An online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace.

Abstract

This invention discloses an online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace, belonging to the field of copper concentrate engineering technology. It includes a batching system, a smelting furnace, and a data processing unit. The batching system and the smelting furnace are connected via a conveyor belt. A belt sampler is installed in the feeding section of the conveyor belt, along with a belt load sensor and a speed sensor. The smelting furnace is equipped with a plug-in thermocouple and an in-situ infrared sensor. The belt load sensor, speed sensor, plug-in thermocouple, and in-situ infrared sensor are all connected to the data processing unit. The matrix composition-detection parameter correction model can output optimal detection parameters within one second without manual intervention, adapting to the detection of samples with different matrix compositions and enabling adaptive adjustment. It can effectively counteract the interference of matrix components such as silicon, aluminum, and calcium on the laser spectrum, significantly improving the detection accuracy of impurity elements and specifically addressing the matrix effect; balancing efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of copper concentrate engineering technology, specifically to an online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace. Background Technology

[0002] The side-blown smelting furnace is the core equipment for copper smelting. The quality parameters of the mixed copper concentrate fed into the furnace (including the content of main elements such as copper and sulfur, the content of impurities such as lead and zinc, and physical properties such as moisture and particle size distribution) directly determine the stability of the smelting process and production efficiency. For example, too low sulfur content will lead to insufficient furnace temperature, too high impurities will easily cause flue gas duct nodules, excessive moisture will affect the permeability of the material layer, and uneven particle size distribution will cause abnormal agitation of the melt.

[0003] Currently, the industry mostly adopts offline detection mode, which involves manual sampling followed by wet chemical analysis for elemental analysis, drying analysis for moisture analysis, and laser diffraction analysis for particle size analysis. A single test takes more than 120 minutes and has serious time lag. When the raw material composition fluctuates, it is impossible to adjust the batching ratio and smelting parameters in time, which often leads to problems such as furnace temperature fluctuation, abnormal liquid level, and slag deterioration. This not only reduces copper recovery rate but also increases energy consumption and pollutant emissions. Although online detection technologies such as laser spectroscopy are used in mineral processing, there is still a lack of solutions for continuous sampling and pretreatment of the charge in side-blown smelting furnaces, simultaneous detection of multiple parameters, and process linkage control. Summary of the Invention

[0004] To address or partially address the problems existing in related technologies, this invention provides an online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace. This method enables continuous sampling and pretreatment of the material fed into the side-blown smelting furnace, simultaneous detection of multiple parameters, and process linkage control.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: An online detection and control system for mixed copper concentrate fed into a side-blown smelting furnace is provided. Its innovation lies in that it includes a batching system, a smelting furnace, and a data processing unit. The batching system and the smelting furnace are connected via a conveyor belt. A belt sampler is installed in the feeding section of the conveyor belt, along with a belt load sensor and a speed sensor. A plug-in thermocouple and an in-situ infrared sensor are installed inside the smelting furnace. The belt load sensor, speed sensor, plug-in thermocouple, and in-situ infrared sensor are all connected to the data processing unit. It also includes a near-infrared spectroscopy submodule and a laser-induced breakdown spectroscopy module. The sampler is connected to a first crusher, which is connected to a divider. The divider is divided into two paths: one path is connected to the near-infrared spectroscopy submodule and the laser-induced breakdown spectroscopy module, and the other path is connected to a second crusher. The second crusher is connected to a microwave moisture meter and a laser particle size analyzer. The laser-induced breakdown spectroscopy module, the microwave moisture meter, the laser particle size analyzer, and the second crusher are all connected to a data processing unit. The data processing unit is also connected to the sampler and the batching system.

[0006] Preferably, the data processing unit has a built-in LSTM prediction model and a matrix composition-detection parameter correction model. The matrix composition-detection parameter correction model is input based on the data from the near-infrared spectroscopy submodule, and the output data is used as the parameters of the laser-induced breakdown spectroscopy module.

[0007] This invention also provides an online detection and control method for mixed copper concentrate fed into a side-blown smelting furnace according to the above system, the innovation of which lies in the following steps: Step S1, Dynamic Compensation Automatic Sampling: The belt load and conveying speed are collected in real time by the belt load sensor and speed sensor. The sampling frequency and sampling amount are determined according to the belt load, and the sample is taken by the sampler. Step S2, Matrix Correction - Elemental Co-detection: S21. Sample pretreatment: The sample taken by the sampler is crushed by the first crusher, and then the crushed sample is reduced to 100-200g by the divider. After removing impurities, it is divided into two parts, A and B. Part A is used for elemental detection and part B is used for physical property detection. S22, Adaptive correction of matrix effect: The matrix composition of sample A is detected in real time by the near-infrared spectroscopy module. Based on the matrix composition of sample A, the matrix composition-detection parameter correction model is called to dynamically adjust the laser energy and spectral integration time of the laser-induced breakdown spectrum. After adjustment, sample A is scanned by the laser-induced breakdown spectroscopy module to detect the content of copper, sulfur, iron main elements and lead, zinc and arsenic impurity elements in sample A. Step S3, Block Management - Comprehensive Physical Property Detection: S31, Agglomeration Crushing-Screening Linkage: Send sample B into the second crusher, crush it, sieve it, collect the sieved sample, and record the agglomeration rate; S32. Physical property testing: Take the sieved sample and determine the moisture content using a microwave moisture meter, and at the same time use a laser particle size analyzer to detect the particle size distribution; store the data of agglomeration rate, moisture content and particle size distribution together. Step S4: Predictive data fusion analysis: S41 Outlier Removal and Standard Comparison: Based on element content, agglomeration rate, moisture content, and particle size data, outliers are removed using the 3σ criterion and compared with preset standard values ​​to determine the sample pass rate. The data processing unit automatically generates a test report based on the comparison results. The report includes the test data and comparison results for element content, agglomeration rate, moisture content, and particle size distribution. S42. Multi-parameter trend prediction and early warning: Based on historical 6 months of detection data and real-time data, including element content, agglomeration rate, moisture, and particle size data, the LSTM prediction model is called to predict the component qualification rate, smelting furnace temperature trend, and agglomeration rate change in the next 1-2 hours; if the predicted value exceeds the preset standard value, an early warning is triggered, and the data processing unit automatically matches the cause of the deviation. S43. Update the process optimization knowledge base: Link and store the test data, early warning results, subsequent control measures and smelting indicators, update the process optimization knowledge base, and use it to recommend control parameters for similar working conditions in the future. Step S5: Dual closed-loop feedback control of furnace operating conditions: S51. Basic Parameter Control: Based on the comparison and prediction results, the data processing unit outputs a detection report and control signals for the feeding rate, oxygen enrichment concentration, and batching ratio, which are transmitted to the batching system and the instrumentation and control station of the smelting furnace to adjust the material ratio, feeding rate, and oxygen enrichment concentration in real time; among which, the feeding rate adjustment range is 0-5t / h, and the oxygen enrichment concentration adjustment range is 21-30%; S52. Coordinated Control of Furnace Operating Conditions: Real-time temperature inside the smelting furnace is collected via insertion thermocouples, and SO2 concentration data in flue gas is collected via in-situ infrared sensors, forming a dual closed loop with the detection data. When the sulfur content is greater than 35% and the SO2 concentration in the flue gas is greater than 2000 ppm, the proportion of high-sulfur concentrate should be reduced, and the oxygen enrichment concentration should be reduced by 1-2% simultaneously. When the agglomeration rate is greater than 3% and the temperature fluctuation inside the furnace is greater than 30℃ / h, increase the crushing roller speed of the crusher and simultaneously reduce the feed rate by 0.5-1t / h.

