multi-source process data. The model building module is communicatively connected to the stage division module. The model building module is used to filter and extract key feature parameters for each sub-stage, and build a prediction model for each sub-stage based on the key feature parameters and a deep neural network. The model building module is also used to perform hyperparameter tuning on the prediction models for each sub-stage based on the verification accuracy during the verification phase. Furthermore, the model building module is used to optimize the prediction models for each sub-stage based on the prediction accuracy during the testing phase. Finally, the model building module is used to integrate the optimized prediction models into a unified prediction system framework and configure a unique stage invocation identifier for each prediction model. The model invocation module is communicatively connected to the model building module and the data sensors mounted on the copper converter. The model invocation module is used to collect real-time multi-source process data and, based on this data, identify the current sub-stage of the copper converter's blowing process. The module is also used to match the corresponding stage invocation identifier and invoke the prediction model for the current sub-stage. Furthermore, the module is used to select the key feature parameters of the current sub-stage from the real-time multi-source process data as model input. The endpoint prediction module is communicatively connected to the model invocation module. The model invocation module is used to output the core reaction progress index of the current sub-stage based on the invoked prediction model, and to complete the model switching based on the core reaction progress index. The endpoint prediction module is also used to obtain and output the prediction result of the copper converter blowing endpoint based on the core reaction progress index corresponding to the copper production phase III when switching to copper production phase III.

[0007] Furthermore, when the stage division module divides the entire copper converter blowing process into five sub-stages based on the historical multi-source process data, it includes: When the melt temperature is greater than or equal to the first melt temperature threshold, the SO2 concentration is greater than or equal to the first SO2 concentration threshold, and the time for maintaining the current state is greater than or equal to the first preset duration, the stage division module determines that the current sub-stage is the first stage of slag making. When the SO2 concentration is less than the first SO2 concentration threshold, and the SO2 concentration is greater than the second SO2 concentration threshold, and the melt temperature is greater than or equal to the second melt temperature threshold, and the iron-silicon ratio in the slag is greater than or equal to the first iron-silicon ratio threshold, the stage division module determines that the current sub-stage is the second stage of slag making. When the SO2 concentration is less than or equal to the second SO2 concentration threshold, the melt temperature is greater than or equal to the third melt temperature threshold, and the iron-silicon ratio in the slag is greater than or equal to the second iron-silicon ratio threshold, the stage division module determines that the current sub-stage is copper production phase one. When the sulfur content of the melt is less than or equal to the first sulfur content threshold of the melt, and the CO concentration is in the first CO concentration range, and the cumulative air supply is greater than or equal to the first cumulative air supply threshold, the stage division module determines that the current sub-stage is the second phase of copper production. When the sulfur content of the melt is less than or equal to the second melt sulfur content threshold, the melt temperature is greater than or equal to the fourth melt temperature threshold, and the cumulative air supply is greater than or equal to the second cumulative air supply threshold, the stage division module determines that the current sub-stage is the third phase of copper production.

[0008] Furthermore, when the model building module filters and extracts key feature parameters for each sub-stage, it includes: If the current sub-stage is in the first stage of slag formation, then the key characteristic parameters are SO2 concentration, SO2 concentration change rate, furnace melt temperature, air supply volume and quartz stone addition rate. If the current sub-stage is in the second stage of slag formation, then the key characteristic parameters are SO2 concentration, furnace melt temperature fluctuation value, slag FeO content, slag SiO2 content, and oxygen supply concentration. If the current sub-stage is the first stage of copper production, then the key characteristic parameters are CO concentration, CO2 to CO ratio, furnace melt temperature, oxygen supply concentration, and copper matte conversion rate. If the current sub-stage is in the second phase of copper production, then the key characteristic parameters are: stable CO concentration, real-time crude copper grade, quantitative characteristics of melt color, melt temperature in the furnace, and oxygen supply concentration. If the current sub-stage is in the third phase of copper production, the key characteristic parameters are the real-time value of crude copper grade, CO concentration, CO2 to CO ratio, furnace melt temperature fluctuation value, and O2 concentration stability value.

[0009] Furthermore, when the model building module builds a prediction model for each sub-stage based on the key feature parameters and the deep neural network, it includes: The model building module constructs a corresponding key feature parameter dataset for each sub-stage based on the key feature parameters extracted from each sub-stage. The model building module preprocesses the key feature parameter datasets of each sub-stage and divides the preprocessed key feature parameter datasets into training set, validation set and test set. The model building module designs prediction models for each sub-stage based on the feature dimensions and prediction targets of each sub-stage. The model building module trains the prediction model corresponding to each sub-stage based on the training set of each sub-stage. The model building module validates the prediction model corresponding to each sub-stage based on the validation set of each sub-stage, obtains the validation accuracy, and performs hyperparameter tuning on the prediction model of each sub-stage based on the validation accuracy. The model building module tests the prediction model corresponding to each sub-stage based on the test set of each sub-stage, obtains the prediction accuracy, and optimizes the prediction model of each sub-stage based on the prediction accuracy.

[0010] Furthermore, when the model building module performs hyperparameter tuning on the prediction model for each sub-stage based on the verification accuracy, it includes: When the verification accuracy of a certain sub-stage is less than the corresponding verification accuracy threshold within a consecutive preset number of rounds, the model building module performs hyperparameter tuning on the prediction model corresponding to the current sub-stage, and retrains and verifies the prediction model based on the training set and the validation set until the verification accuracy is greater than or equal to the corresponding verification accuracy threshold within a consecutive preset number of rounds. If the verification accuracy of a certain sub-stage is greater than or equal to the corresponding verification accuracy threshold within a consecutive preset number of rounds, the model building module keeps the hyperparameters of the current prediction model unchanged.

[0011] Furthermore, when the model building module optimizes the prediction model for each sub-stage based on the prediction accuracy, it includes: When the prediction accuracy of a certain sub-stage is greater than or equal to the prediction accuracy threshold corresponding to the current sub-stage, the model building module keeps the network architecture and hyperparameters of the current prediction model unchanged. When the prediction accuracy of a certain sub-stage is less than the prediction accuracy threshold corresponding to the current sub-stage, the model building module readjusts the network architecture and hyperparameters of the prediction model, and retrains, validates and tests the prediction model until the prediction accuracy of the current sub-stage is greater than or equal to the prediction accuracy threshold corresponding to the current sub-stage.

[0012] Furthermore, when the endpoint prediction module outputs the core response progress indicators for the current sub-stage through the invoked prediction model, it includes: The core reaction progress indicator corresponding to the first phase of slag formation is the FeO removal rate; The core reaction progress indicator corresponding to the second phase of slag making is the slag phase qualification rate; The core reaction progress indicator corresponding to the first phase of copper production is the copper matte conversion rate; The core reaction progress indicator corresponding to the second phase of copper production is the real-time value of crude copper grade. The core reaction progress indicator corresponding to the third phase of copper production is the probability of achieving the endpoint target.

[0013] Furthermore, when the endpoint prediction module switches models based on the core response progress indicator, it includes: When the FeO removal rate of the first stage of slag making is greater than or equal to the preset rate threshold, the endpoint prediction module switches to call the prediction model corresponding to the second stage of slag making. When the slag phase qualification rate of the second stage of slag making is greater than or equal to the qualification rate threshold, the endpoint prediction module switches to call the prediction model corresponding to the first stage of copper making. When the copper matte conversion rate of the first phase of copper production is greater than or equal to the conversion threshold, the endpoint prediction module switches to call the prediction model corresponding to the second phase of copper production. When the real-time value of the crude copper grade in the second phase of copper production is greater than or equal to the real-time grade threshold, the endpoint prediction module switches to call the prediction model corresponding to the third phase of copper production.

[0014] Furthermore, when the endpoint prediction module switches to the third phase of copper production, it obtains and outputs the prediction result of the copper converter blowing endpoint based on the core reaction progress index corresponding to the third phase of copper production, including: When the probability of reaching the endpoint is greater than or equal to the preset endpoint probability threshold, the real-time value of the crude copper grade is greater than or equal to the target grade threshold, and the time for maintaining the current state is greater than the second preset duration, the endpoint prediction module determines that the copper converter blowing endpoint has been reached and outputs the prediction result of the blowing endpoint being reached.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: By meticulously dividing the entire copper converter smelting process into five sub-stages, and building dedicated deep neural network prediction models for each sub-stage based on its unique process characteristics, reaction patterns, and prediction requirements, the model training, validation, hyperparameter tuning, and optimization are completed by combining key feature parameters of each sub-stage. This completely abandons the traditional approach of adapting a single model to the entire smelting process, ensuring that the model is highly adapted to the reaction characteristics of each sub-stage. This effectively avoids the bias caused by uniform prediction of the entire process, fundamentally improving the prediction accuracy of reaction progress and smelting endpoint of each sub-stage, and achieving refined and highly accurate prediction of the copper converter smelting process.

