Intelligent sorting system and method for upgrading of fluorite

Abstract

The present application belongs to the technical field of mineral separation, and particularly relates to an intelligent separation system and method for upgrading of fluorite. The system comprises an ore online detection unit, an intelligent decision center, a medium density dynamic regulation unit and a separation execution and product collection unit. The ore online detection unit collects multi-dimensional physical property data, the intelligent decision center calculates the optimal separation medium density through a deep neural network model, the medium density dynamic regulation unit accurately adjusts the medium density, and the separation execution unit completes the separation and collects the products. The system adopts a multi-level control architecture and has a model self-updating mechanism, can adapt to fluctuations in ore properties in real time, and solves the problems of low efficiency and unstable concentrate grade in traditional fixed density separation, and improves the separation precision, resource recovery rate and stability of concentrate grade.

Description

Technical Field

[0001] This invention belongs to the field of mineral sorting technology, specifically relating to an intelligent sorting system and method for fluorite upgrading. Background Technology

[0002] Fluorite, as an important strategic non-metallic mineral resource, requires efficient sorting and upgrading, which are crucial steps in mineral processing. The grade and recovery rate of fluorite concentrate directly affect the product quality and cost of downstream industries such as fluorochemicals, metallurgy, and new materials. Therefore, developing efficient sorting technologies is of great significance for improving resource utilization.

[0003] Gravity separation technology is widely used in fluorite separation due to its large processing capacity and relatively low cost, especially in heavy medium hydrocyclone separation processes. This technology separates minerals by specific gravity by controlling the density of the medium, and theoretically can separate fluorite and gangue minerals with relatively coarse particle size and significant differences in specific gravity.

[0004] Existing technologies generally use media with fixed density or limited adjustment range for separation, which is difficult to adapt to the complex composition and large grade fluctuations of fluorite ore. Media with fixed density cannot achieve optimal separation, resulting in high-grade fluorite particles being lost in tailings because their density is close to that of the media, or low-grade ore being mixed into the concentrate product because the density difference is not significant.

[0005] Traditional systems lack real-time detection and feedback mechanisms for the grade and mineral composition of the feed ore, resulting in a significant lag in density adjustment compared to changes in feed properties. This mismatch between the static sorting mode and dynamic feed conditions leads to low sorting efficiency, unstable concentrate grades, and resource waste, particularly when dealing with complex and variable mine feed. Therefore, there is an urgent need to develop a fluorite upgrading technology capable of real-time sensing of ore characteristics and intelligent adjustment of sorting parameters. Summary of the Invention

[0006] The technical problem this invention aims to solve is to overcome the shortcomings of existing fluorite gravity separation technologies, which, due to the use of fixed or limited adjustable media densities, cannot adapt to the complex composition and grade fluctuations of raw ore, resulting in low separation efficiency, unstable concentrate grade, and low resource recovery rate. The purpose of this invention is to provide an intelligent separation system and method for fluorite upgrading. This system can achieve real-time sensing of the characteristics of the feed ore and intelligently adjust the separation parameters accordingly, thereby improving separation accuracy and resource utilization.

[0007] To achieve the above objectives, the present invention provides an intelligent sorting system for fluorite upgrading, which includes an online ore detection unit, an intelligent decision-making center, a media density dynamic control unit, and a sorting execution and product collection unit.

[0008] The online ore detection unit is deployed downstream of the feed inlet of the sorting system to collect multi-dimensional physical property data of the fluorite ore to be sorted in real time.

[0009] The intelligent decision center communicates with the online ore detection unit, receives and processes the multi-dimensional physical characteristic data, and calculates the current optimal density setting value of the sorting medium through the built-in intelligent analysis model.

[0010] The media density dynamic control unit is connected to the intelligent decision center, receives the optimal sorting media density setting value, and drives the actuator to make precise and rapid adjustments to the density of the circulating media in the sorting system.

[0011] The sorting and product collection unit includes a heavy medium cyclone separator and a concentrate and tailings collection device. It completes the ore sorting under controlled medium density conditions and collects high-grade fluorite concentrate and low-grade tailings respectively.

[0012] Furthermore, the online ore detection unit integrates multiple sensing and detection modules. This unit includes a hyperspectral imaging module, an online particle size analysis module, and a real-time density detection module. The hyperspectral imaging module continuously scans the surface of the feed ore at a preset scanning frequency to acquire its spectral reflectance characteristic data. The online particle size analysis module uses the principle of laser diffraction to measure the particle size distribution of the ore in real time. The real-time density detection module, based on gamma-ray transmission technology, non-contactly determines the overall average density of the feed ore.

[0013] These three modules work synchronously to form a multi-dimensional physical property data stream that characterizes the properties of the feed ore.