[0008] Preferably, in step S1, the unit of belt load is t / h, and the unit of conveying speed is m / s; the standard for determining the sampling frequency and sampling amount based on the belt load is as follows: when the load is ≤10t / h, samples are taken at 10-minute intervals, with a sampling amount of 200g / batch; when 10t / h < load ≤20t / h, samples are taken at 8-minute intervals, with a sampling amount of 350g / batch; when the load is >20t / h, samples are taken at 5-minute intervals, with a sampling amount of 500g / batch; the belt load-sampling parameter correspondence is recorded synchronously during the sampling process.

[0009] Preferably, in step S1, dynamic sampling compensation calibration is performed during the sampling process. Specifically, the belt load sensor data is calibrated every 8 hours using a belt scale. When the load detection error is not within 3%, the data processing unit automatically corrects the sampling frequency and sampling amount to ensure that the sampling representativeness matches the dynamic load of the belt.

[0010] Preferably, in step S21, the particle size of the crushed sample is ≤3mm; the sample is reduced by a gravity-type divider with an adjustable reduction ratio of 1:5-1:10.

[0011] Preferably, in step S22, the matrix composition is the content of silicon, aluminum and calcium. The laser energy is dynamically adjusted to 50-100mJ and the spectral integration time is 10-30ms by a matrix composition-detection parameter correction model. The detection time by laser-induced breakdown spectroscopy is ≤5 minutes. The detection limit for arsenic is ≤0.05%, the repeatability RSD is ≤2.5%, and the repeatability RSD for other elements is ≤3%.

[0012] Preferably, in step S22, the construction process of the matrix composition-detection parameter correction model is as follows: S221. Prepare matrix standard samples and construct a diverse standard sample library. Matrix composition gradient design: Based on the content gradient of silicon (0.1%, 0.5%, 1%, 3%, or 5%), aluminum (0.05%, 0.1%, 0.5%, 2%, or 3%), and calcium (0.02%, 0.05%, 0.1%, 1%, or 2%), samples with more than m different matrix ratios are formed by combining two or three substances in different proportions, where m > 500. Sample homogeneity and stability treatment: After each group of samples is crushed and mixed, the homogeneity is verified by sampling-testing-comparison to ensure that the homogeneity error is within 1%. The samples are sealed and stored to avoid moisture and component shift, ensuring the stability of sample characteristics. Target element content fixation: In each set of standard samples, the content of main elements such as copper, sulfur, and iron, as well as impurity elements such as lead, zinc, and arsenic, is fixed according to industrial standards. Only the proportion of matrix components is changed to avoid the target element content from interfering with model modeling. S222. Multi-parameter combination calibration to obtain matrix-parameter-accuracy correlation data. For each of the more than m groups of standard samples in step S221, the detection parameters are calibrated to establish the correspondence between the matrix composition and the optimal detection parameters. Specifically: Detection parameter variable settings: Laser energy of 50mJ, 60mJ, 70mJ, 80mJ, 90mJ, and 100mJ, and spectral integration time of 10ms, 15ms, 20ms, 25ms, and 30ms are selected as core variables, forming 6×5=30 parameter combinations; Full parameter combination detection: For each group of standard samples, laser-induced breakdown spectroscopy detection was performed using 30 parameter combinations. The detection results of the target elements under each parameter combination were recorded, including the detection value of each element, repeatability RSD, and detection time. Optimal parameter screening criteria: For each sample group, the parameter combination with the smallest detection repeatability RSD and detection time ≤ 5 minutes is selected as the optimal detection parameters for that matrix component; Data recording and organization: The final result is a three-dimensional associated dataset of m groups of matrix components, optimal detection parameters, and detection accuracy, which serves as the basic data for model training; where the matrix components are the contents of silicon, aluminum, and calcium; the optimal detection parameters are the corresponding laser energy and integration time; and the detection accuracy is the detection repeatability RSD. S223. Model building, training and optimization, establishing nonlinear mapping relationships. Based on the 3D correlation dataset from step S222, a machine learning algorithm is used to construct a model to achieve a nonlinear mapping from matrix components to optimal detection parameters, specifically: Algorithm selection: First, a linear relationship between matrix components and detection parameters is initially established through multiple linear regression, and then nonlinear errors are corrected through gradient boosting trees; thus forming a model that can achieve nonlinear mapping between matrix components and optimal detection parameters. Model input and output definitions: Input features: silicon content, aluminum content, and calcium content in the associated dataset from step S222; Output labels: Optimal laser energy and optimal spectral integration time in the associated dataset of step S222.

[0013] Training and validation split: The m sets of 3D association datasets are split in a 7:3 ratio, with 70% used as the training set for model fitting and 30% used as the validation set for accuracy verification; the training set is first input into the model formed by the combination algorithm of multiple linear regression and gradient boosting tree to initially fit the mapping relationship between matrix components and detection parameters; Model optimization objectives: The goal is to achieve a deviation of ≤5% between predicted parameters and actual optimal parameters in the validation set, an RSD of ≤2.5% for arsenic detection repeatability, and an RSD of ≤3% for other elements. The model is optimized by adjusting the algorithm hyperparameters to ensure that the model's parameter prediction accuracy on the validation set is ≥95%. The algorithm hyperparameters adjustment includes adjusting the tree depth and learning rate of the gradient boosting tree, and the regularization coefficient of the multiple linear regression. The model error is iteratively corrected until the model meets the detection accuracy requirements, thus completing the training optimization.