[0016] On the other hand, this invention proposes a method for predicting the endpoint of copper converter blowing based on deep neural networks, including: Historical multi-source process data is acquired, and the entire copper converter blowing process is divided into five sub-stages based on the historical multi-source process data; Key feature parameters are selected and extracted for each sub-stage, and a prediction model is built for each sub-stage based on the key feature parameters; in the verification stage, the hyperparameters of the prediction models for each sub-stage are tuned based on the verification accuracy; in the testing stage, the prediction models for each sub-stage are optimized based on the prediction accuracy; the optimized prediction models are integrated into a unified prediction system framework, and a unique stage call identifier is configured for each prediction model. Collect real-time multi-source process data, and based on the real-time multi-source process data, identify the current sub-stage of copper converter blowing; match the corresponding stage call identifier, and call the prediction model of the current sub-stage in a targeted manner; select the key feature parameters of the current sub-stage from the real-time multi-source process data as model input; The prediction model is invoked to output the core reaction progress indicators of the current sub-stage, and the model is switched based on the core reaction progress indicators. When switching to the third stage of copper production, the prediction results of the copper converter blowing endpoint are obtained and output based on the core reaction progress indicators corresponding to the third stage of copper production.

[0017] It is understandable that the aforementioned method and system for predicting the end point of copper converter blowing based on deep neural networks have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 Functional block diagram of a copper converter blowing endpoint prediction system based on deep neural networks provided in an embodiment of the present invention; Figure 2 A flowchart of a copper converter blowing endpoint prediction method based on deep neural networks provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art. It should be noted that, without conflict, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It is important to note that: The five sub-stages are: slag production stage I, slag production stage II, copper production stage I, copper production stage II, and copper production stage III.

[0021] The historical multi-source process data includes: SO2 concentration, CO concentration, melt temperature, cumulative air supply, melt sulfur content, and iron-silicon ratio in slag; the real-time multi-source process data also includes: SO2 concentration, CO concentration, melt temperature, cumulative air supply, melt sulfur content, and iron-silicon ratio in slag.

[0022] First melt temperature threshold < Second melt temperature threshold < Third melt temperature threshold < Fourth melt temperature threshold, First SO2 concentration threshold > Second SO2 concentration threshold, First iron-silicon ratio threshold < Second iron-silicon ratio threshold, First melt sulfur content threshold > Second melt sulfur content threshold, First cumulative air supply threshold < Second cumulative air supply threshold.

[0023] SO2 concentration refers to the SO2 concentration in the flue gas during the copper converter smelting process. It is used to divide the slag formation stage into two sub-stages and serves as a key characteristic parameter for both stages. The SO2 concentration is collected in real time by installing an SO2 concentration sensor in the copper converter flue gas pipeline. The collected real-time SO2 concentration value is the required SO2 concentration.

[0024] CO concentration refers to the carbon monoxide content in the flue gas inside the copper converter during copper smelting. It is used to divide the copper production process into two sub-stages and serves as a key characteristic parameter for each stage. The carbon monoxide content is collected by a copper converter flue gas monitoring sensor, and abnormal fluctuations are filtered out. The stable value is the real-time CO concentration.

[0025] Melt temperature refers to the real-time temperature value of the molten material inside the copper converter during the blowing process. It is used to divide all five blowing sub-stages and is the basic parameter for determining and predicting each stage. The melt temperature is collected in real time by a high-temperature resistant temperature sensor inside the furnace, and outliers are removed. The processed value is the real-time melt temperature.

[0026] Cumulative air supply volume refers to the cumulative air supply volume from the start of copper converter smelting to the current moment. It is used to divide the copper production phase II and copper production phase III into sub-stages and to determine the progress of the smelting stage. The instantaneous air supply volume is collected by the air supply flow sensor and calculated by accumulating the data. The total accumulated air volume value is the real-time cumulative air supply volume.

[0027] The sulfur content in melt refers to the sulfur content in the molten material during copper converter blowing. It is used to divide the copper production process into two sub-stages: the second stage and the third stage. It is a key indicator for determining the copper production stage. The sulfur content is collected and calibrated by a melt composition detection sensor in the furnace. The calibrated value is the real-time sulfur content in melt.

[0028] The iron-silicon ratio in slag refers to the ratio of iron to silicon content in the slag during copper converter blowing. It is used to divide the second stage of slag making and the first stage of copper making into sub-stages and to determine the stage transition node. The iron and silicon contents are collected separately by slag composition detection equipment and the ratio is calculated. The result of dividing the iron content by the silicon content is the iron-silicon ratio in slag.

[0029] The first melt temperature threshold refers to the critical melt temperature value used to divide the first stage of slag formation in copper converter blowing, and is used to determine whether blowing has entered the first stage of slag formation. The first melt temperature threshold is determined by statistically analyzing the critical melt temperature data of the initial stage of the first stage of slag formation in history and completing the calibration.

[0030] The second melt temperature threshold is the critical melt temperature value used to divide the second stage of slag formation in copper converter blowing, and is used to determine whether blowing has entered the second stage of slag formation. The second melt temperature threshold is determined by statistically analyzing the critical melt temperature data of the initial stage of the second stage of slag formation in history and completing the calibration.

[0031] The third melt temperature threshold refers to the critical melt temperature value used to divide the copper converter blowing process into the first stage of copper production. It is used to determine whether the blowing process has entered the first stage of copper production. The third melt temperature threshold is determined by statistically analyzing the critical melt temperature data of the historical first stage of copper production and completing the calibration.

[0032] The fourth melt temperature threshold is the critical melt temperature value used to divide the three sub-stages of copper converter blowing for copper production. It is used to determine whether the blowing process has entered the third stage of copper production. The fourth melt temperature threshold is determined by statistically analyzing the critical melt temperature data of the initial stage of the three stages of copper production in history and completing the calibration.

[0033] The first SO2 concentration threshold refers to the critical value of SO2 concentration that distinguishes between the first and second stages of slag formation in copper converter smelting. It is used to determine the transition from the first to the second stage of slag formation. The first SO2 concentration threshold is determined by statistically analyzing historical critical data of SO2 concentration at the end of the first stage and the beginning of the second stage of slag formation.

[0034] The second SO2 concentration threshold refers to the critical value of SO2 concentration that distinguishes the second stage of slag formation and the first stage of copper production in copper converter smelting. It is used to determine the switch from the second stage of slag formation to the first stage of copper production. The second SO2 concentration threshold is determined by statistically analyzing the critical data of SO2 concentration at the end of the second stage of slag formation and the beginning of the first stage of copper production.

[0035] The first iron-silicon ratio threshold refers to the critical value of the iron-silicon ratio in the slag that divides the second stage of slag formation in copper converter smelting. It is used to determine whether smelting has entered the second stage of slag formation. The first iron-silicon ratio threshold is obtained by statistically analyzing the critical data of the iron-silicon ratio in the slag at the beginning of the second stage of slag formation in history and completing the calibration.

[0036] The second iron-silicon ratio threshold refers to the critical value of the iron-silicon ratio in the slag for dividing the first stage of copper production in copper converter blowing, and is used to determine whether blowing has entered the first stage of copper production. By statistically analyzing the critical data of the iron-silicon ratio in the slag at the beginning of the first stage of copper production in history and completing the calibration, the calibrated ratio value is the second iron-silicon ratio threshold.

[0037] The first melt sulfur content threshold refers to the critical value of melt sulfur content for dividing the second stage of copper converter blowing into copper production, and is used to determine whether the blowing has entered the second stage of copper production. The first melt sulfur content threshold is obtained by statistically analyzing the critical data of melt sulfur content at the beginning of the second stage of copper production in history and completing the calibration.

[0038] The second melt sulfur content threshold refers to the critical value of melt sulfur content for dividing the three sub-stages of copper converter blowing copper production, and is used to determine whether blowing has entered the third stage of copper production. By statistically analyzing the critical data of melt sulfur content at the beginning of the third stage of copper production in history and completing the calibration, the calibrated content value is the second melt sulfur content threshold.

[0039] The first cumulative air supply threshold refers to the critical value of cumulative air supply for dividing the second phase of copper converter smelting into sub-stages of copper production. It is used to determine whether smelting has entered the second phase of copper production. By statistically analyzing the critical data of cumulative air supply at the beginning of the second phase of copper production in history and completing the calibration, the calibrated cumulative value is the first cumulative air supply threshold.

