Intelligent mineral flotation control system and method based on multi-modal perception

Abstract

This invention discloses an intelligent mineral flotation control system and method based on multimodal perception, belonging to the field of mineral processing automation technology. The system includes a flotation cell multimodal perception module, an edge computing processing unit, an intelligent decision control module, and an actuator feedback module. By integrating visual, electrochemical impedance spectroscopy, and bubble dynamics characteristics, a digital twin model of the flotation process is constructed. A deep reinforcement learning algorithm is used to optimize reagent addition strategies and aeration parameters in real time, achieving closed-loop intelligent control of the flotation process. This invention solves the technical problems of traditional flotation control, such as reliance on manual experience, slow response, and large fluctuations in recovery rate. It significantly improves concentrate grade and recovery rate, reduces reagent and energy consumption, enhances system adaptability and operational stability, and is suitable for the flotation separation of non-ferrous metals, ferrous metals, and non-metallic minerals.

Description

Technical Field

[0001] This invention relates to the field of mineral processing automation technology, and in particular to an intelligent mineral flotation control system and method based on multimodal sensing. Background Technology

[0002] Mineral flotation is a mainstream mineral processing technology that separates minerals by utilizing the differences in their surface physicochemical properties and the selective carrying of target mineral particles by bubbles. It is widely used in the separation of non-ferrous metals, ferrous metals, and non-metallic minerals. The flotation process involves complex three-phase flow of gas, liquid, and solid, interfacial electrochemical reactions, reagent adsorption and desorption, bubble mineralization, and other multi-physical field coupling behaviors. It is characterized by strong nonlinearity, large time lag, strong coupling of multiple variables, and significant time-varying disturbances, making it a typical complex industrial process control object.

[0003] Traditional flotation control relies primarily on manual observation of froth conditions or on analog measurements such as level, concentration, and flow rate. This limited sensing methods and insufficient information dimensions make it difficult to capture crucial microscopic states that determine flotation efficiency, such as bubble size distribution, mineral surface electrochemical activity, and flow dynamics. Existing PLC-based PID control or rule-based expert systems generally suffer from slow response, low adjustment accuracy, and weak anti-interference capabilities. When faced with disturbances such as fluctuations in ore properties, changes in feed rate, and equipment aging, they cannot quickly match optimal control parameters, easily leading to large fluctuations in concentrate grade and recovery rate, reagent waste, and high energy consumption.

[0004] In recent years, intelligent technologies such as machine vision, digital twins, and deep reinforcement learning have developed rapidly in the field of industrial process control, but they are still in the stage of localized application in mineral flotation scenarios. Existing technologies have failed to achieve deep integration of multimodal information such as vision, electrochemistry, and kinetics, lack a collaborative architecture of real-time edge computing at the field level and intelligent decision-making in the cloud, make it difficult to build a high-precision digital twin model of flotation conditions, and fail to achieve multi-objective intelligent optimization control with objectives such as grade, recovery rate, reagent consumption, and energy consumption. Consequently, they cannot meet the actual needs of modern mineral processing plants for efficient, stable, low-consumption, and intelligent upgrades. Summary of the Invention

[0005] The purpose of this invention is to propose an intelligent mineral flotation control system and method based on multimodal perception, which aims to solve the problems of single perception, delayed decision-making, poor adaptability, and difficulty in coordinating and optimizing multiple objectives in existing flotation control, so as to achieve efficient, stable, and low-consumption closed-loop intelligent control of the flotation process.

[0006] To achieve the above objectives, this invention proposes an intelligent mineral flotation control system based on multimodal perception, comprising: a flotation cell multimodal perception module, an edge computing processing unit, an intelligent decision control module, and an actuator feedback module; The multimodal sensing module of the flotation cell is used to simultaneously collect visual image information, electrochemical impedance spectroscopy information and bubble surface dynamics information of the slurry in the flotation cell. The edge computing processing unit is communicatively connected to the multimodal sensing module of the flotation cell, and includes a multimodal data fusion unit, a working condition identification unit, and a digital twin construction unit, which are used to perform real-time preprocessing and feature extraction of multimodal data to construct a local digital twin model of the flotation working condition. The multimodal data fusion unit employs a feature fusion network based on a cross-modal Transformer attention mechanism to map visual features, electrochemical features, and dynamic features to a unified high-dimensional feature space. The feature fusion network includes three modality-specific encoders and one shared decoder. The operating condition identification unit classifies and identifies flotation operating conditions based on support vector machine (SVM) or random forest algorithm, including normal operating conditions, over-aeration operating conditions, insufficient reagent operating conditions, and poor mineralization operating conditions. The digital twin building unit establishes a real-time digital mirror of the gas-liquid-solid three-phase flow in the flotation cell based on computational fluid dynamics and discrete element coupled simulation. The intelligent decision control module, based on the digital twin model, uses a deep reinforcement learning algorithm to generate the optimal control strategy for the dosage of the agent, the amount of aeration, and the stirring speed. The actuator feedback module includes a variable frequency chemical pump, an electrically controlled gas charging valve group, and a variable frequency stirrer, which are used to execute control commands and feed back the actual operating parameters to the intelligent decision control module to form a closed-loop control.