[0014] Preferably, in step S31, the crushing roller speed of the second crusher is 50-150 r / min, and after crushing, it is screened through a vibrating screen with a 1 mm aperture. The agglomeration rate is calculated as (mass before screening - mass after screening) / mass before screening × 100%. In step S32, the laser particle size analyzer has a measurement range of 0.02-2000μm, and the repeatability accuracy of the measurement is within 1%. The particle size distribution is detected, including the proportion of particles of -200 mesh. The agglomeration rate, moisture content, and particle size distribution data are stored together. Preferably, in step S41, outliers are values ​​that deviate from the average data by more than three standard deviations; the preset standard values ​​are: copper content ≥15%, sulfur content 8-35%, moisture content ≤12%, -200 mesh particle size >80%, total lead and zinc content <5%, and agglomeration rate ≤3%; In step S43, the smelting indicators include the grade of crude copper and the copper content in the slag.

[0015] This invention provides an online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace, which has the following beneficial effects: (1) In actual detection, the matrix composition-detection parameter correction model of the present invention can obtain the sample matrix composition in real time through the near-infrared spectroscopy submodule. The model can output the optimal detection parameters within 1 second without manual intervention. It can adapt to the detection of samples with different matrix compositions and achieve adaptive adjustment. Through modeling of more than 500 sets of matrix samples with full coverage, it can effectively offset the interference of matrix components such as silicon, aluminum, and calcium on the laser spectrum, significantly improve the detection accuracy of impurity elements (such as arsenic) (detection limit ≤ 0.05%, RSD ≤ 2.5%), and specifically solve the matrix effect. The optimal parameters selected by the model simultaneously meet the requirements of "detection time ≤ 5 minutes" and "minimum repeatability RSD", which not only meets the real-time detection needs of continuous industrial production, but also ensures the reliability of detection data, taking into account both efficiency and accuracy.

[0016] (2) Eliminating detection lag: The detection cycle is shortened from more than 2 hours in the traditional offline mode to ≤5 minutes, realizing real-time monitoring of the quality of mixed copper concentrate, solving the industry pain point of asynchronous raw material fluctuations and process adjustments, and improving detection accuracy. (3) Stable smelting process: This invention, through real-time feedback control, can precisely control the S / Cu ratio, impurity content, and physical properties of the charge, effectively maintaining the furnace temperature and liquid level within a stable range, and reducing the risk of melt splashing and flue gas duct nodule formation. (4) This invention, based on precise detection and ingredient optimization, can increase copper recovery rate by 0.5-1.0 percentage points, while reducing fuel consumption and waste emissions. Combined with automated sampling, it reduces manual operation and significantly lowers overall production costs. (5) The automatic storage and traceability function of the detection data of the present invention can support long-term process optimization analysis and realize intelligent traceability to provide core technical support for the intelligent upgrading of side-blown smelting furnace.

[0017] (6) The model of the present invention is constructed based on 500+ sets of samples, covering common industrial matrix types, and supports continuous updates by supplementing new sample data, adapting to the detection of mixed copper concentrates with different production areas and different compositions. Attached Figure Description

[0018] Figure 1This is a system block diagram of an online detection and control system for mixed copper concentrate fed into a side-blown smelting furnace according to the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings, but the scope of protection of the present invention is not limited to the content described.

[0020] This invention provides an online detection and control system for mixed copper concentrate fed into a side-blown smelting furnace, the system block diagram of which is shown below. Figure 1 As shown, it includes a batching system, a smelting furnace, and a data processing unit. The batching system and the smelting furnace are connected by a conveyor belt. A belt sampler is installed in the material feeding section of the conveyor belt, and a belt load sensor and a speed sensor are installed. A plug-in thermocouple and an in-situ infrared sensor are installed in the smelting furnace. The belt load sensor, speed sensor, plug-in thermocouple, and in-situ infrared sensor are all connected to the data processing unit. The invention also includes a near-infrared spectroscopy submodule and a laser-induced breakdown spectroscopy module. The sampler is connected to a first crusher, which is connected to a particle size analyzer. The particle size analyzer is divided into two paths: one path is connected to the near-infrared spectroscopy submodule and the laser-induced breakdown spectroscopy module, and the other path is connected to a second crusher. The second crusher is connected to a microwave moisture analyzer and a laser particle size analyzer. The laser-induced breakdown spectroscopy module, the microwave moisture analyzer, the laser particle size analyzer, and the second crusher are all connected to a data processing unit. The data processing unit is also connected to the sampler and the batching system.

[0021] The data processing unit incorporates an LSTM prediction model and a matrix composition-detection parameter correction model. Data from the near-infrared spectroscopy submodule is input into the matrix composition-detection parameter correction model, and the output data is used as parameters for the laser-induced breakdown spectroscopy module. The data processing unit is integrated into the PLC controller.

[0022] The present invention also provides an online detection and control method for mixed copper concentrate fed into a side-blown smelting furnace according to the above system, comprising the following steps: Step S1, Dynamic Compensation Automatic Sampling: The belt load and conveying speed are collected in real time by the belt load sensor and speed sensor. The sampling frequency and sampling amount are determined according to the belt load, and the sample is taken by the sampler. The unit for belt load is t / h, and the unit for conveyor speed is m / s. The sampling frequency and sampling amount are determined according to the belt load as follows: when the load is ≤10t / h, samples are taken at 10-minute intervals, with a sampling amount of 200g / batch; when 10t / h < load ≤20t / h, samples are taken at 8-minute intervals, with a sampling amount of 350g / batch; when the load is >20t / h, samples are taken at 5-minute intervals, with a sampling amount of 500g / batch. The belt load-sampling parameter correspondence is recorded synchronously during the sampling process.

[0023] During the sampling process, dynamic sampling compensation calibration is performed. Specifically, the belt load sensor data is calibrated every 8 hours using a belt scale. When the load detection error is not within 3%, the data processing unit automatically corrects the sampling frequency and sampling amount to ensure that the sampling representativeness matches the dynamic load of the belt.

[0024] Step S2, Matrix Correction - Elemental Co-detection: S21. Sample Pretreatment: The sample taken by the sampler is crushed by the first crusher, and then the crushed sample is reduced to 100-200g by the reducer. After removing impurities, it is divided into two parts, A and B. Part A is used for elemental analysis, and part B is used for physical property analysis. The particle size of the crushed sample is ≤3mm. The sample is reduced by a gravity reducer with an adjustable reduction ratio of 1:5-1:10.

