Grinding particle size prediction control method based on deep learning

Abstract

The deep learning-based grinding particle size prediction and control method includes: real-time acquisition of grinding process operating parameters through a sensor network deployed in the grinding system; smoothing and filtering the operating parameters through a sliding window; establishing a deep spatiotemporal feature extraction network; embedding process mechanisms such as mass conservation, energy balance, and particle size dynamics into the network with differentiable constraints; constructing a particle size prediction network based on deep spatiotemporal features and constraints; calculating the optimal control quantity in real time using a sequential quadratic programming algorithm; comparing the actual measured particle size data with the predicted value to trigger an online update mechanism for the particle size prediction network; and continuously evaluating the error index of the particle size prediction network. This method effectively improves the accuracy and stability of grinding particle size prediction, enhances the system's generalization ability under complex operating conditions, and improves its long-term operational reliability. Specifically, it includes six process steps and conditions, possessing the advantages of effectively improving grinding particle size prediction accuracy and control stability, as well as enhancing the system's generalization ability and long-term operational reliability under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of mineral processing control technology, and in particular to a grinding particle size prediction and control method based on deep learning. Background Technology

[0002] The grinding process, a crucial step in mineral processing, directly impacts subsequent beneficiation efficiency and resource recovery rates through precise particle size control. Traditional grinding control systems often rely on empirical formulas, static models, or PID-based feedback mechanisms, making them ill-suited to complex dynamic factors such as ore property fluctuations, equipment wear, and operational disturbances. In recent years, deep learning-based soft sensor and predictive control methods have been introduced into industrial process monitoring, aiming to establish nonlinear mappings between input variables and key quality indicators through data-driven approaches.

[0003] To address the aforementioned issues, several publications have revealed methods, such as CN109856234B "An Online Grinding Particle Size Detection Method and System Based on Machine Vision." This method proposes using a high-definition industrial camera to capture images of the mill discharge port, extracting particle outlines through traditional image processing algorithms (such as edge detection and threshold segmentation), and then statistically analyzing particle size distribution. This solves the problems of high latency and inability to provide real-time particle size feedback in manual sampling and testing, enabling minute-level online monitoring of grinding particle size and reducing labor costs. However, this method relies on lighting conditions and ore color contrast. When there are large fluctuations in slurry concentration or severe bubble interference, image segmentation errors increase. Furthermore, it only provides detection results and lacks integration with the mill itself. Closed-loop linkage control strategies for operating parameters such as feed rate and steel ball filling rate belong to the category of "measuring but not controlling." CN107657392A, "Soft Measurement Method for Key Indicators of Grinding Process Based on Support Vector Regression (SVR)," utilizes the SVR algorithm to establish a nonlinear mapping model between easily measurable variables such as mill power, feed rate, and overflow concentration and grinding particle size. This replaces particle size indicators that are difficult to measure online, solving the problem that grinding particle size is difficult to obtain directly online through physical sensors, reducing hardware investment costs, and providing relatively accurate measurement under relatively stable operating conditions. While predicting particle size trends provides operators with a reference, the SVR model, being a shallow machine learning approach, has limited ability to extract features from high-dimensional, strongly coupled grinding processes. When ore hardness changes abruptly or equipment wears down, the model's generalization ability drops sharply, requiring frequent retraining and making it difficult to adapt to complex dynamic environments. CN110244567B, "A Model Predictive Control (MPC) Method for Grinding and Classification Processes," constructs state-space equations based on a mechanistic model and employs a model predictive control (MPC) algorithm for multivariate coordinated control of feed water and feed rate. This method aims to stabilize grinding concentration and particle size, addressing the issues of slow response and large overshoot in traditional PID control when dealing with systems with large time delays and multivariable coupling. It can improve the stability of the control system and, to some extent, suppress fluctuations caused by operating condition disturbances. However, since this method highly relies on an accurate mechanistic model, and the grinding process involves complex crushing mechanics and fluid mechanics, the mechanistic model suffers from two problems: first, it is difficult to accurately describe all dynamic characteristics, i.e., the "model mismatch" problem; and second, it does not fully utilize the implicit patterns in historical production big data, resulting in insufficient prediction accuracy and control robustness under extreme operating conditions.

[0004] In summary, most existing deep learning models do not fully incorporate process mechanism constraints and rely solely on black-box fitting, resulting in a significant decrease in generalization ability when operating conditions change abruptly. The grinding process involves multi-source heterogeneous sensor data, and traditional models lack the ability to jointly model temporal features and spatial correlations, making it difficult to accurately capture the dynamic laws of particle size evolution.