[0007] Preferably, the flotation cell multimodal sensing module includes an industrial-grade underwater high-definition camera array, an electrochemical impedance spectroscopy sensor array, and a high-frequency pressure sensor network; The industrial-grade underwater high-definition camera array is installed 30-50cm below the surface of the flotation cell to acquire image features of bubble size distribution and bubble-loaded mineral particles. The electrochemical impedance spectroscopy sensor array is distributed at different depths in the flotation cell and is used to detect the redox potential, ion concentration and electrochemical activity of the mineral surface in the slurry. The high-frequency pressure sensor network is attached to the bottom and sidewalls of the flotation cell to collect the dynamic parameters of bubble rise and the distribution of slurry flow field.

[0008] Preferably, the operating condition identification unit classifies and identifies the flotation operating conditions based on the support vector machine (SVM) or random forest algorithm. Classification and recognition are achieved by inputting multimodal fusion feature vectors into the classifier, with the vector formula as follows: ; The SVM multi-class decision function is as follows: ; The decision function for a random forest is as follows: ; in, This is a multimodal fusion feature vector. As a visual feature, It is characterized by electrochemical properties. As a dynamic characteristic, The dimension of the fused feature space. For the output working condition category labels, Index for working condition categories, For the first j Support vector set for similar working conditions For the first j Lagrange multipliers for support vector machines, For the first j Bias terms for support vector machines, For kernel function, For random forest ensemble classification functions, For the total number of decision trees, For the first decision trees x The predicted output, For indicator functions, According to the working condition category, This is a set of operating condition categories.

[0009] Preferably, the digital twin building block, based on computational fluid dynamics and discrete element coupled simulation, establishes a real-time digital mirror of the gas-liquid-solid three-phase flow within the flotation cell, specifically as follows: The real-time mirror evolution of the digital twin model is achieved through a state update operator, as shown in the following formula: ; in, for t The digital twin state vector at any given time. for t Control the input vector at all times. For the parameter set of the digital twin model, To solve operators that couple computational fluid dynamics with discrete element method, This represents the time step of the digital twin model.

[0010] Preferably, the intelligent decision control module includes a deep reinforcement learning unit, a multi-objective optimization unit, and an adaptive adjustment unit; The deep reinforcement learning unit uses the Proximal Policy Optimization (PPO) algorithm or the Actor-Critic (SAC) algorithm, with concentrate grade, recovery rate and reagent consumption as reward functions, to train the agent to learn the optimal control strategy. The near-end policy optimization algorithm updates the policy network parameters by replacing the objective function with truncation, as shown in the following formula: ; in, For policy network parameters, Importance sampling ratio, For a moment t The estimate of the dominance function, To truncate hyperparameters, This is a truncation function. For the expectation operator of empirical sampling; ; in, for t Constantly control the action vector. for t Real-time flotation operation status digital twin. For the policy network in state Down Output Action The probability, For old strategy network parameters; The multi-objective optimization unit uses the non-dominated sorting genetic algorithm NSGA-II to seek the Pareto optimal solution among multiple objectives such as grade, recovery rate, energy consumption, and reagent cost. The non-dominated sorting genetic algorithm ranks individuals in the population based on Pareto dominance and maintains population diversity based on crowding distance, as shown in the following formula: ; ; in, , They are respectively the first in the population i , j The control strategy parameter vector corresponding to each individual. It is a Pareto dominance symbol. For the first m 0 objective functions and , For the first n 0 objective functions and , Let the objective function be the concentrate grade. Let the recovery rate be the objective function. Let the energy consumption objective function be... Let the objective function be the drug cost. For the firsti The crowding distance of each individual The first i Individual in the first m The function values ​​of neighboring individuals on the objective function. , The first m The maximum and minimum values ​​of the objective function; The adaptive adjustment unit dynamically adjusts the exploration-utilization balance parameters of the reinforcement learning strategy network based on fluctuations in ore properties and changes in the grade of incoming ore. Explore - Adaptively adjust the ore grade fluctuation based on balance parameters. The adjustment formula is as follows: ; ; in, for t The amount of fluctuation in ore grade at any given time. for t The ore grade is measured at regular intervals. This represents the historical average grade of the incoming ore. The standard deviation of the ore grade. for t Continuous exploration - utilizing balance parameters, To minimize the exploration rate, To maximize the exploration rate, The attenuation coefficient is... It is a natural exponential function; The entropy regularization of the reinforcement learning policy network is dynamically adjusted based on the following formula: ; in, for t+ Time-1 policy network entropy regularization coefficient, for t The time-space policy network entropy regularization coefficient, This is the ore fluctuation sensitivity coefficient; when the fluctuation of the incoming ore grade increases... Increase to enhance strategy exploration capabilities. Synchronous increase enhances the randomness of the strategy, enabling the agent to quickly adapt to sudden changes in ore properties.

