One-key grinding classification control method based on LightGBM

Abstract

The present application relates to the technical field of ore dressing plant control, and particularly relates to a one-key type grinding and grading control method based on LightGBM, production data is obtained by preprocessing collected iron ore grinding production data through industrial internet of things, a data-driven dynamic prediction model is constructed by using a LightGBM algorithm, time-space lag cross-correlation and flattened data input are combined, the trained light-weight model is deployed to a PLC control layer, and corresponding key process parameter optimization adjustment is completed. The present application has the advantages that: by collecting production data in real time and constructing a dynamic prediction model by using a LightGBM algorithm, key variables such as mill speed and ore supply can be accurately and dynamically adjusted, the grinding process is optimized, mill no-load and overload operation is reduced, compared with a traditional control method, the grinding efficiency is significantly improved; an operator only needs to set target parameters, key parameters such as mill speed and ore supply are optimized in real time through a LightGBM model, the grinding fineness is ensured to be stable and up to standard, and frequent intervention and adjustment are not needed.

Description

Technical Field

[0001] This invention relates to the field of mineral processing plant control technology, and in particular to a one-click grinding and classification control method based on LightGBM. Background Technology

[0002] Grinding systems that rely heavily on manual experience to adjust equipment parameters have the following problems: 1) Delayed operational response: Manual adjustments rely on experience and judgment, making it difficult to respond promptly to fluctuations in ore properties, resulting in low control precision and reduced production efficiency; 2) Inefficient process control: Traditional PID control is difficult to adapt to multi-variable, strongly coupled, and nonlinear grinding processes, and has limited optimization capabilities; 3) High energy and material consumption: Due to the lack of intelligent control methods, the grinding process is often in a suboptimal operating state, resulting in the waste of energy and media (such as steel balls).

[0003] Although next-generation information technologies such as deep learning and digital twins have begun to integrate with mineral processing, existing automated control systems still face three major technical bottlenecks: 1) Insufficient real-time sensing of ore characteristics: It is difficult to dynamically monitor key parameters such as ore hardness and particle size, resulting in poor adaptability of control strategies; 2) Low accuracy of multivariable coupled modeling: Traditional control models (such as PID and fuzzy control) have limited predictive ability for complex operating conditions, and parameter optimization has a lag. 3) Complex human-computer interaction: Operators still need to intervene frequently, the system is not intelligent enough, and it is difficult to achieve "one-click" optimization and control. Summary of the Invention

[0004] The purpose of this invention is to provide a one-click grinding and classification control method based on LightGBM. By deeply integrating the LightGBM algorithm with the industrial control system, it solves the pain points of low efficiency, high energy consumption and reliance on manual labor in traditional grinding processes, and realizes a three-in-one technological innovation of precise control, energy saving and consumption reduction and intelligent operation and maintenance, providing an efficient solution for the green and intelligent transformation of the mineral processing industry.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A one-click grinding and classification control method based on LightGBM specifically includes: S1. Collect production data and preprocess the data; S2, Grinding and Classification Control, is used to control the feed rate, mill water supply, grinding concentration, and hydrocyclone classification process during the grinding and classification process; S3. Lightweight Gradient Boosting Machine (LightGBM) Modeling Based on TLCC Flattened Data Input; S4, the system deployment and real-time control module, is used to deploy the trained prediction model (the prediction model is the one-click grinding and grading control method based on LightGBM) to the PLC control layer to complete millisecond-level real-time decision-making.

[0006] In S1, production data from the grinding and classification process is collected in real time via the Industrial Internet of Things (IIoT). Python is used to load and preprocess the production data, specifically including: S11, Missing value imputation: Missing values ​​in production data are filled using linear interpolation. S12, Outlier Replacement: Outliers in production data are replaced using mean filtering. S13, Outlier Replacement: Outliers are identified by calculating the mean and standard deviation of process parameters and then replaced with the mean of the data sample, as shown in the following formula: ① In formula ①, x i This represents the i-th data point. This represents the average value of the data, and n represents the number of data points. S14. Divide the production data into training and testing sets: After data preprocessing, the production data is divided into training set D1 and test set T1 in a 7:3 ratio. Where D1 ∈ R l×m ; The size of the training set D1 is: R l×m , l represents the number of samples in the training set, m represents the number of feature variables, and R represents the set of real numbers.

[0007] Production data includes the ball mill's rotational speed, current, load, grinding time, mill decibels, classifier overflow concentration, hydrocyclone inlet pressure, return sand ratio, ore density and hardness index, steel ball filling rate and proportion, slurry pH value, actual measurement data from the particle size analyzer, and ore feed particle size.