[0014] Furthermore, the intelligent decision-making center includes a data fusion preprocessing module, a sorting density prediction model, and a decision output module. The data fusion preprocessing module normalizes and extracts features from the hyperspectral data, particle size distribution data, and average density data from the online ore detection unit, eliminating dimensional differences and extracting key feature vectors for decision-making.

[0015] The sorting density prediction model is a regression model based on a deep neural network. Its input is a preprocessed key feature vector, and its output is the predicted optimal sorting medium density value.

[0016] The training data for this model comes from the ore feeding characteristics, the actual density of the medium used, and the corresponding sorting efficiency indicators recorded during the historical sorting process.

[0017] The decision output module uses the density value calculated by the sorting density prediction model as a control command and sends it to the medium density dynamic control unit.

[0018] Furthermore, the medium density dynamic control unit consists of a medium density sensor, an adjustable thickener, and a density control valve assembly.

[0019] The medium density sensor monitors the instantaneous density value of heavy media in the circulating pipeline in real time.

[0020] The adjustable thickener changes the solid concentration of the media suspension by adjusting its underflow discharge rate based on the optimal sorting media density setting received from the intelligent decision center.

[0021] The density control valve assembly includes a water supply valve and a medium supply valve. They receive feedback signals from the medium density sensor and setpoint instructions from the intelligent decision center. Through the coordinated action of proportional-integral-derivative control algorithms, they precisely add water or high-density medium powder to the circulation system, thereby achieving fine-tuning and stable control of the medium density and ensuring that the system's working density quickly converges to the target setpoint.

[0022] Furthermore, the feed pressure of the heavy medium cyclone in the sorting and product collection unit is maintained constant by a frequency converter-driven feed pump.

[0023] The speed of the feed pump is regulated in a closed loop based on the pressure sensor reading at the hydrocyclone inlet.

[0024] The concentrate and tailings collection devices are each equipped with an online grade analyzer to continuously monitor the grade of calcium fluoride in the concentrate product and the loss of fluorite in the tailings.

[0025] This grade data is fed back to the intelligent decision-making center as long-term learning data to optimize the parameters of the sorting density prediction model.

[0026] As one embodiment of the present invention, the updating mechanism of the sorting density prediction model of the intelligent decision-making center is as follows.

[0027] The system periodically collects sorting performance data for each batch of ore, including the actual grade and recovery rate of the concentrate, for example, after each batch of ore has been processed.

[0028] Compare these performance data with the expected sorting performance predicted by the model to calculate the performance deviation.

[0029] By utilizing this performance deviation and the corresponding ore feeding characteristic data, the network weights of the sorting density prediction model are fine-tuned through the backpropagation algorithm to achieve model adaptation and continuous optimization.

[0030] In another embodiment of the present invention, the system operates within a multi-level control architecture. This architecture includes a strategic planning layer, a tactical scheduling layer, and a real-time control layer.

[0031] At the strategic planning level, the overall sorting grade and recovery rate targets are set on a weekly or monthly timescale based on the mine's medium- and long-term mining plan.

[0032] The tactical dispatch layer uses shifts or days as time scales, and adjusts system operating parameters and plans maintenance cycles based on strategic objectives and the expected nature of incoming ore.

[0033] The real-time control layer uses seconds or minutes as the time scale to execute the closed-loop control process of the aforementioned online ore detection, intelligent decision-making, dynamic control of medium density, and sorting.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention integrates an online ore detection unit with an intelligent decision-making center to construct a real-time closed-loop control system that senses feed characteristics and makes decisions on sorting parameters. This system can sensitively capture fluctuations in the composition, particle size, and density of the feed ore, and quickly calculate the optimal sorting medium density using a trained deep neural network model. This completely changes the passive situation of the traditional fixed-density sorting mode being mismatched with dynamic feed conditions, thereby improving the adaptability and response speed of the sorting process to changes in ore properties, and ensuring the stability of concentrate grade and high recovery rate.

[0036] 2. The dynamic density control unit used in this invention achieves rapid and precise adjustment of the density of the heavy medium through the coordinated control of an adjustable thickener and a precision valve assembly. This dynamic control capability ensures that the separation process is always carried out at or near the optimal separation density, minimizing the loss of high-grade fluorite or the mixing of low-grade ore due to unsuitable medium density. This fundamentally improves separation efficiency and resource utilization, while reducing tailings grade and resource waste.