[0011] Preferably, the actuator feedback module includes a variable frequency drug pump group, an electrically controlled gas charging valve group, and a variable frequency stirrer; The variable frequency chemical pump set includes a collector pump, a foaming agent pump, and a conditioning agent pump. Each pump is equipped with a mass flow meter and a servo motor to achieve closed-loop precise control of the chemical flow rate. The electronically controlled inflation valve assembly uses a fuzzy PID control algorithm to adjust the inflation volume and inflation valve opening based on the feedback of bubble size distribution. The fuzzy PID control algorithm uses the deviation between the target bubble diameter and the actual average bubble diameter. As input, the adaptive proportional coefficient is obtained by tuning the PID parameters through fuzzy inference, as shown in the following formula: ; in, The ratio after adaptive tuning via fuzzy inference. This is the initial value of the proportionality coefficient. The total number of fuzzy rules, For the first j The membership degree of a fuzzy rule. For the first j The proportional coefficient correction amount corresponding to the fuzzy rule. The target average bubble diameter, for t The actual average diameter of the bubble detected by the vision sensor at any given time; Output inflation volume control commands and adjust the inflation valve opening according to the inflation volume mapping, as shown in the following formula: ; ; in, for t The inflation volume is controlled at all times. The integral after adaptive tuning via fuzzy inference. These are the differential coefficients after adaptive tuning via fuzzy inference. for t The opening degree of the inflation valve at all times. This is the reference opening degree for the inflation valve. This is the conversion coefficient between the inflation valve opening and the inflation volume. The variable frequency agitator adaptively adjusts the stirring speed according to the slurry concentration and particle size distribution to maintain optimal suspension; wherein, the adaptive adjustment is calculated based on real-time detected values ​​of slurry concentration and particle size, as shown in the following formula: ; in, for t The target speed of the variable frequency stirrer at all times. The reference stirring speed is... This is the slurry concentration compensation coefficient. This is the particle size compensation coefficient. for t Real-time slurry mass concentration test value, This is a reference value for slurry concentration. fort The median particle size of the slurry at any given time was measured. This is the median particle size reference value; When the pulp concentration increases or the particle size becomes coarser... It automatically rises to maintain particle suspension; when the concentration decreases or the particle size becomes finer... Automatically reduce energy consumption and wear.

[0012] This invention also provides an intelligent mineral flotation control method based on multimodal sensing, the steps of which are as follows: Step S1: Simultaneously collect multi-source heterogeneous data in the flotation cell using a visual sensor, electrochemical sensor, and pressure sensor array with a sampling period of 100-500ms. Step S2: On the edge computing nodes deployed at the flotation site, the collected data is denoised, aligned, and feature extracted to construct a real-time digital twin model of the flotation process. Step S3: Based on the deep reinforcement learning algorithm, with the current working condition digital twin state as input, output the optimal control commands for the amount of agent added, the amount of aeration, and the stirring speed; Step S4: Control the actuator to implement control commands and feed back the actual operating parameters to the intelligent decision-making module to form a closed-loop control circuit; Step S5: Based on the deviation between actual production indicators and predicted indicators, update the reinforcement learning strategy network parameters online to achieve continuous evolution of the control strategy.

[0013] Preferably, in step S2, feature extraction includes: Visual feature extraction: An improved YOLOv8 instance segmentation network is used to identify and statistically analyze bubble size distribution, bubble loading rate, and bubble aggregation degree; Electrochemical feature extraction: The solution resistance, double-layer capacitance, and charge transfer resistance of the slurry are extracted by fitting the equivalent circuit of electrochemical impedance spectroscopy. Dynamic feature extraction: Based on wavelet transform analysis of pressure sensor signals, bubble rising velocity, slurry turbulence intensity, and bubble-particle collision frequency are extracted.

[0014] Preferably, in step S3, the reward function of the deep reinforcement learning algorithm is designed as follows: ; Where R is the instant reward value, for t Constant concentration of mineral grade The target value for concentrate grade, For recovery rate, The target value for recovery rate, For the cost of the medicine, This serves as a benchmark for drug costs. For energy consumption, This is the energy consumption baseline value. , , , These are the weighting coefficients.