[0040] The second cumulative air supply threshold refers to the critical value of cumulative air supply for dividing the three sub-stages of copper converter smelting and copper production. It is used to determine whether smelting has entered the third stage of copper production. The cumulative air supply threshold is determined by statistically analyzing the critical data of cumulative air supply at the beginning of the three stages of copper production in history and completing the calibration.

[0041] The first preset duration refers to the critical value for determining the state maintenance time of the first stage of slag formation in copper converter blowing, which is used to confirm that the state of the first stage of slag formation is stable and meets the standards. The first preset duration is determined by statistically analyzing the shortest time data of the first stage of slag formation in history and completing the calibration.

[0042] The first CO concentration range refers to the range of CO concentration values ​​used to define the second phase of copper production in copper converter blowing, and is used to determine whether blowing has entered the second phase of copper production. It is defined as the first CO concentration range by statistically analyzing the upper and lower limits of CO concentration during the stable phase of the second phase of copper production in history and completing the calibration. Validation accuracy refers to the accuracy of a sub-stage prediction model on a validation set. It is used to determine whether the model needs hyperparameter tuning. The accuracy percentage is calculated by comparing the model output results on the validation set with the actual process results. The calculated accuracy percentage is the validation accuracy.

[0043] The preset number of rounds refers to the number of times the verification accuracy is continuously checked to determine whether the model verification effect continues to fail to meet the standard. The number of continuous checks is set according to the model training and debugging requirements, and the set number of rounds is the preset number of rounds.

[0044] The verification accuracy threshold is the critical value used to determine whether the verification effect of the sub-stage prediction model is qualified. It is used to decide whether to start the hyperparameter tuning process. The lowest verification accuracy of the historical qualified sub-stage models is statistically analyzed and calibrated. The calibrated accuracy value is the verification accuracy threshold.

[0045] Prediction accuracy refers to the final prediction accuracy of the sub-stage prediction model on the test set, and is used to determine whether the model needs to be optimized and adjusted. The accuracy percentage is calculated by comparing the model output results on the test set with the actual process results. The calculated accuracy percentage is the prediction accuracy.

[0046] The prediction accuracy threshold is the critical value used to determine whether the test results of the prediction model in a sub-stage are acceptable. It is used to decide whether to retain the current model. The lowest prediction accuracy of historical acceptable sub-stage models is statistically analyzed and calibrated. The calibrated accuracy value is the prediction accuracy threshold.

[0047] FeO removal rate refers to the rate at which FeO is removed from the slag per unit time during the first stage of slag making. It serves as a core reaction progress indicator for the first stage of slag making and is used to determine whether to switch to the second stage of slag making model. The FeO removal rate is calculated by collecting the change in FeO content per unit time and calculating the amount removed per unit time.

[0048] The preset rate threshold is the critical rate value used to determine whether the FeO removal in the first stage of slag making meets the standard and can proceed to the next stage. It is used to trigger the model switch from the first stage of slag making to the second stage of slag making. The preset rate threshold is determined by statistically analyzing the critical FeO removal rate data at the end of the first stage of slag making in history and calibrating it.

[0049] Slag phase qualification refers to the degree to which the slag composition, iron-silicon ratio and other indicators meet the process requirements in the second stage of slag making. As a core reaction progress indicator in the second stage of slag making, it is used to determine whether to switch to the first stage copper making model. The slag phase qualification is obtained by comprehensively scoring and normalizing the key slag indicators according to the process standards.

[0050] The qualification threshold is the critical qualification value used to determine whether the slag phase of the second stage of slag making meets the standards and can enter the next stage. It is used to trigger the model switch from the second stage of slag making to the first stage of copper making. The qualification threshold is determined by statistically analyzing the critical qualification data of the slag phase at the end of the qualified stage of the second stage of slag making in history and calibrating it.

[0051] Copper matte conversion rate refers to the proportion of copper matte converted into the target product in the first stage of copper production. As a core reaction progress indicator in the first stage of copper production, it is used to determine whether to switch to the second stage of copper production model. The ratio is calculated by dividing the difference in copper matte content before and after conversion by the initial copper matte content. The calculated ratio is the copper matte conversion rate.

[0052] The conversion threshold refers to the critical conversion ratio at which the copper matte conversion in the first phase of copper production meets the standard and can proceed to the next stage. It is used to trigger the model switch from the first phase of copper production to the second phase. The conversion threshold is determined by statistically analyzing and calibrating the critical conversion rate data of copper matte at the end of the first phase of copper production in history.

[0053] The real-time value of crude copper grade refers to the quality and purity value of crude copper generated in the furnace in real time during the second and third stages of copper production. As a core indicator of the copper production stage, it is used for stage switching and endpoint determination. The crude copper quality data is collected in real time by the furnace composition detection device and calibrated. The calibrated value is the real-time value of crude copper grade.

[0054] The real-time grade threshold is the critical quality value used to determine whether the grade of crude copper in the second phase of copper production meets the standard and can enter the third phase of copper production. It is used to trigger the model switch from the second phase of copper production to the third phase of copper production. The critical grade data of crude copper at the end of the qualified second phase of copper production in history is statistically analyzed and calibrated. The calibrated grade value is the real-time grade threshold.

[0055] The probability of reaching the target endpoint refers to the probability value that the copper refining process will reach the target endpoint in the three-stage copper refining process. It is used to determine whether the copper converter refining process has reached the target endpoint. The probability of reaching the target endpoint is calculated by combining the model with parameters such as real-time crude copper grade and melt temperature.

[0056] The preset endpoint probability threshold is the critical probability value used to determine whether the probability of meeting the three-stage copper production endpoint is qualified and to confirm the blowing endpoint. It is used to trigger the output of the blowing endpoint determination. The lowest probability data of qualified blowing endpoints in history are statistically analyzed and calibrated. The calibrated probability value is the preset endpoint probability threshold.

[0057] The target grade threshold is the critical quality value used to determine whether the grade of crude copper at the end of the blowing process meets the production requirements. It is used in conjunction with the probability of reaching the end point to confirm the end of the blowing process. The lowest grade data of crude copper at the historical qualified end point is statistically analyzed and calibrated. The calibrated grade value is the target grade threshold.

[0058] The second preset duration refers to the critical time value at which various compliance indicators need to be stably maintained when the end point of the third phase of copper production is determined. It is used to confirm that the end point status is stable and does not rebound. The shortest time data of the historical qualified end point indicators are statistically analyzed and calibrated. The calibrated duration value is the second preset duration.

[0059] Reference Figure 1 As shown in some embodiments of this application, this embodiment provides a copper converter blowing endpoint prediction system based on deep neural networks, including: The stage division module is used to acquire historical multi-source process data and divide the entire copper converter blowing process into five sub-stages based on the historical multi-source process data.

[0060] Specifically, when the stage division module divides the entire copper converter blowing process into five sub-stages based on the historical multi-source process data, it includes: When the melt temperature is greater than or equal to the first melt temperature threshold, the SO2 concentration is greater than or equal to the first SO2 concentration threshold, and the time for maintaining the current state is greater than or equal to the first preset duration, the stage division module determines that the current sub-stage is the first stage of slag making. When the SO2 concentration is less than the first SO2 concentration threshold, and the SO2 concentration is greater than the second SO2 concentration threshold, and the melt temperature is greater than or equal to the second melt temperature threshold, and the iron-silicon ratio in the slag is greater than or equal to the first iron-silicon ratio threshold, the stage division module determines that the current sub-stage is the second stage of slag making. When the SO2 concentration is less than or equal to the second SO2 concentration threshold, the melt temperature is greater than or equal to the third melt temperature threshold, and the iron-silicon ratio in the slag is greater than or equal to the second iron-silicon ratio threshold, the stage division module determines that the current sub-stage is copper production phase one. When the sulfur content of the melt is less than or equal to the first sulfur content threshold of the melt, and the CO concentration is in the first CO concentration range, and the cumulative air supply is greater than or equal to the first cumulative air supply threshold, the stage division module determines that the current sub-stage is the second phase of copper production. When the sulfur content of the melt is less than or equal to the second melt sulfur content threshold, the melt temperature is greater than or equal to the fourth melt temperature threshold, and the cumulative air supply is greater than or equal to the second cumulative air supply threshold, the stage division module determines that the current sub-stage is the third phase of copper production.