[0005] Therefore, it is of great significance to develop a grinding particle size prediction and control method based on deep learning. Summary of the Invention

[0006] The objective of this invention is to overcome the shortcomings of existing methods and provide a deep learning-based method for predicting and controlling grinding particle size. This method can adapt to complex dynamic factors such as fluctuations in ore properties, equipment wear, and operating condition disturbances, and can also establish a nonlinear mapping relationship between input variables and key quality indicators through a data-driven approach.

[0007] To accomplish the above tasks, the present invention adopts the following technical solution:

[0008] The deep learning-based grinding particle size prediction and control method includes: real-time acquisition of grinding process operating parameters through a sensor network deployed in the grinding system; smoothing and filtering the operating parameters using a sliding window; establishing a deep spatiotemporal feature extraction network; embedding process mechanisms such as mass conservation, energy balance, and particle size dynamics into the network with differentiable constraints; constructing a particle size prediction network based on deep spatiotemporal features and constraints; calculating the optimal control quantity in real time using a sequential quadratic programming algorithm; comparing the actual measured particle size data with the predicted values ​​to trigger an online update mechanism for the particle size prediction network; and continuously evaluating the error index of the particle size prediction network. This method effectively improves the accuracy and stability of grinding particle size prediction and enhances the system's generalization ability and long-term operational reliability under complex operating conditions. Specifically, it includes the following process steps and conditions:

[0009] Step 1. Data Acquisition and Preprocessing: Real-time acquisition of operating parameters such as feed rate, mill current, and bearing vibration through sensor network; noise removal using sliding window smoothing filter, outlier removal using 3 sigma criterion, and linear interpolation and time alignment for missing data.

[0010] Step 2. Construct a deep spatiotemporal feature extraction network: Establish an architecture with a bidirectional long short-term memory network (Bi-LSTM) as the core and combined with a convolutional neural network (CNN); use the attention mechanism to extract the temporal dependence and spatial correlation features of multi-sensor data;

[0011] Step 3. Embedding process mechanism constraints: Transform physical equations such as mass conservation, energy balance, and particle size dynamics into differentiable constraints; embed these constraints into the forward propagation process of the neural network using the Lagrange multiplier method to ensure that the model conforms to physical laws;

[0012] Step 4. Construct a particle size prediction network: Based on the extracted spatiotemporal features and mechanistic constraints, construct a sequence-to-sequence (Seq2Seq) prediction model; output a particle size distribution curve covering the entire grinding cycle (e.g., the next 30 sampling points);

[0013] Step 5. Calculate the optimal control quantity: With minimizing granularity deviation, minimizing energy consumption, and ensuring equipment stability as optimization objectives, the optimal control command is calculated in real time and output under the premise of satisfying load constraints using the Sequential Quadratic Programming (SQP) algorithm.

[0014] Step 6. Trigger the online update mechanism: compare the actual measurement granularity with the predicted value in real time; when the deviation exceeds the preset threshold (e.g., 3%), automatically start the online update and use the stochastic gradient descent algorithm with momentum to adjust the network parameters to correct the error;

[0015] Step 7. Performance Evaluation and Retraining: Continuously monitor the root mean square error of the model and the integral absolute error of the controller; when the performance indicators degrade beyond the set limit (e.g., 25%), automatically start the model retraining process to ensure continuous adaptation to changes in operating conditions.

[0016] Step 8. Continuously evaluate error and initiate retraining: Continuously evaluate the root mean square error of the granular prediction network and the integral absolute error of the controller. When the performance index degrades by more than 25%, the granular prediction network retraining process is automatically initiated.

[0017] Compared with the prior art, the innovative points and advantages or beneficial effects of this invention are as follows:

[0018] 1. By using a sensor network to collect the operating parameters of the grinding process in real time and a depth spatiotemporal feature extraction network, the comprehensive capture and efficient extraction of multi-dimensional operating features of the grinding process can be achieved.

[0019] 2. By transforming the mass conservation equation, energy balance relationship, and particle size distribution dynamics model in the grinding process into differentiable constraints, and embedding them into the forward propagation process of a deep neural network using the Lagrange multiplier method, the generalization ability and prediction stability under sudden changes in operating conditions are significantly improved.

[0020] 3. By constructing a granularity prediction network based on deep spatiotemporal features and constraints, end-to-end closed-loop collaboration between granularity prediction and control optimization is achieved, effectively reducing the systematic deviation between control commands and granularity targets.

