Coal heavy medium separation control method based on lightgbm

Abstract

The coal heavy medium separation control method based on LightGBM disclosed in the application comprises the following steps: step 1, collecting the working condition data of the heavy medium separation system, and pre-processing the data to form an original data set; step 2, adopting a cross-validation method based on time series data slicing, the original data set is divided into a training set and a validation set; step 3, based on the time series slicing of the heavy medium separation process data, a separation control model is constructed with the flow valve opening degree as the output and the remaining process data as the input, and the model is trained by using the training set and the validation set; step 4, calling the trained model to connect the field control data, outputting the flow valve control parameters under the current working condition, and introducing a dynamic binning method to bin the model output results. The application solves the problem of the influence of the high-frequency adjustment of the existing traditional PID or fuzzy control algorithm on the service life of the flow valve.

Description

Technical Field

[0001] This invention belongs to the field of coal washing and beneficiation technology, specifically relating to a coal heavy media separation control method based on LightGBM. Background Technology

[0002] In my country's coal preparation industry, heavy media separation is the most widely used process. It relies on the density difference between coal and gangue for coal preparation. Most coal preparation plants use heavy media separation as their core process, supplemented by jigging, flotation, and other techniques to achieve efficient and clean coal preparation. Therefore, the intelligent construction of heavy media coal preparation plants plays a crucial role in the intelligent development of my country's coal industry.

[0003] Heavy media system control methods refer to the precise control measures used to effectively allocate and stabilize the heavy media involved in coal preparation, thereby achieving efficient coal separation. In recent years, with the integration and application of computer technology, automation technology, and artificial intelligence, heavy media system control methods have shown a trend towards diversification and intelligence.

[0004] Currently, the main control methods for recurrent media systems include the following: Traditional PID control: This method is based on the classic proportional-integral-derivative (PI-DE) control principle, achieving stable control of heavy media by adjusting the proportional coefficient, integral time, and derivative time. While PID control has advantages such as simple structure and ease of implementation, in practical applications, the density regulation system has a significant time delay, and the characteristic parameters of actuators such as diverters and clean water valves exhibit substantial nonlinearity. Excessive frequency adjustments often reduce the lifespan of regulating valves such as diverters and make-up water valves, affecting production continuity and increasing costs.

[0005] Fuzzy control: Based on the analysis of the dynamic characteristics of the density and level system of heavy medium suspension, a multivariable fuzzy control method was developed, which uses two independent fuzzy controllers to control the density and level of heavy medium suspension respectively. The method is simple to design and the control effect is good, as shown by the experiment in the network control system environment. However, it still has problems such as the excessive control frequency similar to PID control.

[0006] Neural network control: Neural network control utilizes the powerful self-learning and adaptive capabilities of neural networks to automatically adjust network parameters through training sample data, achieving intelligent control of the heavy medium system. Neural network controllers can learn online and adapt to various changes in the coal preparation process, thereby improving control accuracy and stability. However, neural network control requires a large amount of training data and computational resources, and suffers from problems such as overfitting. Summary of the Invention

[0007] The purpose of this invention is to provide a coal heavy medium separation control method based on LightGBM, which solves the problem of the impact of high-frequency regulation of existing traditional PID or fuzzy control algorithms on the life of the diverter valve.

[0008] The technical solution adopted in this invention is a coal heavy media separation control method based on LightGBM, comprising the following steps: Step 1: Collect the operating data of the heavy medium sorting system and preprocess the data to form the raw dataset; Step 2: The original dataset is divided into a training set and a validation set using a cross-validation method based on time-series data slicing. Step 3: Based on the heavy media sorting process data of time series slices, construct a sorting control model with the diversion valve opening as the output and the other process data as the input, and train the model using the training set and validation set; Step 4: Call the trained model to connect with the field control data, output the diversion valve control parameters under the current operating conditions, and introduce a dynamic binning method to bin the model output results.

[0009] The invention is further characterized in that, The operating data in step 1 includes suspension density, tank position, diversion valve opening, water supply valve opening, and magnetic material content; The preprocessing method uses the content of magnetic materials as the data consistency benchmark and achieves data consistency based on the Pearson correlation coefficient, as shown in the following expression: ; In the formula, X and Y represent two different variables, and r is the Pearson correlation coefficient between the two variables.