[0061] Specifically, other conditions can also be used to distinguish sub-stages, as follows: When the melt temperature is less than the first melt temperature threshold, or the SO2 concentration is less than the first SO2 concentration threshold, or the time to maintain the current state is less than the first preset duration, the current sub-stage is determined to be the initial stage of blowing. When the SO2 concentration is greater than or equal to the first SO2 concentration threshold, or the SO2 concentration is less than or equal to the second SO2 concentration threshold, or the melt temperature is less than the second melt temperature threshold, or the iron-silicon ratio in the slag is less than the first iron-silicon ratio threshold, the current sub-stage is determined to be the first stage of slag making. When the SO2 concentration is greater than the second SO2 concentration threshold, or the melt temperature is less than the third melt temperature threshold, or the iron-silicon ratio in the slag is less than the second iron-silicon ratio threshold, the current sub-stage is determined to be the second stage of slag making. When the sulfur content of the melt is greater than the first melt sulfur content threshold, or the CO concentration is not in the first CO concentration range, or the cumulative air supply is less than the first cumulative air supply threshold, the current sub-stage is determined to be the first stage of copper production. When the sulfur content of the melt is greater than the second melt sulfur content threshold, or the melt temperature is less than the fourth melt temperature threshold, or the cumulative air supply is less than the second cumulative air supply threshold, the current sub-stage is determined to be the second phase of copper production. It can be understood that the slag-forming phase I, slag-forming phase II, copper-forming phase I, copper-forming phase II, and copper-forming phase III of copper converter smelting each have unique and stable physicochemical reaction characteristics. Slag-forming phase I is characterized by achieving the target melt temperature and a significant and stable increase in SO2 concentration. Slag-forming phase II is characterized by a decrease in SO2 concentration and simultaneous achievement of the target melt temperature and iron-silicon ratio in the slag. Copper-forming phase I is characterized by a decrease in SO2 concentration to a low level and the melt temperature and iron-silicon ratio in the slag meeting the initial requirements for copper production. Copper-forming phase II is characterized by a decrease in melt sulfur content, a stable CO concentration, and achievement of the target cumulative blast volume. The core characteristics of the three stages of copper production are extremely low sulfur content in the melt, and the melt temperature and cumulative air volume reaching the final standard. This invention targets the unique core characteristics of each stage and selects a multi-parameter "AND" joint judgment condition that can be satisfied simultaneously only in the corresponding stage. This can accurately match the specific reaction state of each stage. Other judgment conditions can only reflect some non-core characteristics and are easily affected by parameter fluctuations and overlapping stage transitions, resulting in judgment errors. Therefore, by adopting this multi-parameter synchronous standard-reaching judgment method, the five sub-stages can be accurately and unambiguously divided, which is completely in line with the actual reaction process law of copper converter blowing.

[0062] For example: the first melt temperature threshold is set to 1200℃, the second melt temperature threshold is set to 1230℃, the third melt temperature threshold is set to 1250℃, the fourth melt temperature threshold is set to 1280℃, the first SO2 concentration threshold is set to 15%, the second SO2 concentration threshold is set to 5%, the first iron-silicon ratio threshold is set to 1.2, the second iron-silicon ratio threshold is set to 0.8, the first melt sulfur content threshold is set to 5%, the second melt sulfur content threshold is set to 1%, and the first cumulative air supply threshold is set to 800 Nm. 3 The second cumulative air supply threshold is set to 1500 Nm. 3 The first preset duration was set to 10 seconds, and the first CO concentration range was set to 0.2%–0.5%. Specific data was selected from a set of historical data from the past 300 heats of copper converter blowing, including historical melt temperatures of 1200℃, 1230℃, 1250℃, and 1280℃; historical SO2 concentrations of 15%, 10%, and 5%; historical iron-silicon ratios in slag of 1.2 and 0.8; historical melt sulfur content of 5% and 1%; historical CO concentration of 0.3%; and historical cumulative air supply of 800 Nm³. 3 1500Nm 3The current state is maintained for 10 seconds. When the historical melt temperature is 1200℃, the historical SO2 concentration is 15%, and the current state is maintained for 10 seconds, the current blowing process is determined to be in the first stage of slag formation. When the historical SO2 concentration is 10%, the historical melt temperature is 1230℃, and the historical iron-silicon ratio in the slag is 1.2, the current blowing process is determined to be in the second stage of slag formation. When the historical SO2 concentration is 5%, the historical melt temperature is 1250℃, and the historical iron-silicon ratio in the slag is 0.8, the current blowing process is determined to be in the first stage of copper production. When the historical melt sulfur content is 5%, the historical CO concentration is 0.3%, and the historical cumulative air supply is 800 Nm³, the current blowing process is determined to be in the first stage of copper production. 3 At that time, it was determined that the current smelting was in the second stage of copper production; when the historical melt sulfur content was 1%, the historical melt temperature was 1280℃, and the historical cumulative air supply was 1500 Nm. 3 At that time, it was determined that the current smelting process was in the third sub-stage of copper production.

[0063] The model building module is communicatively connected to the stage division module. The model building module is used to filter and extract key feature parameters for each sub-stage, and build a prediction model for each sub-stage based on the key feature parameters and a deep neural network. The model building module is also used to perform hyperparameter tuning of the prediction models for each sub-stage based on the verification accuracy during the verification phase. Furthermore, the model building module is used to optimize the prediction models for each sub-stage based on the prediction accuracy during the testing phase. Finally, the model building module is used to integrate the optimized prediction models into a unified prediction system framework and configure a unique stage invocation identifier for each prediction model.

[0064] Specifically, when the model building module filters and extracts key feature parameters for each sub-stage, it includes: If the current sub-stage is in the first stage of slag formation, then the key characteristic parameters are SO2 concentration, SO2 concentration change rate, furnace melt temperature, air supply volume and quartz stone addition rate. If the current sub-stage is in the second stage of slag formation, then the key characteristic parameters are SO2 concentration, furnace melt temperature fluctuation value, slag FeO content, slag SiO2 content, and oxygen supply concentration. If the current sub-stage is the first stage of copper production, then the key characteristic parameters are CO concentration, CO2 to CO ratio, furnace melt temperature, oxygen supply concentration, and copper matte conversion rate. If the current sub-stage is in the second phase of copper production, then the key characteristic parameters are: stable CO concentration, real-time crude copper grade, quantitative characteristics of melt color, melt temperature in the furnace, and oxygen supply concentration. If the current sub-stage is in the third phase of copper production, the key characteristic parameters are the real-time value of crude copper grade, CO concentration, CO2 to CO ratio, furnace melt temperature fluctuation value, and O2 concentration stability value.

[0065] It can be understood that the slag-forming stages I, II, I, II, and III of copper converter smelting each have distinct core physicochemical reaction characteristics. Slag-forming stage I focuses on oxidative de-ironization and slag formation; therefore, selecting SO2 concentration, SO2 concentration change rate, furnace melt temperature, air supply volume, and quartz addition rate can accurately characterize the core states of the de-ironization oxidation reaction intensity, reaction temperature baseline, oxidation rate control, and slag composition adjustment. Slag-forming stage II focuses on deep de-ironization and slag formation; therefore, selecting SO2 concentration, furnace melt temperature fluctuation value, slag FeO content, slag content, and oxygen supply concentration can directly reflect the deep slag formation effect, reaction stability, slag product indicators, and oxygen supply control requirements. Copper formation stage I focuses on the initial conversion of copper matte into copper; therefore, selecting CO concentration, SO2 to CO ratio, furnace melt temperature, oxygen supply concentration, and copper matte conversion rate can accurately reflect the initial reaction atmosphere, conversion process, and basic reaction conditions. Copper formation stage II focuses on crude... The core objective is to achieve continuous and stable copper production. Therefore, stable CO concentration, real-time blister copper grade, quantitative characteristics of melt color, furnace melt temperature, and oxygen concentration are selected to comprehensively characterize the quality of blister copper production, the stability of the reaction atmosphere, the visual state of the melt, and the conditions for ensuring the reaction. The third stage of copper production focuses on the approaching end of the blowing process and the near completion of the reaction. Therefore, real-time blister copper grade, CO concentration, SO2 to CO ratio, furnace melt temperature fluctuation, and stable SO2 concentration are selected to accurately reflect the final product compliance, the final reaction state, process stability, and the degree of reaction completion. This invention is based on the unique core reaction characteristics of each sub-stage. It selectively selects parameters that can only characterize the reaction process of the corresponding stage and have no redundant interference as key feature parameters, and eliminates parameters that are irrelevant to the stage reaction. This ensures that the feature inputs are completely consistent with the reaction rules of each stage, providing accurate and suitable feature support for the construction of the stage prediction model, and fundamentally improving the accuracy of stage prediction and endpoint determination.

[0066] Specifically, when the model building module builds prediction models for each sub-stage based on the key feature parameters and deep neural networks, it includes: The model building module constructs a corresponding key feature parameter dataset for each sub-stage based on the key feature parameters extracted from each sub-stage. The model building module preprocesses the key feature parameter datasets of each sub-stage and divides the preprocessed key feature parameter datasets into training set, validation set and test set. The model building module designs prediction models for each sub-stage based on the feature dimensions and prediction targets of each sub-stage. The model building module trains the prediction model corresponding to each sub-stage based on the training set of each sub-stage. The model building module validates the prediction model corresponding to each sub-stage based on the validation set of each sub-stage, obtains the validation accuracy, and performs hyperparameter tuning on the prediction model of each sub-stage based on the validation accuracy. The model building module tests the prediction model corresponding to each sub-stage based on the test set of each sub-stage, obtains the prediction accuracy, and optimizes the prediction model of each sub-stage based on the prediction accuracy.