[0015] Preferably, in step S5, the online update of the reinforcement learning strategy network parameters is specifically performed as follows: an experience replay mechanism and a priority sampling strategy are adopted to prioritize learning the control experience when flotation conditions change abruptly, thereby accelerating strategy convergence, establishing a knowledge graph of the flotation process, and integrating mineralogical characteristics, reagent action mechanisms, and equipment operating parameters to provide prior knowledge constraints for deep reinforcement learning and avoid dangerous conditions during the exploration process.

[0016] Therefore, this invention proposes an intelligent mineral flotation control system and method based on multimodal sensing, the beneficial effects of which are as follows: (1) By integrating visual, electrochemical impedance spectroscopy and bubble dynamics multimodal information, this invention can accurately characterize bubble mineralization, mineral surface activity and flow field state, breaking through the information limitations of traditional single parameter detection.

[0017] (2) The present invention uses edge computing and deep reinforcement learning to achieve second-level decision-making, which is significantly better than traditional manual and PID control, effectively suppressing operating condition fluctuations and improving system stability.

[0018] (3) This invention takes grade, recovery rate, reagent consumption and energy consumption as multiple optimization objectives, avoiding the overall benefit loss caused by optimizing a single indicator, and taking into account both production indicators and economic benefits.

[0019] (4) The online learning mechanism in this invention enables the system to automatically adapt to time-varying factors such as fluctuations in ore properties, equipment aging, and changes in ambient temperature, and has the ability to learn online and continuously evolve, with no degradation in long-term operation performance.

[0020] (5) The present invention relies on the flotation knowledge graph to provide prior constraints, thereby avoiding dangerous working conditions such as over-gasification and excessive chemical addition, and ensuring continuous and safe production.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall architecture of an intelligent mineral flotation control system based on multimodal perception according to the present invention. Figure 2 This is a schematic diagram showing the deployment of the multimodal sensing module in the flotation cell of the present invention; Figure 3 This is a flowchart of the data processing of the edge computing processing unit in this invention; Figure 4 This is a flowchart of an intelligent mineral flotation control method based on multimodal sensing according to the present invention; Figure 5 The above are comparison diagrams of copper ore flotation effects in embodiments of the present invention; wherein, (a) is a comparison diagram of concentrate grade stability, (b) is a comparison diagram of recovery rate, (c) is a comparison diagram of reagent consumption, and (d) is a comparison diagram of overall performance. Detailed Implementation

[0023] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0025] Example 1 like Figures 1-2 As shown, the present invention provides an intelligent mineral flotation control system based on multimodal perception, comprising: a flotation cell multimodal perception module, an edge computing processing unit, an intelligent decision control module, and an actuator feedback module; The flotation cell multimodal sensing module includes an industrial-grade underwater high-definition camera array, an electrochemical impedance spectroscopy sensor array (EIS), and a high-frequency pressure sensor network, which are used to simultaneously acquire visual image information, electrochemical impedance spectroscopy information, and bubble surface dynamics information of the slurry in the flotation cell. An industrial-grade underwater high-definition camera array is installed 30-50cm below the surface of the flotation cell to acquire image features of bubble size distribution and bubble-loaded mineral particles. An electrochemical impedance spectroscopy (EIS) sensor array is distributed at different depths in the flotation cell. It uses an offline calibration combined with a simplified online high-frequency measurement method to detect the redox potential, ion concentration, and electrochemical activity of the mineral surface in the slurry. A high-frequency pressure sensor network is attached to the bottom and sidewalls of the flotation cell to collect parameters of bubble rising dynamics and slurry flow field distribution. The edge computing processing unit is communicatively connected to the multimodal sensing module of the flotation cell, including a multimodal data fusion unit, a working condition identification unit, and a digital twin construction unit. It is used to perform real-time preprocessing and feature extraction of multimodal data and construct a local digital twin model of the flotation working condition. The edge computing processing unit synchronously receives data from three types of sensors—visual, electrochemical, and kinetic—via Gigabit Ethernet using the PTP protocol, with a sampling period of 100-500ms. The multimodal data fusion unit employs an attention-based feature fusion network to map visual features, electrochemical features, and kinetic features to a unified high-dimensional feature space. The operating condition identification unit, based on the support vector machine (SVM) or random forest algorithm, classifies and identifies the flotation operating conditions, including normal operating conditions, over-aerated operating conditions, insufficient reagent operating conditions, and poor mineralization operating conditions. The digital twin building block, based on computational fluid dynamics and discrete element coupled simulation, establishes a real-time digital mirror of the gas-liquid-solid three-phase flow in the flotation cell; The intelligent decision control module includes a deep reinforcement learning unit, a multi-objective optimization unit, and an adaptive adjustment unit. Based on a digital twin model, it uses a deep reinforcement learning algorithm to generate the optimal control strategy for the dosage of the agent, the aeration rate, and the stirring speed. The deep reinforcement learning unit uses the proximal policy optimization algorithm or the actor-critic algorithm (SAC) with concentrate grade, recovery rate and reagent consumption as reward functions to train the agent to learn the optimal control policy. The multi-objective optimization unit uses the non-dominated sorting genetic algorithm NSGA-II to seek the Pareto optimal solution among multiple objectives such as grade, recovery rate, energy consumption, and reagent cost; The adaptive adjustment unit dynamically adjusts the exploration-utilization balance parameter of the reinforcement learning strategy network based on fluctuations in ore properties and changes in the grade of incoming ore.