[0021] 4. By comparing the actual measured particle size data with the predicted value, when the deviation exceeds the preset threshold of 3%, the online update mechanism of the particle size prediction network is triggered to ensure the adaptability and reliability of the system in long-term operation, and finally achieve accurate prediction of grinding particle size in complex and ever-changing industrial environments. Attached Figure Description

[0022] The specific structure of the invention is shown in the following figures.

[0023] Figure 1 This is a schematic diagram of the process flow of a grinding particle size prediction and control method based on deep learning proposed in this invention.

[0024] The present invention will be further described in detail below with reference to the accompanying drawings. Detailed Implementation

[0025] 1. For example Figure 1 As shown, the grinding particle size prediction and control method based on deep learning provided by this invention includes: real-time acquisition of grinding process operating parameters through a sensor network deployed in the grinding system; smoothing and filtering the operating parameters through a sliding window; establishing a deep spatiotemporal feature extraction network; embedding process mechanisms such as mass conservation, energy balance, and particle size dynamics into the network with differentiable constraints; constructing a particle size prediction network based on deep spatiotemporal features and constraints; calculating the optimal control quantity in real time using a sequential quadratic programming algorithm; comparing the actual measured particle size data with the predicted value to trigger an online update mechanism for the particle size prediction network; and continuously evaluating the error index of the particle size prediction network. This method can effectively improve the grinding particle size prediction accuracy and control stability, enhance the system's generalization ability under complex working conditions, and improve long-term operational reliability. Specifically, it includes the following process steps and conditions:

[0026] Step 1. Data Acquisition and Preprocessing: Real-time acquisition of operating parameters such as feed rate, mill current, and bearing vibration through sensor network; noise removal using sliding window smoothing filter, outlier removal using 3 sigma criterion, and linear interpolation and time alignment for missing data.

[0027] Step 2. Construct a deep spatiotemporal feature extraction network: Establish an architecture with a bidirectional long short-term memory network (Bi-LSTM) as the core and combined with a convolutional neural network (CNN); use the attention mechanism to extract the temporal dependence and spatial correlation features of multi-sensor data;

[0028] Step 3. Embedding process mechanism constraints: Transform physical equations such as mass conservation, energy balance, and particle size dynamics into differentiable constraints; embed these constraints into the forward propagation process of the neural network using the Lagrange multiplier method to ensure that the model conforms to physical laws;

[0029] Step 4. Construct a particle size prediction network: Based on the extracted spatiotemporal features and mechanistic constraints, construct a sequence-to-sequence (Seq2Seq) prediction model; output a particle size distribution curve covering the entire grinding cycle (e.g., the next 30 sampling points);

[0030] Step 5. Calculate the optimal control quantity: With minimizing granularity deviation, minimizing energy consumption, and ensuring equipment stability as optimization objectives, the optimal control command is calculated in real time and output under the premise of satisfying load constraints using the Sequential Quadratic Programming (SQP) algorithm.

[0031] Step 6. Trigger the online update mechanism: compare the actual measurement granularity with the predicted value in real time; when the deviation exceeds the preset threshold (e.g., 3%), automatically start the online update and use the stochastic gradient descent algorithm with momentum to adjust the network parameters to correct the error;

[0032] Step 7. Performance Evaluation and Retraining: Continuously monitor the root mean square error of the model and the integral absolute error of the controller; when the performance indicators degrade beyond the set limit (e.g., 25%), automatically start the model retraining process to ensure continuous adaptation to changes in operating conditions.

[0033] Step 8. Continuously evaluate error and initiate retraining: Continuously evaluate the root mean square error of the granular prediction network and the integral absolute error of the controller. When the performance index degrades by more than 25%, the granular prediction network retraining process is automatically initiated.

[0034] The present invention may be further defined as follows:

[0035] The sensor network in step one includes at least an electromagnetic flowmeter, a vibration acceleration sensor, a power transmitter, and a laser particle size analyzer.

[0036] The data preprocessing or operating parameters in step one include time alignment of multi-source data and the use of a dynamic time warping algorithm to eliminate minor offsets in the timestamps collected by different sensors.

[0037] The differentiable constraints in step three include at least the crushing rate function, the classification efficiency curve, and the material balance equation. The crushing rate function adopts the first-order crushing kinetic model, the classification efficiency curve adopts the Whiten model, and the material balance equation involves the feeding, discharging, and circulating load.

[0038] The prediction network in step four uses an encoder and a decoder.

[0039] The encoder in step four consists of a four-layer bidirectional long short-term memory network, with 64, 128, 256, and 128 hidden units in each layer, respectively.