[0067] It is understandable that the five sub-stages of copper converter blowing have fundamentally different characteristics, physicochemical reaction laws, and prediction targets. The slag-forming stage focuses on predicting the iron removal and slag-forming process, the copper-forming stage focuses on predicting copper matte conversion and crude copper generation, and the final stage aims at predicting the achievement of the blowing endpoint. Therefore, it is necessary to design appropriate prediction models based on the unique characteristics and specific prediction targets of each sub-stage to avoid prediction bias caused by the inability of a general model to match the stage characteristics. First, hyperparameter tuning is carried out based on the validation accuracy. During the model training process, the model hyperparameters are corrected in real time through the validation set to prevent overfitting or underfitting and ensure the basic fitting effect and training stability of the model. Then, the final optimization is carried out based on the prediction accuracy. The actual generalization ability and field applicability of the model are tested through an independent test set to further improve the model network architecture and parameters. The two optimizations are layered and progressive, each with its own function. The modeling is adapted to the differences of each stage, and the two optimizations ensure the accuracy of the model at each stage. In this way, the prediction model of each sub-stage can accurately match the reaction characteristics and prediction requirements of the corresponding stage, providing stable and reliable model support for subsequent accurate use and endpoint prediction.

[0068] For example, when preprocessing the key feature parameter datasets for each sub-stage: for the key feature parameter datasets of slag production stage I, slag production stage II, copper production stage I, copper production stage II, and copper production stage III, outliers in each parameter data are removed using the 3σ criterion. For the data after outlier removal, missing values ​​are filled using linear interpolation. Then, the min-max normalization method is used to map each parameter data to the [0,1] interval to eliminate the influence of differences in the magnitude of different parameters. Subsequently, the normalized dataset is subjected to a stationarity test. Data that does not meet the stationarity requirements are subjected to differencing to achieve a stationary state. Finally, the processed key feature parameter datasets for each sub-stage are... Data cleaning removes redundant and duplicate data, and preprocesses the datasets of key feature parameters for each sub-stage. When designing deep neural network prediction models adapted to the feature dimensions and prediction targets for each sub-stage: the key feature parameters for the first stage of slag formation have 5 dimensions (corresponding to SO2 concentration, SO2 concentration change rate, furnace melt temperature, air supply volume, and quartz stone addition rate), and the prediction target is the FeO removal rate. The designed deep neural network prediction model has 5 neurons in the input layer, 2 hidden layers initially (128 neurons per layer), and 1 neuron in the output layer. The ReLU activation function is used, and a linear activation function is used in the output layer to adapt to the rate. Prediction: The key feature parameters for the second phase of slag making are 5-dimensional (corresponding to SO2 concentration, furnace melt temperature fluctuation, slag FeO content, slag SiO2 content, and oxygen supply concentration). The prediction target is slag phase qualification. The designed model consists of an input layer of 5 neurons, an initial 2-layer hidden layer (128 neurons per layer), and an output layer of 1 neuron. The activation function is the ReLU function, and the output layer uses the Sigmoid activation function to predict the qualification percentage. The key feature parameters for the first phase of copper production are 5-dimensional (corresponding to CO concentration, CO2 to CO ratio, furnace melt temperature, oxygen supply concentration, and copper-matte conversion rate). The prediction target is copper-matte conversion rate. The designed model consists of an input layer of 5 neurons, an initial 2-layer hidden layer (128 neurons per layer), and an output layer of 1 neuron. The activation function is the ReLU function, and the output layer uses the Sigmoid activation function to predict the qualification percentage. The model consists of 5 neurons in the input layer, 2 initial hidden layers (128 neurons per layer), and 1 neuron in the output layer. The ReLU activation function is used, and the Sigmoid activation function is used in the output layer to adapt for conversion rate percentage prediction. The key feature parameters of the second phase of copper production are 5-dimensional (corresponding to five parameters: stable CO concentration, real-time value of crude copper grade, quantitative features of melt color, melt temperature in the furnace, and oxygen supply concentration). The prediction target is the real-time value of crude copper grade. The designed model consists of 5 neurons in the input layer, 2 initial hidden layers (128 neurons per layer), and 1 neuron in the output layer. The ReLU activation function is used, and the Sigmoid activation function is used in the output layer to adapt for grade percentage prediction.The key feature parameters for the third phase of copper production are 5-dimensional (corresponding to real-time values ​​of crude copper grade, CO concentration, CO2 to CO ratio, furnace melt temperature fluctuation, and stable O2 concentration). The prediction target is the probability of achieving the final target. The designed model consists of 5 neurons in the input layer, 2 initial hidden layers (128 neurons per layer), and 1 neuron in the output layer. The ReLU activation function is used, and the Sigmoid activation function is used in the output layer to adapt to the probability prediction. Each sub-stage model is adapted to its own feature dimensions and prediction target to ensure the relevance and accuracy of the model's predictions.

[0069] Specifically, when the model building module performs hyperparameter tuning on the prediction model for each sub-stage based on the verification accuracy, it includes: When the verification accuracy of a certain sub-stage is less than the corresponding verification accuracy threshold within a consecutive preset number of rounds, the model building module performs hyperparameter tuning on the prediction model corresponding to the current sub-stage, and retrains and verifies the prediction model based on the training set and the validation set until the verification accuracy is greater than or equal to the corresponding verification accuracy threshold within a consecutive preset number of rounds. If the verification accuracy of a certain sub-stage is greater than or equal to the corresponding verification accuracy threshold within a consecutive preset number of rounds, the model building module keeps the hyperparameters of the current prediction model unchanged.

[0070] It is understandable that during the training process of the prediction model for each sub-stage of copper converter blowing, the validation accuracy of a single round is easily affected by random data fluctuations and local feature perturbations, failing to truly reflect the degree of fit between the model's hyperparameters and the feature dimensions and reaction patterns of the corresponding sub-stage. Therefore, by setting a continuous preset round as a stability judgment period, the validation accuracy of the model on the validation set is continuously monitored. When the validation accuracy of the model for consecutive preset rounds is lower than the validation accuracy threshold of the corresponding sub-stage, it indicates that the current model's hyperparameters cannot adapt to the process data characteristics and reaction patterns of that sub-stage, and the model has overfitting, underfitting, or unstable fitting problems. The hyperparameters must be adjusted and refitted. Training and validation are conducted continuously until the accuracy meets the target after a series of preset rounds, thereby locking in the optimal basic hyperparameters for that sub-stage. When the model's accuracy meets the target after a series of preset rounds, it proves that the current hyperparameters perfectly match the data characteristics and training requirements of the sub-stage, and there is no need to adjust the hyperparameters, avoiding ineffective tuning that would waste computing power and damage model stability. This mechanism accurately identifies hyperparameter fitting problems and corrects them in a targeted manner through stability checks in consecutive rounds, building a stable basic fitting capability for the prediction model in each sub-stage, avoiding misjudgments and ineffective tuning caused by accuracy fluctuations in a single round, and ensuring that each sub-stage model has the optimal hyperparameter configuration that suits its own characteristics.

[0071] For example, the preset number of rounds is set to 3, and the verification accuracy thresholds for each sub-stage are set as follows: 90% for slag-making stage 1, 91% for slag-making stage 2, 92% for copper-making stage 1, 93% for copper-making stage 2, and 94% for copper-making stage 3. In the verification phase of the deep neural network prediction model for each sub-stage, for the slag-making stage 1 model, if the verification accuracy of the model is consistently less than 90% for 3 consecutive rounds, then the hyperparameters of the deep neural network prediction model corresponding to slag-making stage 1 are tuned. Specifically, the initial learning rate is set to 0.01, the hidden layers are set to 2 layers with 128 neurons per layer. If the verification accuracy fails to meet the standard for 3 consecutive rounds, the learning rate is reduced to 0.005, the hidden layers are increased to 3 layers with 256 neurons in the first layer, 128 in the second layer, and 64 in the third layer. The model is then retrained using the training and validation sets from slag-making stage 1. If the verification still fails to meet the standard, the learning rate is further reduced to 0.001, maintaining a 3-layer hidden layer structure. The number of neurons in each layer is adjusted to 256, 192, and 64 respectively, until the verification accuracy of the model is consistently greater than or equal to 90% for 3 consecutive rounds. When the verification accuracy of the first phase of the slag-making model is consistently greater than or equal to 90% for 3 consecutive rounds, the hyperparameters of the first phase of the slag-making model remain unchanged, and it directly enters the testing phase. Similarly, the second phase of the slag-making model, the first phase of the copper-making model, the second phase of the copper-making model, and the third phase of the copper-making model are judged based on the corresponding verification accuracy thresholds of 91%, 92%, 93%, and 94% and the preset number of rounds of 3, respectively. The initial values ​​of the hyperparameters are uniformly set to a learning rate of 0.01 and 2 hidden layers (128 neurons per layer). If the standard is not met, the same adjustment logic is used as described above (the learning rate is gradually reduced to 0.005 and 0.001, the number of hidden layers is increased to 3, and the number of neurons is gradually optimized) to complete the hyperparameter tuning or directly enter the testing phase.