[0026] The actuator feedback module includes a variable frequency chemical pump, an electrically controlled gas charging valve group, and a variable frequency agitator. It is used to execute control commands and feed back the actual operating parameters to the intelligent decision control module to form a closed-loop control.

[0027] The variable frequency chemical pump set includes a collector pump, a foaming agent pump, and a conditioning agent pump. Each pump is equipped with a mass flow meter and a servo motor to achieve closed-loop precise control of the chemical flow rate. The electronically controlled inflation valve assembly uses a fuzzy PID control algorithm to adjust the inflation volume and inflation valve opening based on the feedback of bubble size distribution. The variable frequency agitator adaptively adjusts the stirring speed according to the slurry concentration and particle size distribution to maintain the optimal suspension state.

[0028] Example 2 like Figure 4As shown, the present invention also provides an intelligent mineral flotation control method based on multimodal sensing, the specific steps of which are as follows: Step S1: Simultaneously collect multi-source heterogeneous data in the flotation cell using a visual sensor, electrochemical sensor, and pressure sensor array with a sampling period of 100-500ms. Step S2: On the edge computing nodes deployed at the flotation site, the collected data is denoised, aligned, and feature extracted to construct a real-time digital twin model of the flotation process. Feature extraction includes: Visual feature extraction: An improved YOLOv8 instance segmentation network is used to identify and statistically analyze bubble size distribution, bubble loading rate, and bubble aggregation degree; Electrochemical feature extraction: The solution resistance, double-layer capacitance, and charge transfer resistance of the slurry are extracted by fitting the equivalent circuit of electrochemical impedance spectroscopy. Dynamic feature extraction: Based on wavelet transform analysis of pressure sensor signals, bubble rising velocity, slurry turbulence intensity, and bubble-particle collision frequency are extracted.

[0029] Step S3: Based on the deep reinforcement learning algorithm, with the current working condition digital twin state as input, output the optimal control commands for the amount of agent added, the amount of aeration, and the stirring speed; The reward function design for deep reinforcement learning algorithms is as follows: ; Where R is the instant reward value, for t Constant concentration of mineral grade The target value for concentrate grade, For recovery rate, The target value for recovery rate, For the cost of the medicine, This serves as a benchmark for drug costs. For energy consumption, This is the energy consumption baseline value. , , , These are the weighting coefficients.

[0030] Step S4: Control the actuator to implement control commands and feed back the actual operating parameters to the intelligent decision-making module to form a closed-loop control circuit; Step S5: Based on the deviation between actual production indicators and predicted indicators, update the reinforcement learning strategy network parameters online to achieve continuous evolution of the control strategy. Specifically, the online update of the reinforcement learning strategy network parameters involves: adopting an experience replay mechanism and a priority sampling strategy to prioritize learning the control experience when flotation conditions undergo abrupt changes, thereby accelerating strategy convergence; establishing a knowledge graph of the flotation process, integrating mineralogical characteristics, reagent action mechanisms, and equipment operating parameters to provide prior knowledge constraints for deep reinforcement learning and avoid dangerous conditions during the exploration process.

[0031] The technical solution of the present invention will be further illustrated below through specific implementation examples.

[0032] Example 3 like Figure 3 As shown in the figure, this embodiment uses the system of the present invention to intelligently transform the roughing operation of a copper mine beneficiation plant.

[0033] System Deployment: 1. In the coarse separation tank (volume 40m³) 3 Three underwater high-definition cameras (resolution 2048×1536, frame rate 30fps) were installed, distributed at the midpoint of the long side and the two short sides of the tank. 2. Install 12 electrochemical impedance spectroscopy sensors, one every 50cm along the depth of the tank, with a measurement frequency range of 10mHz-100kHz. 3. Install 8 high-frequency pressure sensors (sampling frequency 1kHz), 4 on the bottom and 4 on the side wall; 4. The edge computing nodes use NVIDIA Jetson AGX Xavier industrial computers and are deployed in explosion-proof cabinets at the edge of the slot; 5. The intelligent decision-making module is deployed on the server in the central control room of the ore dressing plant, equipped with an NVIDIA RTX A6000 GPU; 6. The actuators include 3 variable frequency chemical pumps (1 each of butyl xanthate collector, MIBC foaming agent, and lime slurry modifier), an electrically controlled air filling valve assembly (including an air filling valve opening adjustment mechanism), and a variable frequency agitator (55kW power).