[0040] The fourth decoder consists of three layers of gated recurrent units with 128, 64 and 32 hidden units respectively. Its attention mechanism adopts an additive attention model with a dimension of 64.

[0041] The granular prediction network in step six adopts a sequence-to-sequence architecture.

[0042] The optimization objective function of the granularity prediction network predictive controller in step six comprehensively considers the granularity qualification rate, energy consumption index and equipment operation stability, with weight coefficients of 0.6, 0.3 and 0.1, respectively.

[0043] General Implementation Examples

[0044] This paper presents a deep learning-based method for predictive control of grinding particle size. This method includes real-time acquisition of operating parameters such as feed rate, mill current, bearing vibration frequency, classifier return sand volume, and slurry concentration during the grinding process via a sensor network deployed in the grinding system. The acquired operating parameters are then smoothed using a sliding window filtering process with a window width of 50 sampling points. Outlier data points are removed using the 3σ or Laida criterion, and missing data is filled using time-series linear interpolation. A deep spatiotemporal feature extraction network is established, employing a bidirectional long short-term memory network with an attention mechanism as the core architecture. Spatial correlation features of multi-sensor data are extracted in parallel through a convolutional neural network module. The mass conservation equation, energy balance relationship, and particle size distribution dynamics model in the grinding process are transformed into differentiable constraints. The Granger multiplier method is embedded into the forward propagation process of a deep neural network. Based on deep spatiotemporal features and constraints, a particle size prediction network is constructed to output the particle size distribution curves for the next 30 sampling points, with the prediction time span covering the entire grinding cycle. The optimization objective is to minimize the deviation between the predicted particle size distribution and the target particle size, while considering mill load constraints and energy consumption limitations. A sequential quadratic programming algorithm is used to calculate the optimal control quantity in real time. The actual measured particle size data is compared with the predicted value. When the deviation exceeds a preset threshold of 3%, the online update mechanism of the particle size prediction network is triggered, and the network parameters are adjusted using a stochastic gradient descent algorithm with the driving quantity. The root mean square error of the particle size prediction network and the integral absolute error of the controller are continuously evaluated. When the performance index degrades by more than 25%, the retraining process of the particle size prediction network is automatically started.

[0045] Specifically, as follows:

[0046] First, a sensor network deployed throughout the mill system collects real-time operating parameters of the grinding process, including feed rate, mill current, bearing vibration frequency, classifier return sand volume, and slurry concentration. The sensor network sampling frequency is no less than 100 Hz, with a data acquisition accuracy of 0.5%. The sensor network includes electromagnetic flowmeters, vibration accelerometers, power transmitters, and laser particle size analyzers. Specifically, the electromagnetic flowmeter measures the feed rate, with a range of 0 to 500 tons per hour and an accuracy of 0.5%; the vibration accelerometer, installed in the mill bearing housing, monitors the bearing vibration frequency, with a frequency response range of 5 to 5 kHz and a sensitivity of 100 mV / g; the power transmitter acquires the mill current signal in real time with a measurement accuracy of 0.2%; and the laser particle size analyzer measures the product particle size distribution online, with a measurement range of 0.1 to 2000 micrometers and a repeatability error of less than 1%. All sensors are connected to a central data acquisition terminal via an industrial Ethernet network to ensure real-time and complete data transmission. Next, the collected operating parameters are processed using a sliding window smoothing filter with a window width of 50 sampling points. A 3-sigma criterion is used to remove outlier data points, and missing data is filled using time-series-based linear interpolation. The sliding window smoothing filter uses a centrally symmetric rectangular window function, executed independently for each sensor channel to suppress high-frequency noise interference. The 3-sigma criterion is calculated based on the mean μ and standard deviation σ of the data within the sliding window. If a data point x satisfies |x - μ| > 3σ, it is identified as an outlier and removed. For missing data caused by communication interruptions or sensor malfunctions, a time-series-based linear interpolation method is used for completion. This involves linearly fitting the nearest valid data points before and after the missing point to generate intermediate missing values. Optionally, to eliminate time offsets caused by sampling and transmission delays between different sensors, the system performs time alignment, using a dynamic time warping algorithm to perform non-linear time axis matching on the data, ensuring that all variables are strictly synchronized under a unified time reference.