[0072] Specifically, when the model building module optimizes the prediction model for each sub-stage based on the prediction accuracy, it includes: When the prediction accuracy of a certain sub-stage is greater than or equal to the prediction accuracy threshold corresponding to the current sub-stage, the model building module keeps the network architecture and hyperparameters of the current prediction model unchanged. When the prediction accuracy of a certain sub-stage is less than the prediction accuracy threshold corresponding to the current sub-stage, the model building module readjusts the network architecture and hyperparameters of the prediction model, and retrains, validates and tests the prediction model until the prediction accuracy of the current sub-stage is greater than or equal to the prediction accuracy threshold corresponding to the current sub-stage.

[0073] It is understandable that prediction accuracy reflects the final generalization ability of the prediction model for each sub-stage on an independent test set. It directly reflects the model's true effectiveness in adapting to the process characteristics, physicochemical reaction laws, and actual production prediction needs of the corresponding sub-stage. Furthermore, the characteristic dimensions, reaction processes, and prediction targets of the five sub-stages of copper converter blowing are fundamentally different. Therefore, a specific prediction accuracy threshold needs to be set for each sub-stage as a qualified entry standard for the model. When the prediction accuracy of a model for a certain sub-stage reaches or exceeds this specific threshold, it indicates that the model has fully learned the data characteristics of this stage and can stably output accurate prediction results, and can be directly retained as the prediction model for formal application in this stage. When the prediction accuracy is lower than the threshold, it means... The current model's network architecture and hyperparameters are unable to match the feature distribution, reaction logic, and prediction requirements of this stage, resulting in insufficient generalization ability and large prediction bias in field applications. Therefore, it is necessary to readjust the network architecture and hyperparameters and completely re-execute the entire training, validation, and testing process. Through iterative optimization, the model structure and parameters are continuously corrected until the prediction accuracy meets the standard. This mechanism sets an accuracy threshold for the final application of each sub-stage model, ensuring that each sub-stage prediction model has stable and reliable actual prediction performance. This fundamentally avoids errors in the debonding stage division and endpoint prediction bias caused by insufficient model accuracy, thus ensuring the accuracy and practicality of the entire prediction system.

[0074] For example, the prediction accuracy thresholds for each sub-stage are set as follows: 92% for slag production stage 1, 93% for slag production stage 2, 94% for copper production stage 1, 95% for copper production stage 2, and 96% for copper production stage 3. During the testing phase of the deep neural network prediction model for each sub-stage, if the prediction accuracy of the slag production stage 1 model is greater than or equal to 92%, the corresponding deep neural network prediction model is directly retained. If the prediction accuracy of the slag production stage 1 model is less than 92%, the model's network architecture and hyperparameters are readjusted, with the learning rate adjusted sequentially to 0.005 and 0.001, and the hidden layer structure changed from 2 layers with 1... The 28 neurons were optimized into 3 layers with the number of neurons increasing to 256, 128, and 64 respectively, until the prediction accuracy of the slag-making phase I model was greater than or equal to 92%. When the prediction accuracy of the slag-making phase II model was greater than or equal to 93%, it was retained. If it was less than 93%, the network architecture and hyperparameters were adjusted in the same way to meet the standard. The copper-making phase I, copper-making phase II, and copper-making phase III models were selected, adjusted, and retained according to the same adjustment logic, with 94%, 95%, and 96% as the judgment criteria, respectively, to ensure that the models retained in each sub-stage all met the preset prediction accuracy requirements.

[0075] The model invocation module is communicatively connected to the model building module and the data sensors mounted on the copper converter. The model invocation module is used to collect real-time multi-source process data and, based on the real-time multi-source process data, identify the current sub-stage of copper converter blowing. The model invocation module is also used to match the corresponding stage invocation identifier and invoke the prediction model of the current sub-stage. The model invocation module is also used to select the key feature parameters of the current sub-stage from the real-time multi-source process data as model input.

[0076] Specifically, the model invocation module acquires data on melt temperature and temperature fluctuations within the copper converter through a high-temperature resistant melt temperature sensor installed inside the furnace; it acquires real-time data on melt sulfur content, copper matte conversion rate, and crude copper grade through a melt composition detection sensor inside the furnace; it acquires data on SO2 concentration, SO2 concentration change rate, CO concentration, and SO2 to CO ratio through a flue gas composition sensor at the converter flue; it acquires data on the iron-silicon ratio, FeO content, and slag content in the slag through a slag composition detection sensor inside the furnace; it acquires data on air volume and cumulative air volume through an air flow sensor in the air supply duct; it acquires data on oxygen concentration through an oxygen supply concentration sensor in the oxygen supply duct; it acquires data on the quartz addition rate through a quartz addition device quartz addition rate sensor; and it acquires quantified feature data of melt color through a furnace mouth image acquisition sensor. This comprehensive collection of real-time multi-source process data is used to identify the current refining sub-stage and selectively invoke the corresponding prediction model.

[0077] The endpoint prediction module is communicatively connected to the model invocation module. The model invocation module is used to output the core reaction progress index of the current sub-stage based on the invoked prediction model, and to complete the model switching based on the core reaction progress index. The endpoint prediction module is also used to obtain and output the prediction result of the copper converter blowing endpoint based on the core reaction progress index corresponding to the copper production phase III when switching to copper production phase III.

[0078] Specifically, when the endpoint prediction module outputs the core response progress indicators for the current sub-stage through the invoked prediction model, it includes: The core reaction progress indicator corresponding to the first phase of slag formation is the FeO removal rate; The core reaction progress indicator corresponding to the second phase of slag making is the slag phase qualification rate; The core reaction progress indicator corresponding to the first phase of copper production is the copper matte conversion rate; The core reaction progress indicator corresponding to the second phase of copper production is the real-time value of crude copper grade. The core reaction progress indicator corresponding to the third phase of copper production is the probability of achieving the endpoint target.

[0079] It can be understood that the five sub-stages of copper converter blowing each undertake completely different core physicochemical reaction tasks. The first stage of slag formation focuses on oxidative de-ironization and slag formation; the FeO removal rate directly and accurately reflects the progress and completion of the de-ironization reaction in this stage, making it the most critical indicator of the slag formation progress. The second stage focuses on deep slag formation and achieving slag composition standards; the slag phase qualification comprehensively reflects whether the slag iron-silicon ratio and composition meet the transition requirements, directly reflecting the completion progress of the slag formation process. The first stage of copper production focuses on the conversion of copper matte into blister copper; the copper matte conversion rate directly quantifies the progress of copper matte conversion, accurately representing the initial stage of copper production. The reaction process is as follows: Phase II of copper production focuses on the continuous generation and steady improvement of crude copper quality. The real-time value of crude copper grade can intuitively reflect the quality of crude copper production and the completion of the stage. Phase III of copper production focuses on predicting the end point of the smelting process. The probability of reaching the end point can comprehensively predict whether the end point of the smelting process has been reached, which is the core basis for the end point determination. This invention is based on the unique core reaction objectives of each sub-stage, and selects parameters that are highly bound to the stage reaction and can uniquely and accurately characterize the degree of completion of the stage reaction as core reaction progress indicators. This ensures that the indicators are fully adapted to the reaction laws of each stage, and provides a unique and reliable basis for subsequent automatic model switching and accurate prediction of the end point.

[0080] Specifically, when the endpoint prediction module switches models based on the core response progress index, it includes: When the FeO removal rate of the first stage of slag making is greater than or equal to the preset rate threshold, the endpoint prediction module switches to call the prediction model corresponding to the second stage of slag making. When the slag phase qualification rate of the second stage of slag making is greater than or equal to the qualification rate threshold, the endpoint prediction module switches to call the prediction model corresponding to the first stage of copper making. When the copper matte conversion rate of the first phase of copper production is greater than or equal to the conversion threshold, the endpoint prediction module switches to call the prediction model corresponding to the second phase of copper production. When the real-time value of the crude copper grade in the second phase of copper production is greater than or equal to the real-time grade threshold, the endpoint prediction module switches to call the prediction model corresponding to the third phase of copper production.