[0034] Optimal control strategy training: 1. Collect the plant's operational data for the past three months, including manual operation records, test results, and equipment operation logs; 2. Construct an initial digital twin model based on historical data; 3. Offline pre-training is performed based on the SAC algorithm, and the reward function weights are set to... =0.3、 =0.4、 =0.2、 =0.1, prioritizing recovery rate; 4. After pre-training, online fine-tuning will be conducted for two weeks, during which key operations will be manually supervised.

[0035] like Figure 5 As shown, after the system was put into operation, the grade of the roughing concentrate increased from 18.5% to 20.2%, and the recovery rate increased from 82.3% to 87.6%; the consumption of butyl xanthate decreased from 120g / t to 95g / t, and the consumption of MIBC decreased from 35g / t to 28g / t; the system response time was shortened from 5-10 minutes of manual operation to less than 30 seconds; when the ore properties fluctuated (the copper grade of the incoming ore fluctuated from 0.8% to 1.2%), the system automatically adjusted the reagent regime, and the fluctuation range of the concentrate grade narrowed from ±2.5% to ±0.8%.

[0036] Example 4 This embodiment uses the system of the present invention to intelligently transform the anion reverse flotation operation of an iron ore beneficiation plant.

[0037] In this embodiment, sensors for pulp pH and temperature are added to address the characteristics of reverse flotation; simultaneously, to prioritize ensuring the grade of iron concentrate, the reward function weights are adjusted to... =0.5、 =0.2、 =0.2、 =0.1, and the digital twin model considers the effect of temperature on the drug adsorption kinetics, while the deployment of other systems remains consistent with Example 3.

[0038] After the system was put into operation, the iron concentrate grade increased from 65.2% to 67.8%, reaching the first-grade standard; the tailings iron grade decreased from 18.5% to 14.2%, reducing metal loss; the consumption of starch inhibitor decreased by 20%, and the consumption of anionic collector decreased by 18%.

[0039] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0040] Therefore, this invention provides an intelligent mineral flotation control system and method based on multimodal perception. Through the deep integration of multimodal perception, edge computing, digital twins and deep reinforcement learning, it realizes full-dimensional perception, real-time intelligent decision-making and closed-loop precise control of the flotation process, significantly improves concentrate grade and recovery rate, reduces reagent consumption and energy consumption, enhances system adaptability and operational stability, and brings outstanding economic and safety benefits.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-modal perception based intelligent mineral flotation control system, characterized in that, include: The flotation cell includes a multimodal sensing module, an edge computing processing unit, an intelligent decision control module, and an actuator feedback module. The multimodal sensing module of the flotation cell is used to simultaneously collect visual image information, electrochemical impedance spectroscopy information and bubble surface dynamics information of the slurry in the flotation cell. The edge computing processing unit is communicatively connected to the multimodal sensing module of the flotation cell, and includes a multimodal data fusion unit, a working condition identification unit, and a digital twin construction unit, which are used to perform real-time preprocessing and feature extraction of multimodal data to construct a local digital twin model of the flotation working condition. The multimodal data fusion unit employs a feature fusion network based on a cross-modal Transformer attention mechanism to map visual features, electrochemical features, and dynamic features to a unified high-dimensional feature space. The feature fusion network includes three modality-specific encoders and one shared decoder. The operating condition identification unit classifies and identifies flotation operating conditions based on support vector machine (SVM) or random forest algorithm, including normal operating conditions, over-aeration operating conditions, insufficient reagent operating conditions, and poor mineralization operating conditions. The digital twin building unit establishes a real-time digital mirror of the gas-liquid-solid three-phase flow in the flotation cell based on computational fluid dynamics and discrete element coupled simulation. The intelligent decision control module, based on the digital twin model, uses a deep reinforcement learning algorithm to generate the optimal control strategy for the dosage of the agent, the amount of aeration, and the stirring speed. The actuator feedback module includes a variable frequency chemical pump, an electrically controlled gas filling valve group, and a variable frequency stirrer, which are used to execute control commands and feed back the actual operating parameters to the intelligent decision control module to form a closed-loop control.

2. The intelligent mineral flotation control system based on multi-modal perception as claimed in claim 1, wherein, The flotation cell multimodal sensing module includes an industrial-grade underwater high-definition camera array, an electrochemical impedance spectroscopy sensor array, and a high-frequency pressure sensor network. The industrial-grade underwater high-definition camera array is installed 30-50cm below the surface of the flotation cell to acquire image features of bubble size distribution and bubble-loaded mineral particles. The electrochemical impedance spectroscopy sensor array is distributed at different depths in the flotation cell and is used to detect the redox potential, ion concentration and electrochemical activity of the mineral surface in the slurry. The high-frequency pressure sensor network is attached to the bottom and sidewalls of the flotation cell to collect the dynamic parameters of bubble rise and the distribution of slurry flow field.