[0008] In S2, grinding and classification control includes: S21. Ore feed rate control: By comparing the actual feed rate signal of the belt scale with the set feed rate, the deviation signal is obtained and transmitted to the frequency converter, which then drives the feeder to adjust the feed rate. S22. Mill water supply control includes the control of the proportioned water supply during ore feeding and the control of the supplementary water supply during ball mill discharge: Water ratio control during ore feeding: Water ratio control during ore feeding is to precisely adjust the water flow rate during the feeding process based on factors such as ore feed rate, moisture content and particle size, to ensure the optimal mixing ratio of water and ore. Control of water replenishment during ball mill discharge: The actual flow rate on the flushing water pipe at the ball mill discharge end is compared with the set flow rate, and the deviation signal is transmitted to the regulating valve to complete the control of water replenishment. S23. Grinding concentration control: With a fixed feed rate, the mill water flow rate is adjusted to maintain the grinding concentration within the set range. S24. Hydrocyclone classification process control: The feed to the hydrocyclone is adjusted by regulating the speed of the switching pump to regulate the inlet pressure of the hydrocyclone, so that the inlet pressure of the hydrocyclone is within the set value range.

[0009] In S3, a lightweight gradient boosting machine (LightGBM) is used for modeling based on TLCC flattened data input, specifically including: The time lag cross-correlation (TLCC) method is used to analyze the time lag difference between production data. A dynamic prediction model is constructed using the LightGBM algorithm, and the model training is optimized through a leaf-based growth strategy, one-sided gradient sampling, and parallel learning.

[0010] The Time Lag Cross-Correlation (TLCC) method calculates the time difference between two time series using the Pearson correlation coefficient, specifically including: 1) Slide the two time series one by one, and calculate the Pearson correlation coefficient each time. The calculation formula is as follows; ① In formula ①, The Pearson correlation coefficient represents the strength and direction of the linear relationship between variables x and y. Its value ranges from 1 to 1, where +1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear relationship. This represents the data for the i-th observation of variable x; variable This represents the data for the i-th observation of variable y; variable The average value of variable x; variable y represents the average value of the variable y; n represents the number of data points; Formula ① is used to determine which of two time series is leading. This method involves sliding the two series sequentially, calculating the Pearson correlation coefficient (r) between them at each slide. xy When the correlation coefficient between two sequences reaches its maximum under a certain time lag, this time lag difference is considered to be the true time lag difference between the two process parameters. 2) When the Pearson correlation coefficient reaches its maximum value, the corresponding lag time difference is the true lag time difference of the production data.

[0011] The leaf growth strategy prioritizes dividing the leaves that have the greatest division gain. One-sided gradient sampling is used to optimize production data, including: sorting by absolute gradient value, selecting the data with the largest absolute gradient value (a100%) and randomly selecting the data with the smallest absolute gradient value (b100%); multiplying the data with the smallest absolute gradient value by a constant (1-a) / b and merging it with the data with the largest absolute gradient value to calculate the gain; Parallel learning is achieved through feature parallelism, data parallelism, and voting parallelism strategies: Features are used in parallel to store all data on different machines, and the segmentation strategy is executed locally. Data parallelism uses a distributed reduction method to merge histograms; The voting process is parallelized by merging histograms through filtering TopK features and voting strategies.

[0012] Flattening the data input is used to convert time series data into window-based input, specifically including: The transformation function retains only the data from the most recent time point for the auxiliary variables within the time series window; The transformed one-dimensional data is input into the prediction model for training and prediction, resulting in a single target. The loss is as follows: ② In formula ②, R w*m : Represents a time series window, a matrix or vector with w*m elements; w* represents the first dimension of the time series window size; M: represents the second dimension of the time series window; rw*m-1 represents the transformed time series window, with w*m-1 elements; ③ In formula ③, Y represents the value of the loss function, used to measure the difference between the model's predictions and the true values; N represents the total number of samples, the number of samples required for the model to train; t' represents the time step for each sample; Y i，j : Represents the true value (label), the true target value at time point j for the i-th sample; This represents the predicted value, which is calculated by the native pointwise prediction model.

[0013] Formula 3 represents the mean squared error between the model's predicted value and the true value; Formula 2 transforms the original window of size w*m into a window of size w*m-1. During the transformation, the labels remain unchanged, while the dimensionality of the auxiliary variables is reduced, retaining only the data from the most recent time point in each window; Unlike conventional time series data input formats, the flattened data input method reconfigures the time series into a window-based input and trains on multiple features (window data), thereby achieving good predictive performance.