[0047] Next, a deep spatiotemporal feature extraction network is established, employing a bidirectional long short-term memory network with an attention mechanism as its core architecture. A convolutional neural network module extracts spatial correlation features from multi-sensor data in parallel. The deep spatiotemporal feature extraction network adopts an encoder-decoder structure. The encoder consists of four layers of bidirectional long short-term memory networks, with 64, 128, 256, and 128 hidden units per layer, extracting deep temporal dependencies of the input sequence layer by layer. The decoder consists of three layers of gated recurrent units, with 128, 64, and 32 hidden units per layer, used to predict future states. An additive attention mechanism with a dimension of 64 is introduced between the encoder and decoder to dynamically assign importance weights for different historical time steps to the current prediction. Optionally, to capture the spatial correlation between multiple sensors, a one-dimensional convolutional neural network module is deployed in parallel. This module receives the original multi-channel input data, uses three convolutional kernels with dimensions of 3, 5, and 7, and outputs 32 channels. After compressing the feature dimension through a max-pooling layer, it is concatenated with the output of the bidirectional long short-term memory network to form a joint representation that integrates spatiotemporal features. Then, the mass conservation equation, energy balance relationship, and particle size distribution kinetic model in the grinding process are transformed into differentiable constraints, which are then embedded into the forward propagation process of a deep neural network using the Lagrange multiplier method. The differentiable constraints include the crushing rate function, the classification efficiency curve, and the material balance equation. The crushing rate function adopts a first-order crushing kinetic model, expressed as:

[0048]

[0049] In the formula M i k represents the mass fraction of the i-th particle size. i Let b be the crushing rate constant for this particle size. ji The distribution coefficient is used to break down materials from particle size j to particle size i. The classification efficiency curve adopts the Whiten model to describe the separation efficiency of the classifier for materials of different particle sizes. The material balance equation comprehensively considers the feed, discharge and circulation load to ensure the conservation of total mass.

[0050] The aforementioned discretization and transformation into a differentiable form are added as soft constraint terms to the loss function of the neural network. An augmented Lagrange function is constructed using the Lagrange multiplier method. During forward propagation, the degree of violation of the theoretical constraints is synchronously calculated, and during backpropagation, the network parameters and Lagrange multipliers are jointly updated. Then, based on deep spatiotemporal features and constraints, a particle size prediction network is constructed, outputting particle size distribution curves for the next 30 sampling points, with a prediction time span covering the entire grinding cycle. The particle size prediction network adopts a sequence-to-sequence architecture. The encoder receives input features from the past 60 time steps, including preprocessed multi-source sensor data and extracted spatiotemporal features. The decoder progressively outputs particle size prediction values ​​for the next 30 time steps, with each time step corresponding to a complete particle size distribution vector (e.g., 200 particle size intervals divided at 10-micrometer intervals). The prediction interval is 10 seconds, therefore the total prediction time domain is 300 seconds, sufficient to cover the complete dynamic process of the grinding process from changes in feed to a stable product particle size response. Then, minimizing the deviation between the predicted particle size distribution and the target particle size is used as the optimization objective. Simultaneously considering mill load constraints and energy consumption limitations, a sequential quadratic programming algorithm is employed to calculate the optimal control quantity in real time. The optimization objective function of the particle size prediction network controller comprehensively considers the particle size qualification rate, energy consumption index, and equipment operational stability, with weighting coefficients of 0.6, 0.3, and 0.1, respectively. The particle size qualification rate is defined as the probability integral of the predicted particle size falling within the target range; the energy consumption index is characterized by the sum of squares of the predicted mill current; equipment operational stability is measured by the sum of squares of the rate of change of the control quantity to suppress frequent and large-scale adjustments; load constraints include mill current not exceeding 85% of the rated value, feed rate fluctuations controlled within ±10%, and the physical upper limit of the classifier return sand volume. Next, the actual measured particle size data is compared with the predicted value. When the deviation exceeds a preset threshold of 3%, the online update mechanism of the particle size prediction network is triggered, using a stochastic gradient descent algorithm to adjust the network parameters. The online update mechanism includes two levels: short-term correction and long-term update. At the short-term correction level, a Kalman filter algorithm is used to compensate for the prediction deviation in real time; at the long-term update level, continuous... The cumulative prediction error is assessed. When the cumulative deviation (defined as the average absolute deviation over 5 consecutive sampling points) exceeds 5%, a full update of the network parameters is initiated. The update process employs a stochastic gradient descent algorithm with momentum, a momentum coefficient of 0.9, an initial learning rate of 0.001, and an exponential decay strategy. Finally, the root mean square error (RMSE) of the granular prediction network and the integral absolute error (IARE) of the controller are continuously evaluated. When the performance degradation exceeds 25%, the granular prediction network retraining process is automatically initiated. Performance monitoring metrics also include the mean absolute percentage error (MASE) of the granular prediction network and the overshoot of the controller. The MSE measures the overall accuracy of the granular prediction, the IARE reflects the cumulative effect of control deviation, the MASE assesses the relative error level, and the overshoot characterizes the smoothness of the control response.Optionally, multiple warning thresholds can be set: when the root mean square error or integral absolute error degrades by more than 15% compared to the baseline value, a mild alarm is triggered and a log is recorded; when the mean absolute percentage error exceeds 8% for 24 consecutive hours or the overshoot exceeds 12% for a continuous period of time, the system automatically switches to the backup control strategy and immediately starts the model retraining process.