3. The intelligent mineral flotation control system based on multi-modal perception as claimed in claim 1, wherein, The operating condition identification unit classifies and identifies flotation operating conditions based on support vector machine (SVM) or random forest algorithms. Classification and recognition are achieved by inputting multimodal fusion feature vectors into the classifier, with the vector formula as follows: ； The SVM multi-class decision function is as follows: ； The decision function for a random forest is as follows: ； in, This is a multimodal fusion feature vector. As a visual feature, It is characterized by electrochemical properties. As a dynamic characteristic, The dimension of the fused feature space. For the output working condition category labels, For working condition category index, For the first j Support vector set for similar working conditions For the first j Lagrange multipliers for support vector machines, For the first j Bias terms for support vector machines, For kernel function, For random forest ensemble classification functions, For the total number of decision trees, For the first decision trees x The predicted output, For indicator functions, According to the working condition category, This is a set of operating condition categories.

4. The intelligent mineral flotation control system based on multi-modal perception as claimed in claim 1, wherein, The digital twin building block, based on coupled computational fluid dynamics and discrete element method (DEM) simulation, establishes a real-time digital mirror of the gas-liquid-solid three-phase flow within the flotation cell, specifically as follows: The real-time mirror evolution of the digital twin model is achieved through a state update operator, as shown in the following formula: ； in, for t The digital twin state vector at any given time. for t Control the input vector at all times. For the parameter set of the digital twin model, To solve operators that couple computational fluid dynamics with discrete element method, This represents the time step of the digital twin model.

5. The intelligent mineral flotation control system based on multi-modal perception as claimed in claim 1, wherein, The intelligent decision control module includes a deep reinforcement learning unit, a multi-objective optimization unit, and an adaptive adjustment unit. The deep reinforcement learning unit uses the Proximal Policy Optimization (PPO) algorithm or the Actor-Critic (SAC) algorithm, with concentrate grade, recovery rate and reagent consumption as reward functions, to train the agent to learn the optimal control strategy. The near-end policy optimization algorithm updates the policy network parameters by replacing the objective function with truncation, as shown in the following formula: ； in, For policy network parameters, The importance sampling ratio, For a moment t The estimate of the dominance function, To truncate hyperparameters, This is a truncation function. For the expectation operator of empirical sampling; ； in, for t Constantly control the action vector. for t Real-time flotation operation status digital twin. For the policy network in state Down Output Action The probability, For old strategy network parameters; The multi-objective optimization unit uses the non-dominated sorting genetic algorithm NSGA-II to seek the Pareto optimal solution among multiple objectives such as grade, recovery rate, energy consumption, and reagent cost. The non-dominated sorting genetic algorithm ranks individuals in the population based on Pareto dominance and maintains population diversity based on crowding distance, as shown in the following formula: ； ； in, , They are respectively the first in the population i , j The control strategy parameter vector corresponding to each individual. It is a Pareto dominance symbol. For the first m 0 objective functions and , For the first n 0 objective functions and , Let the objective function be the concentrate grade. Let the recovery rate be the objective function. Let the energy consumption objective function be... Let the objective function be the drug cost. For the first i The crowding distance of each individual The first i Individual in the first m The function values ​​of neighboring individuals on the objective function. , The first m The maximum and minimum values ​​of the objective function; The adaptive adjustment unit dynamically adjusts the exploration-utilization balance parameters of the reinforcement learning strategy network based on fluctuations in ore properties and changes in the grade of incoming ore. Explore - Adaptively adjust the ore grade fluctuation based on balance parameters. The adjustment formula is as follows: ； ； in, for t The amount of fluctuation in ore grade at any given time. for t The ore grade is measured at regular intervals. This represents the historical average grade of the incoming ore. The standard deviation of the ore grade. for t Exploring constantly - utilizing balance parameters, To minimize the exploration rate, To maximize the exploration rate, The attenuation coefficient is... It is a natural exponential function; The entropy regularization of the reinforcement learning policy network is dynamically adjusted based on the following formula: ； in, for t+ Time-1 policy network entropy regularization coefficient, for t The time-space policy network entropy regularization coefficient, This is the ore fluctuation sensitivity coefficient; when the fluctuation of the incoming ore grade increases... Increase to enhance strategy exploration capabilities. Synchronous increase enhances the randomness of the strategy, enabling the agent to quickly adapt to sudden changes in ore properties.