[0014] The system deployment and real-time control module includes a model deployment unit and a real-time decision-making unit. The model deployment unit is used to deploy the trained prediction model to the PLC control layer. The real-time decision-making unit is used to make real-time decisions based on the real-time collected production data through the PLC control system, and the prediction model makes real-time decisions within a millisecond response time.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention collects production data in real time and uses the LightGBM algorithm to build a dynamic prediction model, which can accurately and dynamically adjust key variables such as mill speed and feed rate, optimize the grinding process, reduce mill idling and overload operation, and significantly improve grinding efficiency compared with traditional control methods. 2. The one-click intelligent control mode can automatically optimize key variables in the grinding and classification process, reduce energy consumption in the grinding process, reduce total energy consumption in mineral processing, effectively improve energy utilization, and reduce production costs. 3. The LightGBM algorithm has efficient histogram and other optimization algorithms, which can reduce the calculation latency to the millisecond level, meet the real-time requirements of industrial sites, realize precise control of the grinding and classification process, and improve the grinding fineness qualification rate. 4. The efficient computing power of the LightGBM algorithm in this invention, combined with the edge computing architecture, enables the rapid issuance of control commands (delay <50ms), which far exceeds the response speed of traditional PID control. It has strong robustness to interference factors such as ore hardness fluctuations and media wear, and can stably perform grinding control, ensuring the continuity and stability of the grinding process and reducing production fluctuations caused by interference factors. 5. Operators only need to set target parameters and optimize key parameters such as mill speed and feed rate in real time through the LightGBM model to ensure that the grinding fineness is stable and meets the standard without frequent intervention and adjustment; combined with time lag cross-correlation (TLCC) analysis, the control strategy is dynamically adjusted to effectively cope with changes in ore hardness, particle size, etc., reduce the dependence on operator experience, and reduce the impact of human operation errors on production. 6. By integrating machine vision perception and industrial Internet of Things technology, it is possible to perceive ore characteristics in real time, dynamically adapt to fluctuations in ore properties, realize adaptive control of complex working conditions, and improve the flexibility and adaptability of production. 7. This invention realizes intelligent regulation of the grinding and classification process. Through real-time monitoring and dynamic adjustment, it maintains stable operation of the grinding process, reduces production interruptions caused by equipment failure or process fluctuations, and improves the continuity and stability of production. 8. The intelligent decision-making engine based on the LightGBM algorithm uses a large amount of real-time production data for training and prediction, providing scientific and accurate decision-making basis for the production process, and improving the reliability and quality stability of production. Attached Figure Description

[0016] Figure 1 One-click grinding and classification control process based on LightGBM Figure 1 .

[0017] Figure 2 One-click grinding and classification control process based on LightGBM Figure 2 .

[0018] Figure 3 This is a particle size distribution curve from a grinding and classification test. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings, but it should be noted that the implementation of the present invention is not limited to the following embodiments.

[0020] The following embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments. Unless otherwise specified, the methods used in the following embodiments are conventional methods.

[0021] Example 1

[0022] This invention is based on the principles of iron ore grinding and beneficiation processes, utilizing time gradients in machine learning, industrial big data, and the principles of iron ore grinding and beneficiation. It collects production data in real time through the Industrial Internet of Things (IIoT), constructs a data-driven dynamic prediction model using the LightGBM algorithm, and deploys the trained prediction model to the PLC control layer by combining time-lag cross-correlation and flattened data input. Millisecond-level real-time decision-making is achieved through the OPC protocol, automatically optimizing key variables such as mill speed, feed rate, and classifier overflow concentration. Compared to traditional PID control, this invention significantly improves adaptability to nonlinear conditions, enabling one-click control of the entire process. It exhibits strong robustness to ore hardness fluctuations and equipment wear, providing an efficient solution for intelligent mineral processing. A one-click grinding and classification control method based on LightGBM is also available. Figure 1 Specifically, it includes: S1. Data Acquisition and Preprocessing of Iron Ore Grinding and Processing Production Data: Production data collection and data preprocessing are shown in Tables 1 and 2.

[0023] Table 1 shows the process parameters.

[0024]

[0025] Table 2 shows the basic characteristics of the process parameters.

[0026]

[0027] By collecting real-time production data from the iron ore grinding and beneficiation process via the Industrial Internet of Things (IIoT) (including ball mill speed, current, load, grinding time, mill decibels, classifier overflow concentration, hydrocyclone inlet pressure, return sand ratio, ore density and hardness index, steel ball filling rate and proportion, slurry pH value, actual measurement data from particle size analyzer, ore feed particle size, mill vibration, and media filling rate), a production dataset D for the beneficiation plant is obtained. Python is used to load the production data D and perform data preprocessing, specifically including: S11, Missing value imputation: For missing values ​​in production data, linear interpolation is used to fill them in, for example: The data column is missing one data point: 249.58; 251.41; NaN; 252.83; 253.