[0025] S22, Adaptive correction of matrix effect: The matrix composition of sample A is detected in real time by the near-infrared spectroscopy module. Based on the matrix composition of sample A, the matrix composition-detection parameter correction model is called to dynamically adjust the laser energy and spectral integration time of the laser-induced breakdown spectrum. After adjustment, sample A is scanned by the laser-induced breakdown spectroscopy module to detect the content of copper, sulfur, iron main elements and lead, zinc and arsenic impurity elements in sample A. The matrix composition is determined by the content of silicon, aluminum, and calcium. The laser energy is dynamically adjusted to 50-100 mJ and the spectral integration time is 10-30 ms using a matrix composition-detection parameter correction model. The detection time by laser-induced breakdown spectroscopy is ≤5 minutes. The detection limit for arsenic is ≤0.05%, the repeatability RSD is ≤2.5%, and the repeatability RSD for other elements is ≤3%.

[0026] The construction process of the matrix composition-detection parameter calibration model is as follows: S221. Prepare matrix standard samples and construct a diverse standard sample library. Matrix composition gradient design: Based on the content gradient of silicon (0.1%, 0.5%, 1%, 3%, or 5%), aluminum (0.05%, 0.1%, 0.5%, 2%, or 3%), and calcium (0.02%, 0.05%, 0.1%, 1%, or 2%), samples with more than m different matrix ratios are formed by combining two or three substances in different proportions, where m > 500. Sample homogeneity and stability processing: After crushing and homogenizing each group of samples, homogeneity is verified through a sample division-detection-comparison process to ensure that the homogeneity error is within 1%. The samples are then sealed and stored to prevent moisture and compositional shift, ensuring stable sample characteristics. Sample division-detection-comparison verification is a standard homogeneity verification method in the field of copper concentrate testing. The routine procedure involves randomly dividing the homogenized sample into multiple parallel samples, measuring the same index (such as major element content) using the same detection method, and calculating the relative deviation by comparing the test results of the parallel samples to verify sample homogeneity.

[0027] Target element content fixation: In each set of standard samples, the content of main elements such as copper, sulfur, and iron, as well as impurity elements such as lead, zinc, and arsenic, is fixed according to industrial standards (e.g., copper 15-25%, sulfur 8-35%, arsenic 0.01-0.1%), only the matrix composition ratio is changed to avoid the target element content from interfering with model modeling. S222. Multi-parameter combination calibration to obtain matrix-parameter-accuracy correlation data. For each of the more than m groups of standard samples in step S221, the detection parameters are calibrated to establish the correspondence between the matrix composition and the optimal detection parameters. Specifically: Detection parameter variable settings: Laser energy of 50mJ, 60mJ, 70mJ, 80mJ, 90mJ, and 100mJ, and spectral integration time of 10ms, 15ms, 20ms, 25ms, and 30ms are selected as core variables, forming 6×5=30 parameter combinations; Full parameter combination detection: For each group of standard samples, laser-induced breakdown spectroscopy detection was performed using 30 parameter combinations. The detection results of the target elements under each parameter combination were recorded, including the detection value of each element, repeatability RSD, and detection time. Optimal parameter selection criteria: For each sample group, the parameter combination with the smallest detection repeatability RSD and detection time ≤ 5 minutes is selected as the optimal detection parameters for that matrix component (e.g., for samples with 3%+ silicon, 2%+ aluminum, and 1% calcium, the optimal parameters are laser energy of 80mJ and integration time of 20ms). Data recording and organization: The final result is a three-dimensional associated dataset consisting of m groups of matrix components (Si / Al / Ca content), optimal detection parameters (optimal laser energy and optimal integration time), and detection accuracy, which serves as the basic data for model training; where the matrix components are the contents of silicon, aluminum, and calcium; the optimal detection parameters are the corresponding laser energy and integration time; and the detection accuracy is the detection repeatability RSD. S223. Model building, training and optimization, establishing nonlinear mapping relationships. Based on the 3D correlation dataset from step S222, a machine learning algorithm is used to construct a model to achieve a nonlinear mapping from matrix components to optimal detection parameters, specifically: Algorithm selection: First, a linear relationship between matrix components and detection parameters is initially established through multiple linear regression, and then nonlinear errors are corrected through gradient boosting trees; thus forming a model that can achieve nonlinear mapping between matrix components and optimal detection parameters. Model input and output definitions: Input features: silicon content, aluminum content, and calcium content in the associated dataset from step S222 (matrix composition data in three dimensions, all of which have been normalized). Output labels: Optimal laser energy and optimal spectral integration time (two-dimensional detection parameters, modeled according to the original values) in the associated dataset of step S222.

[0028] Training and validation split: The m sets of 3D association datasets are split in a 7:3 ratio, with 70% used as the training set for model fitting and 30% used as the validation set for accuracy verification; the training set is first input into the model formed by the combination algorithm of multiple linear regression and gradient boosting tree to initially fit the mapping relationship between matrix components and detection parameters; Model optimization objectives: The goal is to achieve a deviation of ≤5% between predicted parameters and actual optimal parameters in the validation set, an RSD of ≤2.5% for arsenic detection repeatability, and an RSD of ≤3% for other elements. The model is optimized by adjusting the algorithm hyperparameters to ensure that the model's parameter prediction accuracy on the validation set is ≥95%. The algorithm hyperparameters adjustment includes adjusting the tree depth and learning rate of the gradient boosting tree, and the regularization coefficient of the multiple linear regression. The model error is iteratively corrected until the model meets the detection accuracy requirements, thus completing the training optimization.

[0029] S224, Model Validation and Deployment Adaptation After completing model training and optimization, this invention verifies the model's effectiveness using actual samples and adapts it to industrial testing scenarios. Experimental verification: Ten groups of actual production samples not involved in modeling were selected, with matrix components within the model's coverage range. The sample matrix components were input into the trained model to obtain the predicted optimal laser energy and spectral integration time. Laser-induced breakdown spectroscopy was performed using these parameters, and the results were compared with laboratory inductively coupled plasma mass spectrometry (ICP-MS) results. The accuracy of the model's predicted parameters was verified if the detection deviation for the target elements copper, sulfur, iron, and impurity elements lead, zinc, and arsenic was ≤3% (e.g., arsenic detection deviation ≤0.005%).