6. The intelligent mineral flotation control system based on multi-modal perception as claimed in claim 1, wherein, The actuator feedback module includes a variable frequency drug pump group, an electrically controlled gas charging valve group, and a variable frequency stirrer; The variable frequency chemical pump set includes a collector pump, a foaming agent pump, and a conditioning agent pump. Each pump is equipped with a mass flow meter and a servo motor to achieve closed-loop precise control of the chemical flow rate. The electronically controlled inflation valve assembly uses a fuzzy PID control algorithm to adjust the inflation volume and inflation valve opening based on the feedback of bubble size distribution. The fuzzy PID control algorithm takes the deviation between the target bubble diameter and the actual average bubble diameter as input As input, the PID parameters are set by fuzzy reasoning to obtain an adaptive proportional coefficient, as follows: ； in, The ratio after adaptive tuning via fuzzy inference. This is the initial value of the proportionality coefficient. The total number of fuzzy rules, For the first j The membership degree of a fuzzy rule. For the first j The proportional coefficient correction amount corresponding to the fuzzy rule. The target average bubble diameter, for t The actual average diameter of the bubble detected by the vision sensor at any given time; Output inflation volume control commands and adjust the inflation valve opening according to the inflation volume mapping, as shown in the following formula: ； ； in, for t The inflation volume is controlled at all times. The integral after adaptive tuning via fuzzy inference. These are the differential coefficients after adaptive tuning via fuzzy inference. for t The opening degree of the inflation valve at all times. This is the reference opening degree of the inflation valve. This is the conversion coefficient between the inflation valve opening and the inflation volume. The variable frequency agitator adaptively adjusts the stirring speed according to the slurry concentration and particle size distribution to maintain optimal suspension; wherein, the adaptive adjustment is calculated based on real-time detected values ​​of slurry concentration and particle size, as shown in the following formula: ； in, for t The target speed of the variable frequency stirrer at all times. The reference stirring speed is... This is the slurry concentration compensation coefficient. This is the particle size compensation coefficient. for t Real-time slurry mass concentration test value, This is a reference value for slurry concentration. for t The median particle size of the slurry at any given time was measured. This is the median particle size reference value; When the pulp concentration is increased or the particle size is coarser, automatically increased to maintain the particles in suspension, automatically decreased to reduce energy consumption and overgrinding.

7. A method of intelligent mineral flotation control based on multi-modal perception, characterized in that, The steps are as follows: Step S1: Simultaneously collect multi-source heterogeneous data in the flotation cell using a visual sensor, electrochemical sensor, and pressure sensor array with a sampling period of 100-500ms. Step S2: On the edge computing nodes deployed at the flotation site, the collected data is denoised, aligned, and feature extracted to construct a real-time digital twin model of the flotation process. Step S3: Based on the deep reinforcement learning algorithm, with the current working condition digital twin state as input, output the optimal control commands for the amount of agent added, the amount of aeration, and the stirring speed; Step S4: Control the actuator to implement control commands and feed back the actual operating parameters to the intelligent decision-making module to form a closed-loop control circuit; Step S5: Based on the deviation between actual production indicators and predicted indicators, update the reinforcement learning strategy network parameters online to achieve continuous evolution of the control strategy.

8. A multi-modal perception based intelligent mineral flotation control method as claimed in claim 7, wherein, In step S2, feature extraction includes: Visual feature extraction: An improved YOLOv8 instance segmentation network is used to identify and statistically analyze bubble size distribution, bubble loading rate, and bubble aggregation degree; Electrochemical feature extraction: The solution resistance, double-layer capacitance, and charge transfer resistance of the slurry are extracted by fitting the equivalent circuit of electrochemical impedance spectroscopy. Dynamic feature extraction: Based on wavelet transform analysis of pressure sensor signals, bubble rising velocity, slurry turbulence intensity, and bubble-particle collision frequency are extracted.

9. The intelligent mineral flotation control method based on multi-modal perception as claimed in claim 7, wherein, In step S3, the reward function of the deep reinforcement learning algorithm is designed as follows: ； Where R is the instant reward value, for t Constant concentration of mineral grade The target value for concentrate grade, For recovery rate, The target value for recovery rate, For the cost of the medicine, This serves as a benchmark for drug costs. For energy consumption, This is the energy consumption baseline value. , , , These are the weighting coefficients.

10. The intelligent mineral flotation control method based on multi-modal perception as claimed in claim 7, wherein, In step S5, the specific operation of updating the reinforcement learning strategy network parameters online is as follows: adopting an experience replay mechanism and a priority sampling strategy, prioritizing the learning of control experience when flotation conditions change abruptly, accelerating strategy convergence, establishing a knowledge graph of the flotation process, integrating mineralogical characteristics, reagent action mechanisms, and equipment operating parameters, providing prior knowledge constraints for deep reinforcement learning, and avoiding dangerous conditions in the exploration process.

Images