[0028] Using the known data from the second and fourth points, see the formula: ; y3 represents the calculated NaN value; y2 and y4 represent the known data of the previous and next data points; x2, x3, and x4 represent the time points of the three data points, and y3 is calculated to be 252.12.

[0029] S12, Outlier Replacement: Outliers in production data are replaced using mean filtering, for example: Check and replace outliers with the average value within a certain range. For example, the feed rate per minute is 249.58; 251.41; 459; 252.83; 253. The third data point has a standard deviation greater than 2 and does not reflect the actual situation on site; therefore, the average value is used instead.

[0030] S13, Outlier Replacement: Outliers are identified by calculating the mean and standard deviation of the production data, and then replaced with the mean of the production data sample. For example: xi This represents the i-th data point. This represents the average value of the data, where n represents the number of data points. Outliers are checked and replaced with the average value within a certain range. For example, the ore feed per minute is 249.58; 251.41; 339; 252.83; 253. The standard deviation of the third data point is greater than 1.5 times, so the average value is used instead.

[0031] S14. Divide the production data into training and testing sets: After data preprocessing, the production dataset D is divided into training set D1 and test set T1 in a 7:3 ratio. Wherein, the training set D1 ∈ R l×m ; The size of the training set D1 is given by the following formula: R l×m ① In formula ①, l represents the number of samples in the training set, m represents the number of feature variables, and R represents the set of real numbers; The test set T1 is used to validate the performance of the prediction model.

[0032] S2. Grinding and Classification Control: This control is used to regulate the feed rate, mill water supply, grinding concentration, and hydrocyclone classification process during grinding and classification. The details are as follows: S21. Ore feed rate control: By comparing the actual feed rate signal of the belt scale with the set feed rate, the deviation signal is obtained and transmitted to the frequency converter, which then drives the feeder to adjust the feed rate. S22. Mill water supply control: Mill water supply control includes controlling the proportion of water added during ore feeding and controlling the amount of water added during ball mill discharge. Water ratio control during ore feeding: Water ratio control during ore feeding is to precisely adjust the water flow rate during the feeding process based on factors such as ore feed rate, moisture content and particle size, to ensure the optimal mixing ratio of water and ore. Control of water replenishment during ball mill discharge: The actual flow rate (detected by an electromagnetic flow meter) on the flushing water pipe at the ball mill discharge end is compared with the set flow rate. The deviation signal is then transmitted to the start regulating valve to control the amount of water replenishment. S23. Grinding concentration control: Slurry concentration is the proportion of the dry weight of the minerals in the mill to the total weight of the slurry. It is related to the feed rate, the proportion of water used, variations in the slurry volume, the friction between the grinding media and the ore, and the mill's transport performance. Depending on the required particle size distribution, the grinding concentration needs to be controlled within a certain range. Grinding concentration cannot be directly measured; with a fixed feed rate, adjusting the mill's water supply helps maintain the grinding concentration within the set range. S24. Hydrocyclone classification process control: The hydrocyclone feeds minerals from the pump tank to the inlet by adjusting the feed pressure. The feed pressure is measured by a pressure transmitter in the hydrocyclone inlet pipe, which adjusts the speed of the converter pump to control the inlet pressure of the hydrocyclone. The information sent from the hydrocyclone inlet pressure sensor has a given pressure value and error. The controller is used to control and adjust the error, and finally the control signal is sent to the actuator of the frequency converter and operated to determine the mineral concentration using a concentration meter.

[0033] S3. Lightweight Gradient Boosting Machine (LightGBM) modeling based on TLCC flattened data input, specifically including: S31. Use the Time Lag Cross-Correlation (TLCC) method to analyze the time lag difference between production data; The Time Lag Cross-Correlation (TLCC) method calculates the time difference between two time series using the Pearson correlation coefficient, specifically including: 1) Slide the two time series one by one, and calculate the Pearson correlation coefficient each time. The calculation formula is as follows; ① In formula ①, The Pearson correlation coefficient represents the strength and direction of the linear relationship between variables x and y. Its value ranges from 1 to 1, where +1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear relationship. This represents the data for the i-th observation of variable x; variable This represents the data for the i-th observation of variable y; variable The average value of variable x; variable y represents the average value of the variable y; n represents the number of data points; Formula ① is used to determine which of two time series is leading. This method involves sliding the two series sequentially, calculating the Pearson correlation coefficient (r) between them at each slide. xyWhen the correlation coefficient between two sequences reaches its maximum under a certain time lag, this time lag difference is considered to be the true time lag difference between the two process parameters. 2) When the Pearson correlation coefficient reaches its maximum value, the corresponding lag time difference is the true lag time difference of the production data.