[0081] It can be understood that the five sub-stages of copper converter smelting follow a sequential and irreversible physicochemical reaction pattern. The core reaction of the previous stage must be fully met before proceeding to the next stage. Therefore, the progress indicators of the core reaction specific to each sub-stage are used as the core criterion. When the FeO removal rate in the first stage of slag formation reaches the preset rate threshold, it indicates that the core task of oxidation, iron removal, and slag formation in this stage has been completed, and the system automatically switches to the second stage of slag formation. When the slag phase qualification rate in the second stage of slag formation reaches the qualification threshold, it indicates that the tasks of deep slag formation and slag composition qualification have been completed, and the system automatically switches to the first stage of copper production. When the copper matte conversion in the first stage of copper production reaches the target value... When the conversion rate reaches the conversion threshold, it signifies the completion of the initial conversion task from copper matte to blister copper, and the system automatically switches to the second-stage copper production model. When the real-time value of the blister copper grade in the second-stage copper production reaches the real-time grade threshold, it signifies that the quality of the blister copper production meets the requirements for the transition stage, and the system automatically switches to the third-stage copper production model. This invention achieves seamless, automatic, and precise model switching through threshold determination of core progress indicators, strictly adhering to the progressive reaction law of the blowing stages, ensuring that each stage only calls the prediction model adapted to its own reaction characteristics, fundamentally avoiding prediction deviations caused by cross-stage model mismatch, and ensuring the continuity and accuracy of prediction throughout the entire process.

[0082] For example: With a preset rate threshold of 0.7% / min, a qualification threshold of 90%, a conversion threshold of 83%, and a real-time grade threshold of 97.8%, the real-time calculation yields the following FeO removal rate for the first stage of slag formation at 0.75% / min, the slag phase qualification rate for the second stage of slag formation at 93%, the copper matte conversion rate for the first stage of copper production at 86%, and the real-time grade of crude copper for the second stage of copper production at 98.1%. When the FeO removal rate for the first stage of slag formation (0.75% / min) is greater than or equal to the preset rate threshold of 0... When the slag phase qualification rate is 0.7% / min, switch to the prediction model corresponding to the second stage of slag making. When the slag phase qualification rate of the second stage of slag making is 93% or greater than or equal to the qualification rate threshold of 90%, switch to the prediction model corresponding to the first stage of copper making. When the copper matte conversion rate of the first stage of copper making is 86% or greater than or equal to the conversion threshold of 83%, switch to the prediction model corresponding to the second stage of copper making. When the real-time value of the crude copper grade of the second stage of copper making is 98.1% or greater than or equal to the real-time grade threshold of 97.8%, switch to the prediction model corresponding to the third stage of copper making.

[0083] Specifically, when the endpoint prediction module switches to the third phase of copper production, based on the core reaction progress indicators corresponding to the third phase of copper production, it obtains and outputs the prediction result of the copper converter blowing endpoint, including: When the probability of reaching the endpoint is greater than or equal to the preset endpoint probability threshold, the real-time value of the crude copper grade is greater than or equal to the target grade threshold, and the time for maintaining the current state is greater than the second preset duration, the endpoint prediction module determines that the copper converter blowing endpoint has been reached and outputs the prediction result of the blowing endpoint being reached.

[0084] It can be understood that the endpoint attainment probability is the probability of the endpoint calculated by the prediction model based on real-time process parameters. It can only reflect the theoretical prediction result and is easily affected by instantaneous data fluctuations and local model deviations. The real-time value of crude copper grade is a direct measured indicator of the quality of crude copper in the furnace. It can verify whether the endpoint quality requirements are truly met from the actual product level. The second preset duration is used to ensure that the endpoint attainment state can be maintained stably, avoiding misjudgment of the endpoint due to instantaneous parameter attainment followed by a rebound and decline. The three form a three-layer judgment logic of "model prediction probability + actual quality measurement + stable duration verification", which complements and verifies each other, completely avoiding the problems of inaccurate prediction, instantaneous attainment, and unstable state that exist in single indicator judgment. It strictly conforms to the process requirements that the endpoint of copper converter blowing must simultaneously meet the requirements of reliable prediction, quality attainment, and stable state, thereby achieving accurate and reliable determination of the blowing endpoint.

[0085] For example: the preset endpoint probability threshold is 95%, the target grade threshold is 98.5%, and the second preset duration is 60s. The real-time endpoint achievement probability of the third stage of copper production is 97%, the real-time value of the crude copper grade is 98.8%, and the duration of continuous maintenance of the achievement state is 70s. When the endpoint achievement probability of the third stage of copper production is 97% greater than or equal to the endpoint probability threshold of 95%, and the real-time value of the crude copper grade is 98.8% greater than or equal to the target grade threshold of 98.5%, and the duration of continuous maintenance of the achievement state is 70s greater than or equal to the second preset duration of 60s, it is determined that the copper converter blowing has reached the endpoint.

[0086] In the above embodiments, the entire copper converter smelting process is finely divided into five sub-stages. For each sub-stage, a dedicated deep neural network prediction model is built based on its unique process characteristics, reaction rules, and prediction requirements. The model is trained, verified, and its hyperparameters are tuned and optimized by combining the key feature parameters of each sub-stage. This completely abandons the traditional approach of adapting a single model to the entire smelting process, allowing the model to be highly adapted to the reaction characteristics of each sub-stage. This effectively avoids the deviation caused by uniform prediction of the entire process, fundamentally improving the prediction accuracy of the reaction progress and smelting endpoint of each sub-stage, and achieving refined and high-precision prediction of the copper converter smelting process.

[0087] Reference Figure 2 As shown, in some embodiments of this application, a method for predicting the endpoint of copper converter blowing based on a deep neural network is also provided, including: Historical multi-source process data is acquired, and the entire copper converter blowing process is divided into five sub-stages based on the historical multi-source process data; Key feature parameters are selected and extracted for each sub-stage, and a prediction model is built for each sub-stage based on the key feature parameters; in the verification stage, the hyperparameters of the prediction models for each sub-stage are tuned based on the verification accuracy; in the testing stage, the prediction models for each sub-stage are optimized based on the prediction accuracy; the optimized prediction models are integrated into a unified prediction system framework, and a unique stage call identifier is configured for each prediction model. Collect real-time multi-source process data, and based on the real-time multi-source process data, identify the current sub-stage of copper converter blowing; match the corresponding stage call identifier, and call the prediction model of the current sub-stage in a targeted manner; select the key feature parameters of the current sub-stage from the real-time multi-source process data as model input; The prediction model is invoked to output the core reaction progress indicators of the current sub-stage, and the model is switched based on the core reaction progress indicators. When switching to the third stage of copper production, the prediction results of the copper converter blowing endpoint are obtained and output based on the core reaction progress indicators corresponding to the third stage of copper production.

[0088] It is understandable that the aforementioned method and system for predicting the end point of copper converter blowing based on deep neural networks have the same beneficial effects, and will not be elaborated further here.

[0089] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A copper converter blowing endpoint prediction system based on deep neural networks, characterized in that, include: The stage division module is used to acquire historical multi-source process data and divide the entire copper converter blowing process into five sub-stages based on the historical multi-source process data. The model building module is communicatively connected to the stage division module. The model building module is used to filter and extract key feature parameters for each sub-stage, and build a prediction model for each sub-stage based on the key feature parameters and a deep neural network. The model building module is also used to perform hyperparameter tuning on the prediction models for each sub-stage based on the verification accuracy during the verification phase. Furthermore, the model building module is used to optimize the prediction models for each sub-stage based on the prediction accuracy during the testing phase. Finally, the model building module is used to integrate the optimized prediction models into a unified prediction system framework and configure a unique stage invocation identifier for each prediction model. The model invocation module is communicatively connected to the model building module and the data sensors mounted on the copper converter. The model invocation module is used to collect real-time multi-source process data and, based on this data, identify the current sub-stage of the copper converter's blowing process. The module is also used to match the corresponding stage invocation identifier and invoke the prediction model for the current sub-stage. Furthermore, the module is used to select the key feature parameters of the current sub-stage from the real-time multi-source process data as model input. The endpoint prediction module is communicatively connected to the model invocation module. The model invocation module is used to output the core reaction progress index of the current sub-stage based on the invoked prediction model, and to complete the model switching based on the core reaction progress index. The endpoint prediction module is also used to obtain and output the prediction result of the copper converter blowing endpoint based on the core reaction progress index corresponding to the copper production phase III when switching to copper production phase III.