[0037] 3. The built-in model self-updating mechanism of this invention endows the system with the ability to continuously optimize. By feeding back the actual sorting effect to the intelligent decision center and using it to fine-tune the sorting density prediction model, the system can continuously accumulate operational experience, adapt to the long-term and slow changes in ore characteristics during mining, making sorting decisions increasingly accurate, exhibiting good self-learning and adaptive characteristics, and maintaining high technical and economic indicators under long-term operation. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall technical solution architecture of the intelligent sorting system for fluorite quality improvement proposed in this invention;

[0039] Figure 2 This is a schematic diagram of the core principle framework of the intelligent decision-making center in this invention;

[0040] Figure 3This is a flowchart illustrating the multi-dimensional data acquisition logic of the online ore detection unit in this invention.

[0041] Figure 4 This is a schematic diagram of the control logic and execution framework of the medium density dynamic control unit in this invention;

[0042] Figure 5 This is a schematic diagram of the sorting process and feedback mechanism of the sorting execution and product collection unit in this invention;

[0043] Figure 6 This is a schematic diagram of the hierarchical interaction relationship and data flow of the multi-level control architecture in this invention; Detailed Implementation

[0044] Please refer to the attached document. Figure 1 This embodiment details a specific implementation of an intelligent sorting system for fluorite upgrading. The system aims to overcome the shortcomings of traditional gravity separation technology, which is unable to adapt to the complex composition and grade fluctuations of raw ore due to fixed or limited media density, by intelligently adjusting sorting parameters in real time based on the characteristics of the feed ore. This improves sorting accuracy and resource utilization. The core components of the system include an online ore detection unit, an intelligent decision-making center, a media density dynamic control unit, and a sorting execution and product collection unit.

[0045] These units interact with each other through close communication and data exchange, forming a complete closed-loop control system that encompasses ore characteristic perception, intelligent decision-making, parameter control, sorting execution, and effect feedback.

[0046] The online ore detection unit is deployed downstream of the feed inlet of the sorting system. Its core function is to collect multi-dimensional physical property data of the fluorite ore to be sorted in real time. This unit integrates a hyperspectral imaging module, an online particle size analysis module, and a real-time density detection module.

[0047] Please refer to the attached document. Figure 3 The hyperspectral imaging module continuously scans the surface of the feed ore at a preset scanning frequency to acquire its spectral reflectance characteristics. This module employs pushbroom imaging technology, with a spectral range covering the visible to short-wave infrared bands, specifically 400nm to 2500nm. During the scanning process, each scan line contains 1000 spatial pixels, and each pixel records spectral intensity values ​​for 256 bands.

[0048] After dark current correction and white light reference correction, these raw spectral data form a high-dimensional data cube that reflects the mineral composition and distribution characteristics of the ore surface.

[0049] The online particle size analysis module uses the principle of laser diffraction to measure the particle size distribution of ore particles in real time.

[0050] The module emits a 632.8 nm laser beam that passes through the flowing ore stream, and the laser signals scattered by the particles are received by a ring detector array. The detector array contains 31 detector rings, each corresponding to a particle size measurement range from 0.1 μm to 2000 μm.

[0051] The module's built-in Mie scattering theory algorithm inverts the scattered light energy distribution into a volume-based particle size distribution, and outputs characteristic parameters including D10, D50, D90, and the distribution width index.

[0052] The real-time density detection module uses gamma-ray transmission technology to non-contactly measure the overall average density of the feed ore.

[0053] This module uses cesium-137 as the radiation source, emitting gamma rays with an energy of 660 keV. After penetrating the material flow, the rays are received by a sodium iodide scintillator detector. By measuring the attenuation of the ray intensity and combining it with the known material flow thickness and the ore mass absorption coefficient, the bulk density of the ore is calculated.

[0054] Density calculations are updated 10 times per second, with a measurement accuracy of ±0.5%.

[0055] These three modules, through a precise time synchronization mechanism, ensure that their data acquisition is completed within millisecond-level time deviations, and together constitute a multi-dimensional physical characteristic data stream characterizing the properties of the feed ore.

[0056] The data stream is transmitted to the intelligent decision-making center via industrial Ethernet at a rate of 100MT / s.

[0057] The intelligent decision center communicates with the online ore detection unit, and is responsible for receiving and processing multi-dimensional physical characteristic data. It calculates the current optimal density setting value of the sorting medium through the built-in intelligent analysis model.

[0058] Please refer to the attached document. Figure 2 The intelligent decision-making center includes a data fusion preprocessing module, a sorting density prediction model, and a decision output module.

[0059] The data fusion preprocessing module first normalizes the hyperspectral data, particle size distribution data, and average density data from the online ore detection unit.

[0060] For hyperspectral data, the module extracts the reflectance values ​​of each pixel in four characteristic bands: 520nm, 1400nm, 1900nm, and 2200nm. These four bands correspond to the diagnostic spectral characteristics of typical associated minerals such as fluorite, calcite, and quartz.

[0061] The extracted spectral features, together with the average density of the ore and the D50 particle size value, form a 7-dimensional original feature vector.