[0034] S32. A dynamic prediction model is constructed using the Lightweight Gradient Boosting Machine (LightGBM) algorithm. This model is capable of real-time updates and adjustments as time or environment changes. It not only relies on static data for prediction but also handles changes in data flow and makes corresponding adjustments or retraining when new data enters, in order to maintain the accuracy and real-time performance of the prediction. The Lightweight Gradient Boosting Machine (LightGBM) connects many weak learners in series. Each weak learner is only responsible for the accumulated loss of the previous weak learner. Then, the accumulated loss of the superimposed model is made to decrease in the direction of the negative gradient. Finally, all the weak learners are linearly superimposed (with different weights) to obtain a strong learner with excellent performance and generalization ability.

[0035] Gradient Boosting Trees (GBDT) is a gradient boosting algorithm that utilizes ensemble models, employing decision trees as weak learners. Decision trees are unstable learners, significantly affected by fluctuations in training data; however, statistically, using them as weak learners makes sense. Unstable learners typically produce large errors, but statistical characteristics of experimental results from ensemble models show that the large errors of weak learners help enhance the accuracy and generalization ability of strong learners (models). Therefore, using decision trees as weak learners achieves better generalization ability than using more stable methods like linear regression.

[0036] A dynamic prediction model is constructed using the Lightweight Gradient Boosting Machine (LightGBM) algorithm. The model training is optimized through a leaf growth strategy, one-sided gradient sampling, and parallel learning. The details are as follows: 1) Leaf growth strategy: Find the leaf with the largest splitting gain among all the current layers, and then split only on this leaf. Continue this process at a certain depth (to avoid overfitting). With the same number of splits, better accuracy can be obtained while reducing the amount of computation. 2) One-sided gradient sampling To avoid directly discarding all data with small gradients, which would severely impact the overall data distribution, we first sort all values ​​of the splitting feature from largest to smallest absolute value. Then, we select the a*100% of data with the largest absolute gradient value and randomly select the b*100% of data with the smallest absolute gradient value. Next, we multiply these smaller b*100% of data by (1-a) / b (a constant). Finally, we use the selected (a+b)*100% of data to calculate the gain. In this way, data that originally had almost negligible impact on the model will receive more attention from the model without significantly affecting the overall data distribution, thus effectively reducing the amount of training data while maintaining the accuracy of the algorithm. 3) Parallel learning is achieved through feature parallelism, data parallelism, and voting parallelism strategies: a) Feature Parallelism All data is stored on different datasets, and the segmentation strategy can be executed locally after obtaining the optimal split point to reduce communication costs between different machines. See [link / reference needed]. Figure 3 ; b. Data parallelism By using a distributed reduction method, the task of merging histograms is completed on different machines, which reduces communication overhead. At the same time, the difference processing of histograms is performed, which further reduces communication overhead. e. Parallel voting When faced with large datasets, this strategy only performs histogram merging of some features: 1) Find the top K most influential features on the local machine (TopK features), and then use a voting strategy to select the possible optimal split point features; 2) When merging histograms, only merge the selected features; S33. Flattened data input is used to convert time series data into window-based input and train on multiple features (window data), specifically including: a. Use conversion function ② to convert the time series window of A*B into window data of (A+B-1)*1; this conversion function keeps the labels in the time series window unchanged, while the auxiliary variables in the window only save the most recent time point; The transformed one-dimensional data is input into the model method for training and prediction, resulting in a single target. The loss is as follows: ② b. Input the one-dimensional data into the prediction model (this model) It is a simple point-by-point prediction model that predicts a single target. Its loss is as follows: ③ Model Validation: The test set T1 is an independently selected dataset from the overall dataset, typically accounting for 20%-30% of the total dataset. There is no overlap between the test set T1 and the training set during model training. The test set is then fed into the one-click grinding model of this invention, and the mean squared error (MSE) is calculated. After several calculations, the resulting MSE is consistently less than 1.

[0037] ④ Where n represents the total number of data points; yi represents the true value of the i-th data point; This represents the predicted value for the i-th data point.

[0038] S4, System Deployment and Real-Time Control Module, is used to deploy the trained prediction model to the PLC control layer to complete millisecond-level real-time decision-making. Its contents are as follows: 1) Model Deployment Unit The trained pre-model is deployed to the PLC control layer and uses the OPC protocol to interact with the field control system in real time. The model provides real-time decision support and automatically adjusts key control variables such as mill speed, feed rate, water addition, and classifier overflow concentration.

[0039] 2) Real-time decision-making unit Through the PLC control system, based on real-time acquired data, the model method makes real-time decisions within a millisecond response time; the PLC control system automatically adjusts parameters according to feedback information to ensure that the entire grinding and classification process operates under optimal conditions.