[0010] Based on the preprocessed original dataset, features are constructed, and the constructed feature vectors include density change rate, bucket change rate, daily flow accumulation, and single flow accumulation.

[0011] The expression for the time series data slice in step 2 is shown in the following formula: ; In the formula, For a slice subset, This represents a continuous feature slice from time tk to t. is the label value at time t; k is the sliding window length; After time series slicing, multiple training subsets are formed. One subset is selected as the training set, and all other subsets are used as the validation set.

[0012] The cross-validation method based on time-series data slicing involves extracting continuous slices of random length along the time axis, with the minimum slice length corresponding to a 1.5-hour work cycle and the maximum slice length corresponding to a shift work cycle. Then, five random slices are executed along the time dimension to divide the ratio of the training set to the validation set.

[0013] The input layer expression of the sorting control model is shown in the following equation: ; In the formula, The density value at time t. Bucket height, Magnetic content, The rate of change of density, The rate of change of barrel position. Daily traffic volume accumulation, This is a single-transaction flow accumulation; The output layer mapping is shown in the following equation: ; In the formula, Let t be the predicted value of the diversion valve opening. For the k-th decision tree, For tree sets; The loss function of the sorting control model uses the absolute average error as its core, while introducing an L2 regularization term to prevent overfitting. The expression is shown in the following formula: ; In the formula, The predicted value of the diversion valve opening at time t Let t be the actual value of the diversion valve opening. The weights of the leaf nodes are... is the coefficient of the regularization term.

[0014] The training parameters for the sorting control model include: num_leaves=67, max_depth=13, learning_rate=0.03, n_estimators=1050, feature_fraction=0.6, reg_alpha=0, reg_lambda=0.75, subsample=0.8.

[0015] In step 4, the output diverter valve control parameters are denoised using wavelet denoising and then divided into bins using a dynamic binning method.

[0016] The beneficial effects of this invention are: (1) This invention introduces the LightGBM model and a dynamic binning mechanism. It uses a time-series LightGBM model to predict the opening of the diverter valve, and combines wavelet denoising and dynamic binning to discretize the continuous output into a stable interval. It utilizes the model's nonlinear learning of historical regulation behavior to identify changes in operating conditions requiring adjustment, avoiding frequent adjustments caused by threshold sensitivity in PID or fuzzy control. Experiments show that dynamic binning reduces the number of valve actions by more than 30%, reduces mechanical wear, improves production continuity, reduces maintenance costs, and significantly extends valve life.

[0017] (2) This invention uses time-series data slicing, combined with features such as density change rate and barrel position change rate, to enable the model to capture the dynamic characteristics of the heavy medium system at different time scales. Compared with the high demand for data volume and computing resources of neural networks, LightGBM can achieve high-precision modeling with small samples, and solves the problem of poor generalization of neural networks by suppressing overfitting through L2 regularization and feature subsampling, making it suitable for working conditions with frequent fluctuations in coal quality. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of the coal heavy media separation control method based on LightGBM of the present invention. Figure 2a This is a correlation coefficient matrix diagram of the present invention, which combines the magnetic material content; Figure 2b This is a correlation coefficient matrix diagram of the diverter valve opening degree combined according to the present invention; Figure 2c This is a correlation coefficient matrix diagram of the present invention based on the opening degree of the water supply valve; Figure 3 This invention uses a partial construction of feature vector change trend diagrams; Figure 4 This is a schematic diagram of the cross-validation slicing method for time series data in this invention; Figure 5 This is a comparison diagram of the output binning before and after binning in the model of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Example 1 This invention relates to a coal heavy media separation control method based on LightGBM, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect operating data of the heavy medium separation system, including suspension density, combined medium tank position, diversion valve opening, water supply valve opening, and magnetic material content; and preprocess the data to form the raw dataset. The preprocessing method uses the content of magnetic materials as the data consistency benchmark and achieves data consistency based on the Pearson correlation coefficient, as shown in the following formula (1): (1); In the formula, X and Y represent two different variables, and r is the Pearson correlation coefficient between the two variables. The original parameters are solved iteratively to obtain... Figure 2a , Figure 2b , Figure 2c The correlation matrix shown shows that when the magnetic content is used as the data consistency benchmark, the diversion valve opening degree and the density of the mixing tank have a good correlation, which is consistent with the actual situation on site. Therefore, the magnetic content data is finally selected to merge the original data to form the original dataset.