[0051] In addition, it should be noted that the method of this invention is deployed on an industrial edge computing platform and adopts a distributed architecture to process the parallel monitoring tasks of multiple mills. In the actual application of a copper mine beneficiation plant, this method improved the stability of the -200 mesh particle size qualification rate from the original ±8% to ±2.5%, reduced the energy consumption per ton of ore by 7.3%, and increased the annual economic benefits by more than 12 million yuan.

[0052] Example 1

[0053] A large copper mine beneficiation plant is processing porphyry copper ore with a high Mohs hardness (6-7), and the raw ore grade fluctuates frequently. Traditional PID control often leads to mill over-grinding or bloating, resulting in a particle size qualification rate of only 78%.

[0054] Implementation Process: Data Acquisition: Access the DCS system and collect historical data from the past two years, including feed rate, mill power, return sand volume, overflow concentration, and manually analyzed particle size data, with a sampling frequency of 1 minute / time; Model Training: Utilize the method of this invention to construct a prediction model, focusing on enhancing the extraction of time-series features of the lag relationship between "mill power and feed rate," and introducing the crushing energy consumption mechanism formula as the regularization term of the loss function; Closed-Loop Control: Deploy the model to an edge computing gateway to output the optimal feed rate setpoint and water replenishment command in real time.

[0055] Implementation results: After three months of continuous operation, the pass rate of grinding particle size (-200 mesh percentage) increased to 92.5%, the unit energy consumption of the mill decreased by 4.8%, and the shutdown accident caused by sudden change in ore hardness was effectively avoided.

[0056] Example 2

[0057] A certain iron ore beneficiation production line operates under a multivariate strongly coupled condition. The production line has a multivariate coupling problem, where the overflow concentration of the classifier and the mill load interfere with each other, and the slurry viscosity varies significantly with seasonal temperature.

[0058] Implementation process: Feature fusion: Ambient temperature and slurry viscosity estimates are added as auxiliary input features to the model of this invention, and the correlation matrix between multiple variables is extracted using the spatial convolution module; Dynamic correction: The online adaptive mechanism of the model is enabled, and the model weights are fine-tuned every 24 hours using the latest production data to adapt to seasonal viscosity changes; Control strategy: Multi-variable collaborative control is implemented to synchronously adjust the feed belt speed and the classifier hydrocyclone pressure.

[0059] Implementation results: The system successfully decoupled the interference of concentration and load, reduced the standard deviation of particle size distribution from 3.5% to 1.2%, improved the classification efficiency by 6%, and significantly reduced the over-grinding of qualified fine particles.

[0060] Example 3

[0061] In a gold mine's pre-cyanidation grinding operation, a small-sample cold start condition was encountered. This production line was a newly expanded project, lacking long-term historical data (only 3 months of data were available), and the ore properties were extremely unstable, making it difficult for conventional data-driven models to converge. Implementation process: Transfer learning: Utilizing the transfer learning function supported by this invention, model parameters pre-trained on a similar gold mine dataset were loaded as initial weights; Mechanism guidance: The weight coefficients of the mechanism constraint layer were significantly enhanced, forcing the model output to conform to physical conservation laws (such as mass balance and energy conservation), compensating for the lack of data; Human-machine collaboration: In the initial stage of operation, a "model recommendation + manual confirmation" mode was adopted, switching to fully automatic control after the model confidence exceeded 90%.

[0062] Implementation results: The model can be adapted and put into automatic operation in just 2 weeks, which shortens the debugging cycle by 80% compared with the traditional modeling method. The initial granularity control accuracy reaches 88% and quickly climbs to over 91%.

[0063] Comparative Example 1

[0064] Based on traditional PID feedback control (corresponding to the operating condition in Implementation Case 1), classic single-loop PID control is used, with particle size analysis values ​​as feedback to adjust the feed rate. However, due to a lag of approximately 2 hours in the analysis results in a severe delay in controller response.