[0030] On-site adaptation: During on-site testing, the near-infrared spectroscopy submodule collects the matrix composition of the copper concentrate entering the furnace in real time and inputs it into the model. The model outputs the optimal detection parameters within 1 second and synchronizes them to the laser-induced breakdown spectroscopy module to automatically execute the detection. On-site detection data is continuously collected. If the detection deviation is greater than 3%, the sample data of the corresponding matrix composition is supplemented, and the model parameters are retrained and optimized to achieve accurate adaptation of the model to the industrial site conditions.

[0031] The core objective of this model is to eliminate the interference of the "matrix effect"—that is, the difference in the content of matrix components such as silicon (Si), aluminum (Al), and calcium (Ca) in the sample will lead to different spectral signal intensities and stability under the same detection parameters, thus affecting the detection accuracy of elements such as copper, sulfur, and arsenic. The model can automatically match the optimal detection parameters based on the real-time acquired matrix component data to ensure detection repeatability and accuracy.

[0032] Step S3, Block Management - Comprehensive Physical Property Detection: S31. Agglomeration Crushing-Sieving Linkage: Feed sample B into the second crusher, crush it, and then sieve it. Collect the sieved sample and record the agglomeration rate. The crushing roller speed of the second crusher is 50-150 r / min, and the roller spacing is 0.5-2 mm. After crushing, it is sieved through a 1 mm aperture vibrating screen. The sieving frequency of the vibrating screen is 50-60 Hz. Agglomeration rate = (mass before sieving - sieved mass) / mass before sieving × 100%.

[0033] S32. Physical property testing: The moisture content of the sieved sample is determined by a microwave moisture meter, and the particle size distribution is detected by a laser particle size analyzer. The agglomeration rate, moisture content, and particle size distribution data are stored together. The laser particle size analyzer has a measurement range of 0.02-2000μm and a repeatability accuracy within 1%. The particle size distribution includes the proportion of particles down to -200 mesh. The agglomeration rate, moisture content, and particle size distribution data are stored together.

[0034] Step S4: Predictive data fusion analysis: S41 Outlier Removal and Standard Comparison: Based on element content, agglomeration rate, moisture content, and particle size data, outliers are removed using the 3σ criterion and compared with preset standard values ​​to determine the sample pass rate. The data processing unit automatically generates a test report based on the comparison results. The report includes the test data and comparison results for element content, agglomeration rate, moisture content, and particle size distribution.

[0035] Outliers are values ​​that deviate from the average data by more than 3 standard deviations; the preset standard values ​​are copper content ≥15%, sulfur content 8-35%, moisture content ≤12%, -200 mesh particle size >80%, total lead and zinc content <5%, and agglomeration rate ≤3%.

[0036] S42. Multi-parameter trend prediction and early warning: Based on historical 6 months of detection data and real-time data, including element content, agglomeration rate, moisture, and particle size data, the LSTM prediction model is called to predict the component qualification rate, smelting furnace temperature trend, and agglomeration rate change in the next 1-2 hours; if the predicted value exceeds the preset standard value, an early warning is triggered, and the data processing unit automatically matches the cause of the deviation.

[0037] The early warning system of this invention can define and set gradient thresholds according to the type of smelting furnace (e.g., 30t / h, 50t / h side-blown furnace), and classify it into three levels: yellow, orange, and red, based on the severity of the deviation of the predicted value from the preset standard. For example: Yellow alert: Predicted values ​​slightly deviate from the threshold (e.g., copper content 14.0%-14.5%, agglomeration rate 4%-5%), indicating a risk of slight fluctuations in the process. Orange alert: Predicted values ​​deviate moderately from the threshold (e.g., copper content 13.5%-14.0%, agglomeration rate 5%-6%), indicating significant process fluctuations, requiring timely adjustments; Red alert: Predicted values ​​deviate significantly from the threshold (e.g., copper content < 13.5%, agglomeration rate > 6%), and abnormal furnace conditions are likely to occur (e.g., sudden drop in furnace temperature, flue gas nodule formation), requiring emergency control.

[0038] Early warning cause analysis: Automatically generated by the system, based on the correlation analysis between the input data (copper / sulfur / iron content, moisture, particle size, etc. in the past hour) and the real-time data of the batching system (various concentrate ratios, feed amount, etc.), the system automatically matches the causes of deviations and outputs standardized analysis results (such as the correlation between low copper content and high copper concentrate ratio in the example).

[0039] S43. Process optimization knowledge base update: The detection data, early warning results (including trigger thresholds and cause analysis), subsequent control measures (such as feed ratio, oxygen enrichment concentration, and feed amount adjustment parameters) and smelting indicators (crude copper grade and slag copper content) are linked and stored to form a complete data chain of "detection-early warning-control-results". The process optimization knowledge base is then updated for recommending control parameters for similar operating conditions in the future.

[0040] Each data chain is tagged with operating conditions based on the furnace model, belt load range, matrix composition type, and key parameter range (such as sulfur content 8-35%, moisture ≤12%) to facilitate rapid matching of similar scenarios in the future.

[0041] Optimization and iteration of control parameters: If the current control measures improve smelting indicators (such as increased copper recovery rate and reduced copper content in slag), then update the "optimal control parameter library" for this operating condition; if the control effect is not good, record the reasons for the deviation (such as the parameter adjustment range being too small) and supplement the improvement plan.

[0042] Dynamic calibration of warning thresholds: By combining historical data and new feedback, the trigger thresholds for the three-level warnings are optimized (e.g., the warning threshold for clumping rate under a certain working condition is adjusted from 5% to 4.5%), thereby improving the accuracy of warnings.

[0043] Case library expanded: Added complete working condition cases, including "initial working condition - detection data - early warning information - control measures - smelting results - experience summary", providing control examples that can be directly referenced for subsequent similar working conditions.

[0044] Step S5: Dual closed-loop feedback control of furnace operating conditions: S51. Basic Parameter Control: Based on the comparison and prediction results, the data processing unit outputs a detection report and control signals for the feeding rate, oxygen enrichment concentration, and batching ratio, which are transmitted to the batching system and the instrumentation and control station of the smelting furnace to adjust the material ratio, feeding rate, and oxygen enrichment concentration in real time; among which, the feeding rate adjustment range is 0-5t / h, and the oxygen enrichment concentration adjustment range is 21-30%; S52. Coordinated Control of Furnace Operating Conditions: Real-time temperature inside the smelting furnace is collected via insertion thermocouples (measurement range 1100-1300℃, accuracy within 5℃), with the thermocouples inserted 0.5-1m into the furnace wall; SO2 concentration data in flue gas is collected via in-situ infrared sensors (accuracy ±50ppm), with flue gas data collected every 5-10 minutes, forming a dual closed loop with the detection data. When the sulfur content is greater than 35% and the SO2 concentration in the flue gas is greater than 2000 ppm, the proportion of high-sulfur concentrate should be reduced, and the oxygen enrichment concentration should be reduced by 1-2% simultaneously. When the agglomeration rate is greater than 3% and the temperature fluctuation inside the furnace is greater than 30℃ / h, increase the crushing roller speed of the crusher and simultaneously reduce the feed rate by 0.5-1t / h to avoid local overheating caused by agglomeration.