[0021] Step 2: The original dataset is divided into a training set and a validation set using a cross-validation method based on time-series data slicing. Step 3: Based on the heavy media sorting process data of time series slices, construct a sorting control model with the diversion valve opening as the output and the other process data as the input, and train the model using the training set and validation set; Step 4: Call the trained model to connect with the field control data, output the diversion valve control parameters under the current operating conditions, and introduce a dynamic binning method to bin the model output results.

[0022] Example 2 This invention relates to a coal heavy media separation control method based on LightGBM, comprising the following steps: Step 1: Collect the main process parameters of the heavy media separation process in the coal preparation plant, and improve the time consistency of the original data through data standardization technology to form a complete original dataset of the heavy media separation process; Furthermore, based on the preprocessed original dataset, feature construction is performed to enhance the model's ability to perceive the original data. The constructed feature vectors include density change rate, bucket change rate, daily flow accumulation, and single flow accumulation.

[0023] Specifically, see equations (2) to (5) below: (1); In the formula, ρ represents the measurement data of the combined medium density meter, and t represents time; (2); In the formula, L represents the liquid level measurement data of the combined tank, and t represents time; (3); In the formula, Daily traffic volume accumulation; It is at 0:00 every day. It is 24:00 every day. The opening degree of the shunt valve at each sampling time. The sampling interval is 3 seconds in this embodiment. The diameter of the shunt pipe (fixed parameter); (4); In the formula, This is a single-transaction flow accumulation; The opening time of the diversion valve. This refers to the moment when the diverter valve closes for the first time since it was last opened. The opening degree of the shunt valve at each sampling time. The sampling interval is 3 seconds in this embodiment. The diameter of the shunt pipe (fixed parameter); like Figure 3 As shown, the curves of the training data samples after feature construction are presented within a time period.

[0024] Step 2: The training data is divided into multiple training subsets using a cross-validation method based on time-series data slicing. Each time, one subset is selected as the training set and all other subsets are used as the validation set. The training ends after this process, and the results of multiple training sessions are summarized to obtain the final index. Step 3: Based on the heavy medium sorting process data of time series slices, with the diversion valve opening as the model output and other process data as the model input, construct a sorting control model based on LightGBM, and form a model file after the target is met; Step 4: Call the trained algorithm model file and connect it with the field control data to provide the control parameters of the diversion valve under the current operating conditions; introduce the dynamic binning method to bin the model output results.

[0025] Example 3 Based on the above embodiment 2, the expression for the time-series data slice in step 2 of this invention is as follows (5): (5); In the formula, For a slice subset, This represents a continuous feature slice from time tk to t. is the label value at time t; k is the sliding window length (in this embodiment, k=1440, corresponding to a 30-minute working condition); The specific steps are as follows: S1, Time-series slice generation: Extract continuous slices of random length according to the time axis. Minimum slice length: Lmin=300 (corresponding to 1.5 hours), maximum slice length: Lmax=2880 (corresponding to 1 shift). S2, Training / Validation Set Partition: Training set index is... Continuous data segments preserve the complete evolution of operating conditions; the test set is Continuous data segments; S3, Five-fold cross-cutting: Executes 5 random slices in the time dimension, such as... Figure 4 The diagram illustrates the principle of random slicing; at the same time, the training / validation set ratio is randomly assigned.

[0026] Example 4 Based on Embodiment 2 above, the input layer expression of the sorting control model in step 3 of this invention is as follows (6): (6); In the formula, The density value at time t. Bucket height, Magnetic content, The rate of change of density, The rate of change of barrel position. Daily traffic volume accumulation, This is a single-transaction flow accumulation; The output layer mapping is shown in equation (7): (7); In the formula, Let t be the predicted value of the diversion valve opening. For the k-th decision tree, For tree sets; The loss function of the sorting control model uses the absolute average error as the core, and introduces an L2 regularization term to prevent overfitting. The expression is shown in the following equation (8): (8); In the formula, The predicted value of the diversion valve opening at time t Let t be the actual value of the diversion valve opening. The weights of the leaf nodes are... This is the regularization coefficient (in this example, it is 0.75).