[0040] The method proposed in this invention was used to calculate the field observation data of a one-click grinding and classification loop. The results showed convergence. The calculated predicted values ​​basically matched the actual field values, indicating that within the allowable error range, the process observation data of the grinding and classification loop is reliable, balanced, and representative. This demonstrates the practicality of this algorithm in grinding and classification control systems.

[0041] See Figure 3 The particle size distribution diagram shows the particle size distribution of the products from different equipment in the mineral processing or grinding process, and compares the experimental data (“Experiment”) and equilibrium data (“Equilibrium”).

[0042] Example 2

[0043] A one-click grinding and classification control method based on LightGBM is the same as in Example 1, except that a one-click grinding and classification control process based on LightGBM is added to Example 1.

[0044] A one-click grinding and classification control process based on LightGBM, comprising a multivariable control loop between the ball mill and the classifier, see [link to relevant documentation]. Figure 2 The system optimizes the grinding process through multi-parameter coordinated control, specifically including: 1) A feed setting module monitors the current feed rate, mill accelerator signal, and power data in real time, dynamically adjusting the ball mill load to ensure stable operation; 2) An inlet pressure control module adjusts the pressure of the hydrocyclone and classifier in a coordinated manner to ensure material classification accuracy; 3) A makeup water system intelligently adjusts the water volume and controls the slurry concentration based on the outlet of the mill and rotary mill; 4) An operating pressure module implements closed-loop control of the hydrocyclone inlet pressure, forming a dual guarantee in conjunction with the classifier parameters. The system constructs adaptive control logic through sensor data feedback, optimizing energy consumption while ensuring product particle size meets standards. The correlation analysis between the accelerator signal and power can prevent the risk of material blockage.

[0045] In the LightGBM-based one-click grinding and classification control method, process parameters (such as feed rate, swirl feed pressure, and water feed rate) are key input variables. By monitoring these process parameters in real time, dynamic prediction and control of the grinding process can be achieved. The following is an example of how to substitute specific data into a LightGBM-based model and describe the control process: Project Background Suppose that an automatic control of an ore grinding process is in progress, and the goal is to optimize the grinding and classification process by adjusting input parameters (such as feed rate, water rate, etc.) in real time, so that the output particle size (overflow particle size) is kept within a suitable range, thereby improving ore processing efficiency and ensuring stable operation of the equipment.

[0046] Process parameter description: Gkl (Ore Feed): refers to the amount of ore input.

[0047] Xg_yl (feeding pressure): The pressure in the feeding system used to control the feeding speed.

[0048] Gsl (Water Supply): Water supply is used to adjust the concentration of the slurry.

[0049] my (grinding sound): The sound of a mill, used to indirectly reflect the load condition of the mill.

[0050] Qmj_dl (Ball Mill Current): The current of the ball mill, reflecting the load condition of the mill.

[0051] Xg_ll (Spinning flow): The flow rate supplied to the system.

[0052] Gk_nd (feed concentration): The concentration of the feed material.

[0053] Bc_yw (Pump Pool Level): The level of the pump pool affects the stability of the pumping system.

[0054] Bc_jsl (Pump Pool Water Addition): The amount of water added to the pump pool directly affects the overflow particle size.

[0055] Bc_nd (Pump Pool Concentration): The concentration of the material in the pump pool.

[0056] LightGBM model construction and control process.

[0057] 1. Data Acquisition and Preprocessing First, real-time data acquisition of these process parameters is required. Each process parameter (such as feed rate, water feed rate, etc.) will be input as a feature into the LightGBM model as training data. Simultaneously, historical operational data will be used to train the model to ensure it learns the patterns within the process.

[0058] Example data (assuming each row represents a single point in time):

[0059] 2. Model Training Based on this data, the LightGBM algorithm was used for training. The goal of the model is to predict the optimal process parameters using the input process parameters (such as feed rate, vortex feed pressure, water feed rate, etc.).

[0060] Control strategy: Feedback debugging is carried out based on the test results of the final overflow particle size.

[0061] Excessive overflow particle size: If the predicted overflow particle size is larger than the set range, it indicates that the grinding effect is insufficient. It is necessary to increase the water feed (Gsl), adjust the swirl feed pressure (Xg_yl), or increase the ore feed (Gkl) to enhance the grinding effect.

[0062] Overflow particle size too small: If the overflow particle size is smaller than the set range, it indicates excessive grinding. It may be necessary to reduce the water feed (Gsl) or ore feed (Gkl), or reduce the swirl pressure (Xg_yl) to prevent excessive grinding.

[0063] During each control cycle, the model predicts and adjusts the following process parameters in real time: Feed rate (Gkl): The feed rate is dynamically adjusted based on the predicted overflow particle size to ensure that the mill operates under optimal load.