[0045] Step S6, Calibration System Upgrade: S61. Pre-monitoring of standard sample status: The laser-induced breakdown spectroscopy module and microwave moisture meter are calibrated every 24 hours. The moisture content (deviation ≤ 0.1%) and main element content (deviation ≤ 0.2%) of the standard sample are quickly detected by the near-infrared spectroscopy module. After confirming that the standard sample is not damp and has no compositional shift, it is then used for the calibration of the laser-induced breakdown spectroscopy module and microwave moisture meter. S62. Cross-device calibration comparison: Select 3 groups of samples each month and compare them with laboratory ICP-MS (inductively coupled plasma mass spectrometry) and Karl Fischer moisture determination methods through steps S1-S5. If the deviation is >3%, retrain and update the matrix composition-detection parameter correction model and the calibration coefficient of the microwave moisture meter. The control logic for dynamic sampling compensation is implemented through a PLC controller; The control logic for dynamic sampling compensation in step S1 is implemented through a PLC controller, specifically as follows: Data acquisition and real-time monitoring: The PLC controller reads the load data from the belt load sensor (0-50t / h, accuracy ±0.2t / h) and the conveying speed data from the speed sensor in real time. The sampling frequency is synchronized with the detection cycle (once every 5-10 minutes) and the real-time data is stored.

[0046] Error judgment triggering condition: The belt scale calibration process is triggered every 8 hours. The PLC compares the calibrated standard load value with the real-time detection value of the load sensor and calculates the load detection error (error = |detection value - standard value| / standard value × 100%). When the error is > ±3%, the sampling parameter correction program is started.

[0047] Automatic formula correction logic: The PLC pre-stores the basic sampling formula (the correlation rule between frequency / sampling quantity and load). During correction, the coefficient is adjusted according to the error ratio: if the detected value is higher than the standard value, the sampling frequency coefficient is increased and the sampling quantity is increased; if the detected value is lower, the corresponding coefficient is decreased to ensure that the sampling quantity matches the actual belt load. The correction response time is ≤1 second.

[0048] Parameter execution and feedback: The corrected sampling frequency (5 / 8 / 10 minute interval) and sampling amount (200 / 350 / 500g / batch) signals are output by the PLC to the belt sampler, and the execution status is fed back in real time to form a closed-loop control.

[0049] The overall detection cycle of this invention is ≤5 minutes. The furnace temperature of the smelting furnace is stabilized at 1150-1250℃ and the melt level is stabilized at 1000-1250mm. The material optimization based on this method can increase the copper recovery rate by 0.5-1.0 percentage points. All detection data, control commands and early warning information are automatically stored and traceable, supporting the process optimization and intelligent upgrading of side-blown smelting production.

[0050] Based on the above system and method, the present invention has been implemented in the following embodiments: This embodiment is based on a side-blown smelting production line with an annual output of 150,000 tons of crude copper. It focuses on online detection and control of the mixed copper concentrate fed into the furnace. The specific implementation steps are as follows: 1. Automatic sampling A belt sampler was installed on the unloading section of the Φ1200mm wide mixed copper concentrate conveyor belt (5m from the furnace inlet). The sampling head width was set to 500mm (1 / 2.4 of the belt width), and the sampling speed was matched with the belt running speed (1.2m / s). The load sensor collected a load of 25t / h, so it was set to automatically intercept a sample every 8 minutes, with a single sample amount of 350g, which was sent to the first crusher via a closed screw conveyor.

[0051] 2. Sample pretreatment The sample first enters the roller crusher (first crusher) (roller diameter 300mm, speed 300r / min) and is crushed to a particle size ≤3mm; then it enters the gravity divider and retains 200g of sample at a 1:7 reduction ratio. After removing iron filings and other impurities by the magnetic iron remover, it is divided into two samples, A and B.

[0052] 3. Element content detection The first sample A's matrix composition was detected in real time using a near-infrared spectroscopy module. The matrix composition-detection parameter correction model was used to dynamically adjust the laser energy of the laser-induced breakdown spectrum to 100 mJ and the spectral integration time to 20 ms. Using the laser-induced breakdown spectroscopy module (wavelength range 200-900 nm, resolution 0.05 nm), an Nd:YAG laser (pulse width 10 ns) was used to perform a 10-point scan on the surface of sample A. The following results were obtained simultaneously: copper content 22.3%, sulfur content 28.5%, iron content 21.7%, lead content 1.2%, zinc content 0.8%, and arsenic content 0.03%. The detection time was 3.5 minutes, the repeatability RSD was 3.2%, and the RSD for arsenic was 2.3%.

[0053] 4. Physical property testing Sample B was fed into a second crusher with a crushing roller speed of 120 r / min and a roller spacing of 1 mm. After crushing, the sample was sieved through a vibrating screen with a 1 mm aperture and a sieving frequency of 60 Hz. The agglomeration rate was calculated. The crushed sample B was divided into two parts. One part was sent to a microwave moisture analyzer (frequency 2450 MHz, power 50 W) and heated for 30 seconds. The moisture content was found to be 8.7%, with an accuracy of ±0.03%. The other part was treated with a dispersant (0.1% sodium hexametaphosphate solution) and then injected into a laser particle size analyzer (measurement range 0.02-2000 μm). The analysis showed that the proportion of -200 mesh particles was 89.3%, and the particle size distribution repeatability was ±0.8%.

[0054] 5. Data fusion analysis The data processing unit receives the data on element content, agglomeration rate, moisture, and particle size, and uses the 3σ criterion to remove one abnormally fluctuating sulfur content data (deviation > 8%). Then, it compares the data with the preset standard values ​​(copper ≥ 15%, sulfur 8-35%, moisture ≤ 12%, -200 mesh > 80%, lead and zinc < 5%) to determine that the batch of samples has a 100% pass rate and generates a test report.