[0065] Comparison results: Particle size qualification rate: 78.2%; Fluctuation: When the ore hardness changed abruptly, the system experienced violent oscillations, causing the mill to stop twice.

[0066] Conclusion: Traditional PID controllers cannot solve problems involving large time delays and nonlinearity, and have poor anti-interference capabilities.

[0067] Comparative Example 2

[0068] The soft measurement open-loop guidance based on support vector regression (SVR) (corresponding to the working condition of implementation case 2) uses the SVR algorithm to build a granular prediction model, which is only displayed on the operator's reference screen and does not perform automatic closed-loop control, and no mechanistic constraints are introduced.

[0069] Comparison results: Prediction accuracy (R²): 0.82; Generalization ability: When the slurry viscosity changes seasonally, the prediction error of the SVR model increases sharply, requiring manual re-collection of data to train the model, which takes about 3 days.

[0070] Conclusion: Shallow machine learning models have limited feature extraction capabilities, lack adaptive mechanisms, and the open-loop model cannot eliminate differences in human operation.

[0071] Comparative Example 3

[0072] Pure data-driven LSTM neural network control (without mechanistic constraints) (corresponding to implementation case 3) constructs a standard LSTM deep learning model for predictive control, but removes the "mechanistic constraint layer" in this invention, relying entirely on data fitting.

[0073] Comparison results: Small sample performance: With insufficient data (only 3 months), the model exhibits severe overfitting, with high accuracy on the training set but failure on the test set; Physical interpretability: Under certain extreme conditions, the model outputs control commands that violate physical common sense (such as negative flow and overpower operation), causing the system to trigger a safety interlock shutdown.

[0074] Conclusion: Pure black-box deep learning lacks physical boundary constraints, has poor robustness in small sample scenarios, and poses security risks; this invention effectively solves this problem by embedding mechanistic constraints.

[0075] Comparison of effects between comparative examples and overall embodiments - table

[0076] Indicators / Solutions Traditional PID control (Comparative Example 1) SVR Soft Measurement (Comparative Example 2) Pure data-driven LSTM (Comparative Example 3) General Embodiments of the Invention Particle size qualification rate 78.2% 81.5% (Reference value) 85.0% (Unstable) 92.5% Anti-interference capability weak middle middle powerful Small sample adaptability not applicable Difference Poor (prone to overfitting) Excellent (transfer learning and mechanistic constraints) Response lag time >2 hours Real-time prediction but no control Real-time control Real-time closed-loop control Physical security high high Low (may output outliers) High (mechanism constraint guarantee) Debugging cycle Short but ineffective Long duration (requires frequent retraining) Long (requires a large amount of data) Short (adaptive and transfer)

[0077] Comparison of the effects of each embodiment

[0078] Example number Operating conditions and background characteristics Core Implementation Strategy Summary of Implementation Results Example 1 High-hardness ore working conditions (a large copper mine, where the grade of the raw ore fluctuates greatly, making it prone to "bloating" or over-grinding). 1. Enhance the extraction of time-series characteristics of the lag relationship between "power and feed rate". 2. Introduce the crushing energy consumption mechanism formula as a regularization term for the loss function. The particle size qualification rate increased from 78% to 92.5%; unit energy consumption decreased by 4.8%; and downtime accidents caused by sudden changes in hardness were effectively avoided. Example 2 Multivariable strongly coupled operating conditions (a certain iron mine, where concentration and load have large interferences, and slurry viscosity varies with the seasons). 1. Add temperature and viscosity estimates as auxiliary inputs, and extract the correlation matrix using spatial convolution. 2. Enable an online adaptive mechanism to fine-tune the weights every 24 hours to adapt to seasonal changes. The standard deviation of particle size distribution decreased from 3.5% to 1.2%; the classification efficiency improved by 6%; and the concentration and load interference were successfully decoupled, reducing over-grinding. Example 3 Small sample cold start conditions (a gold mine, lacking historical data, with extremely unstable ore properties). 1. Employ transfer learning, loading pre-trained parameters similar to those used in gold mines. 2. Enhance the weights of the mechanistic constraint layer, forcing compliance with physical conservation laws. 3. Initially, use a "model recommendation and manual verification" model. The debugging cycle was shortened by 80% (only 2 weeks); the initial accuracy reached 88% and quickly rose to over 91%; the problem of difficult convergence with small samples was solved.