[0064] Swirling feed pressure (Xg_yl): Adjust the swirling feed pressure according to the current grinding conditions to optimize the speed and stability of ore entering the mill.

[0065] Water supply (Gsl): By adjusting the water supply in real time, the slurry concentration is kept stable, ensuring the dynamic balance of fluids during the grinding process.

[0066] Pump pool level (Bc_yw) and pump pool water supply (Bc_jsl): Based on real-time data of pump pool level and water supply, adjust the water supply to optimize the material concentration in the system and avoid material sedimentation or excessive dilution.

[0067] 3. One-click control execution With a one-click control system, once the model predicts the overflow granularity, the system automatically adjusts the relevant operating parameters without manual intervention. Users can view the system's real-time feedback and suggestions through a simple interface.

[0068] This invention utilizes real-time production data acquisition and the LightGBM algorithm to construct a dynamic prediction model, enabling precise dynamic adjustment of key variables such as mill speed and feed rate. This optimizes the grinding process, reduces mill idling and overload operation, and significantly improves grinding efficiency compared to traditional control methods. The one-click intelligent control mode automatically optimizes key variables in the grinding and classification process, reducing energy consumption in the grinding step, lowering total energy consumption in mineral processing, effectively improving energy utilization, and reducing production costs. The LightGBM algorithm features efficient histogram and other optimization algorithms, reducing computational latency to milliseconds, meeting the real-time requirements of industrial sites, achieving precise control of the grinding and classification process, and improving the grinding fineness qualification rate. The high-efficiency computing power of the LightGBM algorithm, combined with an edge computing architecture, enables rapid control command issuance (latency <50ms), far exceeding the response speed of traditional PID control. It exhibits strong robustness to interference factors such as ore hardness fluctuations and media wear, ensuring stable grinding control, maintaining the continuity and stability of the grinding process, and reducing the impact of interference. This invention addresses production fluctuations caused by various factors. Operators only need to set target parameters, and the LightGBM model optimizes key parameters such as mill speed and feed rate in real time to ensure stable grinding fineness without frequent intervention. Combined with Time Lag Cross-Correlation (TLCC) analysis, the control strategy is dynamically adjusted to effectively cope with changes in ore hardness and particle size, reducing reliance on operator experience and minimizing the impact of human error on production. By integrating machine vision perception and industrial IoT technology, ore characteristics can be perceived in real time, dynamically adapting to fluctuations in ore properties, achieving adaptive control for complex working conditions, and improving production flexibility and adaptability. This invention achieves intelligent regulation of the grinding and classification process, maintaining stable operation of the grinding process through real-time monitoring and dynamic adjustment, reducing production interruptions caused by equipment failure or process fluctuations, and improving production continuity and stability. The intelligent decision engine based on the LightGBM algorithm uses a large amount of real-time production data for training and prediction, providing scientific and accurate decision-making basis for the production process, improving production reliability and quality stability.

Claims

1. A one-click grinding and classification control method based on LightGBM, characterized in that, Specifically, it includes: S1. Collect production data and preprocess the data; S2, Grinding and Classification Control, is used to control the feed rate, mill water supply, grinding concentration, and hydrocyclone classification process during the grinding and classification process; S3. Lightweight Gradient Boosting Machine (LightGBM) Modeling Based on TLCC Flattened Data Input; S4, System Deployment and Real-time Control Module, is used to deploy the trained prediction model to the PLC control layer to complete millisecond-level real-time decision-making. Flattening the data input is used to convert time series data into window-based input, specifically including: The transformation function retains only the data from the most recent time point for the auxiliary variables within the time series window; The transformed one-dimensional data is input into the prediction model for training and prediction, resulting in a single target. The loss is as follows: ② In formula ②, R w*m : Represents a time series window, a matrix or vector with w*m elements; w represents the first dimension of the time series window size; m represents the second dimension of the time series window; This represents the transformed time series window, which has w*m-1 elements; ③ In formula ③, Y represents the value of the loss function, used to measure the difference between the model's predictions and the true values; n represents the total number of samples, the number of samples required for the model's training; t' represents the time step for each sample; i，j : Represents the true value, the true target value at time point j for the i-th sample; This represents the predicted value, which is calculated by the naive pointwise prediction model.

2. The one-click grinding and classification control method based on LightGBM according to claim 1, characterized in that, In S1, production data from the grinding and classification process is collected in real time via the Industrial Internet of Things (IIoT). Python is used to load and preprocess the production data, specifically including: S11, Missing value imputation: Missing values ​​in production data are filled using linear interpolation. S12, Outlier Replacement: Outliers in production data are replaced using mean filtering. S13, Outlier Replacement: Outliers are identified by calculating the mean and standard deviation of process parameters and then replaced with the mean of the data sample, as shown in the following formula: ① In formula ①, x i This represents the i-th data point. This represents the average value of the data, and n represents the number of data points. S14. Divide the production data into training and testing sets: After data preprocessing, the production data is divided into training set D1 and test set T1 in a 7:3 ratio. Where D1 ∈ R l×m ; The size of the training set D1 is: R l×m , l represents the number of samples in the training set, m represents the number of feature variables, and R represents the set of real numbers.