[0027] After model parameter tuning and optimization, the final model output MAE is 1.03, which meets the prediction accuracy requirements.

[0028] The model file was exported using the joblib tool and saved as a .pkl file.

[0029] Example 5 Based on Example 4 above, in step 3 of this invention, the training parameters of the sorting control model include: num_leaves=67, max_depth=13, learning_rate=0.03, n_estimators=1050, feature_fraction=0.6, reg_alpha=0, reg_lambda=0.75, subsample=0.8.

[0030] The num_leaves setting is 67, which ensures the model's ability to capture the dynamic interaction relationships of multiple features such as density change rate and bucket position change rate during the heavy media sorting process, while avoiding the risk of overfitting caused by too many leaf nodes, and adapting to the working condition fluctuation characteristics in the sorting data; max_depth=13 limits the depth of the tree, balancing the model's learning ability and computational efficiency for complex working conditions such as the difference between heavy-load and light-load periods, and preventing the over-mining of invalid associations in noisy data.

[0031] The `learning_rate=0.03` setting, combined with `n_estimators=1050` iterations, forms a training logic of small-step iterations and gradual optimization. A smaller learning rate prevents model parameters from oscillating due to excessively large single update amplitudes, while sufficient iterations ensure the model fully learns the historical patterns of manually adjusted diversion valves, such as the opening adjustment amplitude under different density deviations, ultimately converging to a stable prediction effect. `feature_fraction=0.6` allows the model to randomly select 60% of its features for training in each iteration. By introducing randomness, it reduces dependence on individual noisy features, such as instantaneous fluctuations in magnetic content data, thereby improving the model's generalization ability to different raw coal characteristics and different production periods.

[0032] The regularization parameter combination of reg_alpha=0 and reg_lambda=0.75 indicates that the model mainly restricts the weights of the decision leaf nodes through L2 regularization, effectively suppressing parameter inflation caused by feature collinearity, such as the strong correlation between the density of the mixing tank and the opening of the diversion valve, and further reducing the risk of overfitting. subsample=0.8 enhances the model's adaptability to unsteady conditions such as sudden adjustments to the water supply valve and fluctuations in raw coal quality by randomly sampling 80% of the sample data in each training session, ensuring that the prediction of the diversion valve opening in actual production can both conform to the logic of manual operation and stably cope with changes in operating conditions.

[0033] Example 6 Based on Example 4 above, this embodiment calls the trained algorithm model file in step 4 and connects it with the field control data to provide the specific control parameters of the diversion valve under the current operating conditions, including: During the operation of the heavy medium sorting system, raw data such as the density of the suspension in the mixed medium tank, the tank height, the magnetic content, and the current opening degree of the diversion valve are collected in real time through sensors such as densitometers, level gauges, and magnetic content meters. After being uploaded to the host computer by the PLC logic controller, the data is synchronously transmitted to the control algorithm module via the OPC UA protocol. To meet the input feature requirements of the sorting control model, the algorithm module further constructs dynamic feature parameters based on the received real-time data: the density change rate of the mixed medium tank is calculated by dividing the density difference between the current time and the previous 5 sampling times by 15 seconds (i.e., 5 sampling intervals, each interval being 3 seconds); the tank level change rate of the mixed medium tank is calculated by dividing the tank level difference within the same time window by 15 seconds; the single diversion flow rate accumulation is the product of the opening value of the diversion valve from the current opening time to the current time, the 3-second sampling interval, and the fixed diameter of the diversion pipe; the daily diversion flow rate accumulation starts from 0:00 each day and accumulates the product of the opening, interval, and pipe diameter of all sampling times of the day. These dynamic features, together with the raw real-time data, constitute a complete set of model input parameters, ensuring that the model can capture the temporal variation patterns of the remediation sorting process.

[0034] During the model invocation phase, the algorithm loads the pre-trained .pkl format model file using the joblib tool and connects the constructed feature parameter set to the model input layer. Combined with the system's real-time data sampling frequency of 3 seconds per sampling, the model automatically updates the input parameters every 3 seconds and outputs the predicted value of the diverter valve opening at the corresponding moment, achieving real-time response to the on-site working conditions.