[0079] The above embodiments are merely preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A grinding particle size prediction and control method based on deep learning, characterized in that... This method includes real-time acquisition of grinding process operating parameters through a sensor network deployed in the grinding system; smoothing and filtering the operating parameters using a sliding window; establishing a deep spatiotemporal feature extraction network; embedding process mechanisms such as mass conservation, energy balance, and particle size dynamics into the network with differentiable constraints; constructing a particle size prediction network based on deep spatiotemporal features and constraints; using a sequential quadratic programming algorithm to calculate the optimal control quantity in real time; comparing the actual measured particle size data with the predicted values ​​to trigger an online update mechanism for the particle size prediction network; and continuously evaluating the error index of the particle size prediction network. This method can effectively improve the accuracy of grinding particle size prediction and control stability, enhance the system's generalization ability under complex operating conditions, and improve long-term operational reliability. Specifically, it includes the following process steps and conditions: Step 1. Data Acquisition and Preprocessing: Real-time acquisition of operating parameters such as feed rate, mill current, and bearing vibration through sensor network; noise removal using sliding window smoothing filter, outlier removal using 3 sigma criterion, and linear interpolation and time alignment for missing data. Step 2. Construct a deep spatiotemporal feature extraction network: Establish an architecture with a bidirectional long short-term memory network (Bi-LSTM) as the core and combined with a convolutional neural network (CNN); use the attention mechanism to extract the temporal dependence and spatial correlation features of multi-sensor data; Step 3. Embedding process mechanism constraints: Transform physical equations such as mass conservation, energy balance, and particle size dynamics into differentiable constraints; embed these constraints into the forward propagation process of the neural network using the Lagrange multiplier method to ensure that the model conforms to physical laws; Step 4. Construct a particle size prediction network: Based on the extracted spatiotemporal features and mechanistic constraints, construct a sequence-to-sequence (Seq2Seq) prediction model; output a particle size distribution curve covering the entire grinding cycle (e.g., the next 30 sampling points); Step 5. Calculate the optimal control quantity: With minimizing granularity deviation, minimizing energy consumption, and ensuring equipment stability as optimization objectives, the optimal control command is calculated in real time and output under the premise of satisfying load constraints using the Sequential Quadratic Programming (SQP) algorithm. Step 6. Trigger the online update mechanism: compare the actual measurement granularity with the predicted value in real time; when the deviation exceeds the preset threshold (e.g., 3%), automatically start the online update and use the stochastic gradient descent algorithm with momentum to adjust the network parameters to correct the error; Step 7. Performance Evaluation and Retraining: Continuously monitor the root mean square error of the model and the integral absolute error of the controller; When performance metrics degrade beyond a set limit (e.g., 25%), the model retraining process is automatically initiated to ensure continuous adaptation to changes in operating conditions. Step 8. Continuously evaluate error and initiate retraining: Continuously evaluate the root mean square error of the granular prediction network and the integral absolute error of the controller. When the performance index degrades by more than 25%, the granular prediction network retraining process is automatically initiated.

2. The method according to claim 1, characterized in that: The sensor network in step one includes at least an electromagnetic flowmeter, a vibration acceleration sensor, a power transmitter, and a laser particle size analyzer.

3. The method according to claim 1 or 2, characterized in that: The data preprocessing or operating parameters in step one include time alignment of multi-source data and the use of a dynamic time warping algorithm to eliminate minor offsets in the timestamps collected by different sensors.

4. The method according to claim 1, characterized in that: The differentiable constraints in step three include at least the crushing rate function, the classification efficiency curve, and the material balance equation. The crushing rate function adopts the first-order crushing kinetic model, the classification efficiency curve adopts the Whiten model, and the material balance equation involves the feeding, discharging, and circulating load.

5. The method according to claim 1, characterized in that: The prediction network in step four uses an encoder and a decoder.

6. The method according to claim 1 or 5, characterized in that: The encoder in step four consists of a four-layer bidirectional long short-term memory network, with 64, 128, 256, and 128 hidden units in each layer, respectively.

7. The method according to claim 1 or 5, characterized in that: The fourth decoder consists of three layers of gated recurrent units with 128, 64 and 32 hidden units respectively. Its attention mechanism adopts an additive attention model with a dimension of 64.

8. The method according to claim 1, characterized in that: The granular prediction network in step six adopts a sequence-to-sequence architecture.

9. The method according to claim 1 or 8, characterized in that: The optimization objective function of the granularity prediction network predictive controller in step six comprehensively considers the granularity qualification rate, energy consumption index and equipment operation stability, with weight coefficients of 0.6, 0.3 and 0.1, respectively.

Images