3. The one-click grinding and classification control method based on LightGBM according to claim 2, characterized in that, The production data includes the ball mill's rotational speed, current, load, grinding time, mill decibels, classifier overflow concentration, hydrocyclone inlet pressure, return sand ratio, ore density and hardness index, steel ball filling rate and proportion, slurry pH value, actual measurement data from the particle size analyzer, and ore feed particle size.

4. The one-click grinding and classification control method based on LightGBM according to claim 1, characterized in that, In S2, the grinding and classification control includes: S21. Ore feed rate control: By comparing the actual feed rate signal of the belt scale with the set feed rate, the deviation signal is obtained and transmitted to the frequency converter, which then drives the feeder to adjust the feed rate. S22. Mill water supply control includes the control of the proportioned water supply during ore feeding and the control of the supplementary water supply during ball mill discharge: Water flow control during ore feeding: Adjust the water flow rate during the ore feeding process according to the ore feed rate, moisture content and particle size; Control of water replenishment during ball mill discharge: The actual flow rate on the flushing water pipe at the ball mill discharge end is compared with the set flow rate, and the deviation signal is transmitted to the regulating valve to control the amount of water replenishment. S23. Grinding concentration control: With a fixed feed rate, the mill water flow rate is adjusted to maintain the grinding concentration within the set range. S24. Hydrocyclone classification process control: The feed to the hydrocyclone is adjusted by regulating the speed of the switching pump to regulate the inlet pressure of the hydrocyclone, so that the inlet pressure of the hydrocyclone is within the set value range.

5. The one-click grinding and classification control method based on LightGBM according to claim 1, characterized in that, In S3, the lightweight gradient boosting machine (LightGBM) modeling based on TLCC flattened data input specifically includes: The time lag cross-correlation (TLCC) method is used to analyze the time lag difference between production data. A dynamic prediction model is constructed using the LightGBM algorithm, and the model training is optimized through a leaf-based growth strategy, one-sided gradient sampling, and parallel learning.

6. The one-click grinding and classification control method based on LightGBM according to claim 5, characterized in that, The Time Lag Cross-Correlation (TLCC) method calculates the lag time difference between two time series using the Pearson correlation coefficient, specifically including: 1) Slide the two time series one by one, and calculate the Pearson correlation coefficient each time. The calculation formula is as follows; ① In formula ①, The Pearson correlation coefficient represents the strength and direction of the linear relationship between variables x and y. Its value ranges from 1 to 1, where +1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear relationship. This represents the data for the i-th observation of variable x; variable This represents the data for the i-th observation of variable y; variable The average value of variable x; variable y represents the average value of the variable y; n represents the number of data points; Formula ① is used to determine which of two time series is leading. This method involves sliding the two series sequentially, calculating the Pearson correlation coefficient (r) between them at each slide. xy When the correlation coefficient between two sequences reaches its maximum under a certain time lag, this time lag difference is considered to be the true time lag difference between the two process parameters. 2) When the Pearson correlation coefficient reaches its maximum value, the corresponding lag time difference is the true lag time difference of the production data.

7. The one-click grinding and classification control method based on LightGBM according to claim 5, characterized in that, The leaf growth strategy described prioritizes leaves with the highest splitting gain for splitting. The unilateral gradient sampling described above is used to optimize production data, including: sorting by absolute gradient value, selecting the a100% data with the largest absolute gradient value and randomly selecting the b100% data with the smallest absolute gradient value; multiplying the data with the smallest absolute gradient value by a constant (1-a) / b, and then merging it with the data with the largest absolute gradient value to calculate the gain; Parallel learning is performed using feature parallelism, data parallelism, and voting parallelism strategies: Features are used in parallel to store all data on different machines, and the segmentation strategy is executed locally. Data parallelism uses a distributed reduction method to merge histograms; The voting process is parallelized by merging histograms through filtering TopK features and voting strategies.

8. The one-click grinding and classification control method based on LightGBM according to claim 1, characterized in that, The system deployment and real-time control module includes a model deployment unit and a real-time decision-making unit. The model deployment unit is used to deploy the trained prediction model to the PLC control layer. The real-time decision-making unit is used to make real-time decisions based on the real-time collected production data through the PLC control system, and the prediction model makes real-time decisions within a millisecond response time.

Images