[0035] Since high-frequency sampling is prone to introducing instantaneous fluctuation noise, the prediction results are first processed by wavelet denoising. By selecting a wavelet basis function (such as db4 wavelet) that is suitable for the heavy media sorting time series signal, a 3-5 level decomposition is performed. The high-frequency noise sub-band is then reconstructed after soft threshold shrinkage processing, which effectively filters out the glitch interference in the sampling process and provides more stable basic data for subsequent binning optimization.

[0036] Based on the denoised prediction results, the algorithm introduces an output binning method for further optimization: according to the historical binning effect and real-time operating condition deviation (such as density fluctuation amplitude and magnetic material content stability), the continuous diverter valve opening prediction value is discretized into several intervals (bins), such as... Figure 5As shown, the binning process further controls the model output jitter caused by input parameter fluctuations. The opening fluctuation range within each bin is controlled within a preset threshold, and the bin boundaries are adaptively adjusted according to changes in operating conditions. For example, when the density fluctuation of the mixing tank is small, the bin interval can be appropriately increased to reduce the number of adjustments; when the density fluctuation is severe, the bin interval is reduced to ensure control accuracy. This dynamic binning strategy not only reduces the high-frequency fluctuation of the model output through discretization, keeping the MAE error stably controlled within 1.03, but also reduces the ineffective adjustment frequency of the diversion valve. Compared with the traditional continuous adjustment method, the number of valve actions is reduced by more than 30% after binning, significantly reducing mechanical structure wear and extending equipment life.

[0037] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0038] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A coal heavy media separation control method based on LightGBM, characterized in that, Includes the following steps: Step 1: Collect the operating data of the heavy medium sorting system and preprocess the data to form the raw dataset; Step 2: The original dataset is divided into a training set and a validation set using a cross-validation method based on time-series data slicing. Step 3: Based on the heavy media sorting process data of time series slices, construct a sorting control model with the diversion valve opening as the output and the other process data as the input, and train the model using the training set and validation set; Step 4: Call the trained model to connect with the field control data, output the diversion valve control parameters under the current working condition, and introduce a dynamic binning method to bin the model output results. The operating data in step 1 includes suspension density, tank position, diversion valve opening, water supply valve opening, and magnetic material content. The preprocessing method uses the content of magnetic materials as the data consistency benchmark and achieves data consistency based on the Pearson correlation coefficient, as shown in the following expression: ； In the formula, X and Y represent two different variables, and r is the Pearson correlation coefficient between the two variables; Based on the preprocessed original dataset, feature construction is performed. The constructed feature vector includes density change rate, bucket change rate, daily flow accumulation, and single flow accumulation. The expression for the time series data slice in step 2 is shown in the following formula: ； In the formula, For a slice subset, This represents a continuous feature slice from time tk to t. The label value at time t; k is the length of the sliding window; After time series slicing, multiple training subsets are formed. One subset is selected as the training set, and all other subsets are used as the validation set. The cross-validation method based on time-series data slicing specifically involves extracting continuous slices of random length along the time axis, with the minimum slice length corresponding to a 1.5-hour work cycle and the maximum slice length corresponding to a shift work cycle; then, five random slices are executed along the time dimension to divide the ratio of the training set and the validation set. In step 4, the output diverter valve control parameters are denoised using wavelet denoising and then divided into bins using a dynamic binning method.

2. The coal heavy media separation control method based on LightGBM according to claim 1, characterized in that, The input layer expression of the sorting control model is shown in the following equation: ； In the formula, The density value at time t. Bucket height, Magnetic content, The rate of change of density, The rate of change of barrel position. Daily traffic volume accumulation, This is a single-transaction flow accumulation; The output layer mapping is shown in the following equation: ； In the formula, Let t be the predicted value of the diversion valve opening. For the k-th decision tree, For tree sets; The loss function of the sorting control model uses the absolute average error as its core, while introducing an L2 regularization term to prevent overfitting. The expression is shown in the following formula: ； In the formula, The predicted value of the diversion valve opening at time t Let t be the actual value of the diversion valve opening. The weights of the leaf nodes are... is the coefficient of the regularization term.

3. The coal heavy media separation control method based on LightGBM according to claim 2, characterized in that, The training parameters of the sorting control model include: num leaves =67、max depth =13、learning rate =0.03、n estimators =1050、feature fraction =0.6、reg alpha =0、reg lambda =0.75、subsample=0.8。

Images