2. The copper converter blowing endpoint prediction system based on deep neural networks according to claim 1, characterized in that, When the stage division module divides the entire copper converter blowing process into five sub-stages based on the historical multi-source process data, it includes: When the melt temperature is greater than or equal to the first melt temperature threshold, the SO2 concentration is greater than or equal to the first SO2 concentration threshold, and the time for maintaining the current state is greater than or equal to the first preset duration, the stage division module determines that the current sub-stage is the first stage of slag making. When the SO2 concentration is less than the first SO2 concentration threshold, and the SO2 concentration is greater than the second SO2 concentration threshold, and the melt temperature is greater than or equal to the second melt temperature threshold, and the iron-silicon ratio in the slag is greater than or equal to the first iron-silicon ratio threshold, the stage division module determines that the current sub-stage is the second stage of slag making. When the SO2 concentration is less than or equal to the second SO2 concentration threshold, the melt temperature is greater than or equal to the third melt temperature threshold, and the iron-silicon ratio in the slag is greater than or equal to the second iron-silicon ratio threshold, the stage division module determines that the current sub-stage is copper production phase one. When the sulfur content of the melt is less than or equal to the first sulfur content threshold of the melt, and the CO concentration is in the first CO concentration range, and the cumulative air supply is greater than or equal to the first cumulative air supply threshold, the stage division module determines that the current sub-stage is the second phase of copper production. When the sulfur content of the melt is less than or equal to the second melt sulfur content threshold, the melt temperature is greater than or equal to the fourth melt temperature threshold, and the cumulative air supply is greater than or equal to the second cumulative air supply threshold, the stage division module determines that the current sub-stage is the third phase of copper production.

3. The copper converter blowing endpoint prediction system based on deep neural networks according to claim 2, characterized in that, The model building module, when selecting and extracting key feature parameters for each sub-stage, includes: If the current sub-stage is in the first stage of slag formation, then the key characteristic parameters are SO2 concentration, SO2 concentration change rate, furnace melt temperature, air supply volume and quartz stone addition rate. If the current sub-stage is in the second stage of slag formation, then the key characteristic parameters are SO2 concentration, furnace melt temperature fluctuation value, slag FeO content, slag SiO2 content, and oxygen supply concentration. If the current sub-stage is the first stage of copper production, then the key characteristic parameters are CO concentration, CO2 to CO ratio, furnace melt temperature, oxygen supply concentration, and copper matte conversion rate. If the current sub-stage is in the second phase of copper production, then the key characteristic parameters are: stable CO concentration, real-time crude copper grade, quantitative characteristics of melt color, melt temperature in the furnace, and oxygen supply concentration. If the current sub-stage is in the third phase of copper production, the key characteristic parameters are the real-time value of crude copper grade, CO concentration, CO2 to CO ratio, furnace melt temperature fluctuation value, and O2 concentration stability value.

4. The copper converter blowing endpoint prediction system based on deep neural networks according to claim 3, characterized in that, The model building module, based on the key feature parameters and deep neural network, builds prediction models for each sub-stage, including: The model building module constructs a corresponding key feature parameter dataset for each sub-stage based on the key feature parameters extracted from each sub-stage. The model building module preprocesses the key feature parameter datasets of each sub-stage and divides the preprocessed key feature parameter datasets into training set, validation set and test set. The model building module designs prediction models for each sub-stage based on the feature dimensions and prediction targets of each sub-stage. The model building module trains the prediction model corresponding to each sub-stage based on the training set of each sub-stage. The model building module validates the prediction model corresponding to each sub-stage based on the validation set of each sub-stage, obtains the validation accuracy, and performs hyperparameter tuning on the prediction model of each sub-stage based on the validation accuracy. The model building module tests the prediction model corresponding to each sub-stage based on the test set of each sub-stage, obtains the prediction accuracy, and optimizes the prediction model of each sub-stage based on the prediction accuracy.

5. The copper converter blowing endpoint prediction system based on deep neural networks according to claim 4, characterized in that, When the model building module performs hyperparameter tuning on the prediction model for each sub-stage based on the verification accuracy, it includes: When the verification accuracy of a certain sub-stage is less than the corresponding verification accuracy threshold within a consecutive preset number of rounds, the model building module performs hyperparameter tuning on the prediction model corresponding to the current sub-stage, and retrains and verifies the prediction model based on the training set and the validation set until the verification accuracy is greater than or equal to the corresponding verification accuracy threshold within a consecutive preset number of rounds. If the verification accuracy of a certain sub-stage is greater than or equal to the corresponding verification accuracy threshold within a consecutive preset number of rounds, the model building module keeps the hyperparameters of the current prediction model unchanged.

6. The copper converter blowing endpoint prediction system based on deep neural networks according to claim 5, characterized in that, When the model building module optimizes the prediction model for each sub-stage based on the prediction accuracy, it includes: When the prediction accuracy of a certain sub-stage is greater than or equal to the prediction accuracy threshold corresponding to the current sub-stage, the model building module keeps the network architecture and hyperparameters of the current prediction model unchanged. When the prediction accuracy of a certain sub-stage is less than the prediction accuracy threshold corresponding to the current sub-stage, the model building module readjusts the network architecture and hyperparameters of the prediction model, and retrains, validates and tests the prediction model until the prediction accuracy of the current sub-stage is greater than or equal to the prediction accuracy threshold corresponding to the current sub-stage.

7. The copper converter blowing endpoint prediction system based on deep neural networks according to claim 6, characterized in that, When the endpoint prediction module outputs the core response progress indicators for the current sub-stage through the invoked prediction model, it includes: The core reaction progress indicator corresponding to the first phase of slag formation is the FeO removal rate; The core reaction progress indicator corresponding to the second phase of slag making is the slag phase qualification rate; The core reaction progress indicator corresponding to the first phase of copper production is the copper matte conversion rate; The core reaction progress indicator corresponding to the second phase of copper production is the real-time value of crude copper grade. The core reaction progress indicator corresponding to the third phase of copper production is the probability of achieving the endpoint target.

8. The copper converter blowing endpoint prediction system based on deep neural networks according to claim 7, characterized in that, When the endpoint prediction module switches models based on the core response progress index, it includes: When the FeO removal rate of the first stage of slag making is greater than or equal to the preset rate threshold, the endpoint prediction module switches to call the prediction model corresponding to the second stage of slag making. When the slag phase qualification rate of the second stage of slag making is greater than or equal to the qualification rate threshold, the endpoint prediction module switches to call the prediction model corresponding to the first stage of copper making. When the copper matte conversion rate of the first phase of copper production is greater than or equal to the conversion threshold, the endpoint prediction module switches to call the prediction model corresponding to the second phase of copper production. When the real-time value of the crude copper grade in the second phase of copper production is greater than or equal to the real-time grade threshold, the endpoint prediction module switches to call the prediction model corresponding to the third phase of copper production.

9. A copper converter blowing endpoint prediction system based on deep neural networks according to claim 8, characterized in that, When the endpoint prediction module switches to copper production phase III, based on the core reaction progress indicators corresponding to copper production phase III, it obtains and outputs the prediction result of the copper converter blowing endpoint, including: When the probability of reaching the endpoint is greater than or equal to the preset endpoint probability threshold, the real-time value of the crude copper grade is greater than or equal to the target grade threshold, and the time for maintaining the current state is greater than the second preset duration, the endpoint prediction module determines that the copper converter blowing endpoint has been reached and outputs the prediction result of the blowing endpoint being reached.

10. A method for predicting the endpoint of copper converter blowing based on deep neural networks, characterized in that, The copper converter blowing endpoint prediction system based on deep neural networks, as described in any one of claims 1-9, comprises: Historical multi-source process data is acquired, and the entire copper converter blowing process is divided into five sub-stages based on the historical multi-source process data; Key feature parameters are selected and extracted for each sub-stage, and a prediction model is built for each sub-stage based on the key feature parameters; in the verification stage, the hyperparameters of the prediction models for each sub-stage are tuned based on the verification accuracy; in the testing stage, the prediction models for each sub-stage are optimized based on the prediction accuracy; the optimized prediction models are integrated into a unified prediction system framework, and a unique stage call identifier is configured for each prediction model. Collect real-time multi-source process data, and based on the real-time multi-source process data, identify the current sub-stage of copper converter blowing; match the corresponding stage call identifier, and call the prediction model of the current sub-stage in a targeted manner; select the key feature parameters of the current sub-stage from the real-time multi-source process data as model input; The prediction model is invoked to output the core reaction progress indicators of the current sub-stage, and the model is switched based on the core reaction progress indicators. When switching to the third stage of copper production, the prediction results of the copper converter blowing endpoint are obtained and output based on the core reaction progress indicators corresponding to the third stage of copper production.

Images