[0055] The LSTM prediction model is invoked to predict the component qualification rate, furnace temperature trend, and agglomeration rate changes over the next two hours; if the predicted values ​​do not exceed the preset standard values, no warning is triggered.

[0056] Recommended adjustment: Maintain the current ingredient ratio, but appropriately reduce the oxygen concentration by 0.5%.

[0057] 6. Feedback Control The control signal is transmitted to the batching system and smelting furnace system, adjusting the oxygen enrichment concentration from 26.5% to 26.0% in real time, while keeping the feeding rate constant at 25t / h; the test report is displayed on the central control screen simultaneously for reference by furnace operators.

[0058] Implementation effect Statistics after 30 days of continuous operation show that the detection cycle is stable at 4.2±0.5 minutes, which is 96.5% shorter than traditional offline detection; the copper recovery rate is increased by 0.7 percentage points; the furnace temperature fluctuation range is reduced from ±50℃ to ±20℃; the flue gas duct nodule cleaning cycle is extended to 60 days (originally 30 days); and the daily energy consumption is reduced by 1200kWh.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An online detection and control system for mixed copper concentrate fed into a side-blown smelting furnace, characterized in that: It includes a batching system, a smelting furnace, and a data processing unit. The batching system and the smelting furnace are connected by a conveyor belt. A belt sampler is installed in the unloading section of the conveyor belt, and a belt load sensor and a speed sensor are installed. A plug-in thermocouple and an in-situ infrared sensor are installed in the smelting furnace. The belt load sensor, speed sensor, plug-in thermocouple, and in-situ infrared sensor are all connected to the data processing unit. It also includes a near-infrared spectroscopy submodule and a laser-induced breakdown spectroscopy module. The sampler is connected to a first crusher, which is connected to a divider. The divider is divided into two paths: one path is connected to the near-infrared spectroscopy submodule and the laser-induced breakdown spectroscopy module, and the other path is connected to a second crusher. The second crusher is connected to a microwave moisture meter and a laser particle size analyzer. The laser-induced breakdown spectroscopy module, the microwave moisture meter, the laser particle size analyzer, and the second crusher are all connected to a data processing unit. The data processing unit is also connected to the sampler and the batching system.

2. The online detection and control system for mixed copper concentrate fed into a side-blown smelting furnace according to claim 1, characterized in that: The data processing unit has a built-in LSTM prediction model and a matrix composition-detection parameter correction model. It inputs the data from the near-infrared spectroscopy submodule into the matrix composition-detection parameter correction model and uses the output data as the parameters of the laser-induced breakdown spectroscopy module.

3. A method for online detection and control of mixed copper concentrate fed into a side-blown smelting furnace according to any one of claims 1-2, characterized in that: Includes the following steps: Step S1, Dynamic Compensation Automatic Sampling: The belt load and conveying speed are collected in real time by the belt load sensor and speed sensor. The sampling frequency and sampling amount are determined according to the belt load, and the sample is taken by the sampler. Step S2, Matrix Correction - Elemental Co-detection: S21. Sample pretreatment: The sample taken by the sampler is crushed by the first crusher, and then the crushed sample is reduced to 100-200g by the divider. After removing impurities, it is divided into two parts, A and B. Part A is used for elemental detection and part B is used for physical property detection. S22, Adaptive correction of matrix effect: The matrix composition of sample A is detected in real time by the near-infrared spectroscopy module. Based on the matrix composition of sample A, the matrix composition-detection parameter correction model is called to dynamically adjust the laser energy and spectral integration time of the laser-induced breakdown spectrum. After adjustment, sample A is scanned by the laser-induced breakdown spectroscopy module to detect the content of copper, sulfur, iron main elements and lead, zinc and arsenic impurity elements in sample A. Step S3, Block Management - Comprehensive Physical Property Detection: S31, Agglomeration Crushing-Screening Linkage: Send sample B into the second crusher, crush it, sieve it, collect the sieved sample, and record the agglomeration rate; S32. Physical property testing: Take the sieved sample and determine the moisture content using a microwave moisture meter, and at the same time use a laser particle size analyzer to detect the particle size distribution; store the data of agglomeration rate, moisture content and particle size distribution together. Step S4: Predictive data fusion analysis: S41 Outlier Removal and Standard Comparison: Based on element content, agglomeration rate, moisture content, and particle size data, outliers are removed using the 3σ criterion and compared with preset standard values ​​to determine the sample pass rate. The data processing unit automatically generates a test report based on the comparison results. The report includes the test data and comparison results for element content, agglomeration rate, moisture content, and particle size distribution. S42. Multi-parameter trend prediction and early warning: Based on the historical 6 months of detection data and real-time data, including element content, agglomeration rate, moisture and particle size data, the LSTM prediction model is called to predict the component qualification rate, smelting furnace temperature trend and agglomeration rate change in the next 1-2 hours. If the predicted value exceeds the preset standard value, an early warning is triggered, and the data processing unit automatically matches the cause of the deviation. S43. Update the process optimization knowledge base: Link and store the test data, early warning results, subsequent control measures and smelting indicators, update the process optimization knowledge base, and use it to recommend control parameters for similar working conditions in the future. Step S5: Dual closed-loop feedback control of furnace operating conditions: S51. Basic Parameter Control: Based on the comparison and prediction results, the data processing unit outputs a detection report and control signals for the feeding rate, oxygen enrichment concentration, and batching ratio, which are transmitted to the batching system and the instrumentation and control station of the smelting furnace to adjust the material ratio, feeding rate, and oxygen enrichment concentration in real time; among which, the feeding rate adjustment range is 0-5t / h, and the oxygen enrichment concentration adjustment range is 21-30%; S52. Coordinated Control of Furnace Operating Conditions: Real-time temperature inside the smelting furnace is collected via insertion thermocouples, and SO2 concentration data in flue gas is collected via in-situ infrared sensors, forming a dual closed loop with the detection data. When the sulfur content is greater than 35% and the SO2 concentration in the flue gas is greater than 2000 ppm, the proportion of high-sulfur concentrate should be reduced, and the oxygen enrichment concentration should be reduced by 1-2% simultaneously. When the agglomeration rate is greater than 3% and the temperature fluctuation inside the furnace is greater than 30℃ / h, increase the crushing roller speed of the crusher and simultaneously reduce the feed rate by 0.5-1t / h.

4. The online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace according to claim 3, characterized in that: In step S1, the unit of belt load is t / h, and the unit of conveyor speed is m / s. The sampling frequency and sampling amount are determined according to the belt load as follows: when the load is ≤10t / h, samples are taken at 10-minute intervals, with a sampling amount of 200g / batch; when 10t / h < load ≤20t / h, samples are taken at 8-minute intervals, with a sampling amount of 350g / batch; when the load is >20t / h, samples are taken at 5-minute intervals, with a sampling amount of 500g / batch. The belt load-sampling parameter correspondence is recorded synchronously during the sampling process.

5. The online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace according to claim 3, characterized in that: In step S1, dynamic sampling compensation calibration is performed during the sampling process. Specifically, the belt load sensor data is calibrated every 8 hours using a belt scale. When the load detection error is not within 3%, the sampling frequency and sampling amount are automatically corrected to ensure that the sampling representativeness matches the dynamic load of the belt.

6. The online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace according to claim 3, characterized in that: In step S21, the particle size of the crushed sample is ≤3mm; the sample is reduced by a gravity-type divider with an adjustable reduction ratio of 1:5-1:

10.

7. The online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace according to claim 3, characterized in that: In step S22, the matrix composition is the content of silicon, aluminum and calcium. The laser energy is dynamically adjusted to 50-100mJ and the spectral integration time is 10-30ms through the matrix composition-detection parameter correction model. The detection time by laser-induced breakdown spectroscopy is ≤5 minutes. Among them, the detection limit for arsenic is ≤0.05%, the repeatability RSD is ≤2.5%, and the repeatability RSD for other elements is ≤3%.

8. The online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace according to claim 7, characterized in that: In step S22, the construction process of the matrix composition-detection parameter correction model is as follows: S221. Prepare matrix standard samples and construct a diverse standard sample library. Matrix composition gradient design: Based on the content gradient of silicon (0.1%, 0.5%, 1%, 3%, or 5%), aluminum (0.05%, 0.1%, 0.5%, 2%, or 3%), and calcium (0.02%, 0.05%, 0.1%, 1%, or 2%), samples with more than m different matrix ratios are formed by combining two or three substances in different proportions, where m > 500. Sample homogeneity and stability treatment: After each group of samples is crushed and mixed, the homogeneity is verified by sampling-testing-comparison to ensure that the homogeneity error is within 1%. The samples are sealed and stored to avoid moisture and component shift, ensuring the stability of sample characteristics. Target element content fixation: In each set of standard samples, the content of main elements such as copper, sulfur, and iron, as well as impurity elements such as lead, zinc, and arsenic, is fixed according to industrial standards. Only the proportion of matrix components is changed to avoid the target element content from interfering with model modeling. S222. Multi-parameter combination calibration to obtain matrix-parameter-accuracy correlation data. For each of the more than m groups of standard samples in step S221, the detection parameters are calibrated to establish the correspondence between the matrix composition and the optimal detection parameters. Specifically: Detection parameter variable settings: Laser energy of 50mJ, 60mJ, 70mJ, 80mJ, 90mJ, and 100mJ, and spectral integration time of 10ms, 15ms, 20ms, 25ms, and 30ms are selected as core variables, forming 6×5=30 parameter combinations; Full parameter combination detection: For each group of standard samples, laser-induced breakdown spectroscopy detection was performed using 30 parameter combinations. The detection results of the target elements under each parameter combination were recorded, including the detection value of each element, repeatability RSD, and detection time. Optimal parameter screening criteria: For each sample group, the parameter combination with the smallest detection repeatability RSD and detection time ≤ 5 minutes is selected as the optimal detection parameters for that matrix component; Data recording and organization: The final result is a three-dimensional associated dataset of m groups of matrix components, optimal detection parameters, and detection accuracy, which serves as the basic data for model training; where the matrix components are the contents of silicon, aluminum, and calcium; the optimal detection parameters are the corresponding laser energy and integration time; and the detection accuracy is the detection repeatability RSD. S223. Model building, training and optimization, establishing nonlinear mapping relationships. Based on the 3D correlation dataset from step S222, a machine learning algorithm is used to construct a model to achieve a nonlinear mapping from matrix components to optimal detection parameters, specifically: Algorithm selection: First, a linear relationship between matrix components and detection parameters is initially established through multiple linear regression, and then nonlinear errors are corrected through gradient boosting trees; thus forming a model that can achieve nonlinear mapping between matrix components and optimal detection parameters. Model input and output definitions: Input features: silicon content, aluminum content, and calcium content in the associated dataset from step S222; Output labels: the optimal laser energy and optimal spectral integration time in the associated dataset of step S222; Training and validation split: The m sets of 3D association datasets are split in a 7:3 ratio, with 70% used as the training set for model fitting and 30% used as the validation set for accuracy verification; the training set is first input into the model formed by the combination algorithm of multiple linear regression and gradient boosting tree to initially fit the mapping relationship between matrix components and detection parameters; Model optimization objectives: The goal is to achieve a deviation of ≤5% between predicted parameters and actual optimal parameters in the validation set, an RSD of ≤2.5% for arsenic detection repeatability, and an RSD of ≤3% for other elements. The model is optimized by adjusting the algorithm hyperparameters to ensure that the model's parameter prediction accuracy on the validation set is ≥95%. The algorithm hyperparameters adjustment includes adjusting the tree depth and learning rate of the gradient boosting tree, and the regularization coefficient of the multiple linear regression. The model error is iteratively corrected until the model meets the detection accuracy requirements, thus completing the training optimization.

9. The online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace according to claim 3, characterized in that: In step S31, the crushing roller speed of the second crusher is 50-150 r / min. After crushing, the material is screened through a vibrating screen with a 1 mm aperture. The agglomeration rate is calculated as (mass before screening - mass after screening) / mass before screening × 100%. In step S32, the laser particle size analyzer has a measurement range of 0.02-2000μm, and the repeatability accuracy of the measurement is within 1%. The particle size distribution is detected, including the proportion of particles of -200 mesh. The agglomeration rate, moisture content, and particle size distribution data are stored together.

10. The online detection and control system and method for mixed copper concentrate fed into a side-blown smelting furnace according to claim 1, characterized in that: In step S41, outliers are values ​​that deviate from the average data by more than 3 standard deviations; the preset standard values ​​are copper content ≥15%, sulfur content 8-35%, moisture content ≤12%, -200 mesh particle size >80%, total lead and zinc content <5%, and agglomeration rate ≤3%. In step S43, the smelting indicators include the grade of crude copper and the copper content in the slag.

Images