[0062] The normalization process uses the minimax method to linearly transform the values ​​of each feature dimension to the range of 0-1.

[0063] In the feature extraction stage, principal component analysis algorithm is used to extract three principal components from the original 7-dimensional feature vector. These three principal components contribute more than 95% of the variance information of the original data and constitute the key feature vector for decision-making.

[0064] The sorting density prediction model is a regression model built on a deep neural network.

[0065] The network structure consists of one input layer, three hidden layers, and one output layer. The input layer has three neurons, corresponding to three key feature vectors.

[0066] The first hidden layer contains 64 neurons and uses the rectified linear unit activation function.

[0067] The second hidden layer contains 32 neurons and also uses the rectified linear unit activation function.

[0068] The third hidden layer contains 16 neurons and uses the hyperbolic tangent activation function.

[0069] The output layer consists of a single neuron, employing a linear activation function to directly output the predicted optimal sorting medium density value, typically ranging from 2.6 g / cm³. 3 Up to 2.9 g / cm 3 between.

[0070] The training data for this model comes from the ore feeding characteristics, the actual density of the medium used, and the corresponding sorting efficiency indicators recorded during the historical sorting process.

[0071] The training dataset contains more than 10,000 samples, each containing a key feature vector, the actual density of the medium used, and the actual concentrate grade and recovery rate of that batch.

[0072] The model training uses mean squared error as the loss function and is optimized using the adaptive moment estimation algorithm. The initial learning rate is set to 0.001, the batch size is 32, and the training period is 500 rounds.

[0073] After training, the model weights are fixed and deployed in the industrial computer of the intelligent decision-making center.

[0074] In real-time operation, whenever a new batch of key feature vectors is received, the model calculates the optimal sorting medium density value during forward propagation.

[0075] The decision output module encapsulates the density value into a standard Modbus TCP control command, which includes the target density value, timestamp, and command priority information, and sends it to the media density dynamic control unit via Gigabit Ethernet.

[0076] The transmission frequency is consistent with the data update frequency of the online ore detection unit, typically 10 times per second.

[0077] The media density dynamic control unit is connected to the intelligent decision center, receives the optimal sorting media density setpoint, and drives the actuator to make precise and rapid adjustments to the density of the circulating media in the sorting system.

[0078] Please refer to the attached document. Figure 4 The unit consists of a medium density sensor, an adjustable thickener, and a density control valve assembly.

[0079] The medium density sensor uses a nuclear density meter to monitor the instantaneous density value of heavy media in the circulating pipeline in real time.

[0080] The sensor measures using gamma ray attenuation, with a measurement range of 1.5 g / cm³. 3 Up to 3.5g / cm 3 The measurement accuracy is ±0.005 g / cm³. 3 The data output frequency is 20 times per second.

[0081] The adjustable thickener changes the solid concentration of the media suspension by adjusting its underflow discharge rate based on the optimal sorting media density setting received from the intelligent decision center.

[0082] The thickener has a diameter of 5m and a height of 4m. Its rake frame drive motor uses frequency conversion control, and the speed adjustment range is 0.5PRM to 3PRM. The underflow discharge pump is a diaphragm pump, and its stroke frequency can be continuously adjusted from 0 to 100 times per minute via a 4mA to 20mA analog signal.

[0083] When the target medium density setpoint increases, the control system reduces the stroke frequency of the underflow discharge pump, thereby increasing the solid concentration of the medium inside the thickener and correspondingly increasing the underflow density. Conversely, when the target density decreases, the discharge frequency is increased to dilute the medium suspension.

[0084] The density control valve assembly includes a water supply valve and a medium supply valve. The water supply valve is an electrically adjustable valve with a nominal diameter of 50 mm and uses an equal percentage flow characteristic. Its valve opening can be continuously adjusted between 0% and 100%.

[0085] The replenishing valve is also an electric regulating valve with a diameter of 40mm. It is used to add high-density media powder to the system. Its media powder silo is equipped with a loss-in-weight feeder with a feeding accuracy of ±1%.

[0086] These two valves receive feedback signals from the medium density sensor and setpoint commands from the intelligent decision center. The control algorithm employs proportional-integral-derivative (PID) control. The proportional gain is set to 2.5, the integral time to 60 seconds, and the derivative time to 10 seconds. The control algorithm executes every 0.1 seconds, calculating the coordinated opening commands for the water supply valve and the medium supply valve.

[0087] When the instantaneous medium density is less than the target set value but greater than 0.01 g / cm³ 3 When the instantaneous medium density exceeds the target set value by more than 0.01 g / cm³, the opening of the replenishing medium valve increases according to the proportional-integral-differential algorithm, while the opening of the water replenishing valve decreases accordingly. 3 When the water supply valve opening increases according to the proportional integral differential algorithm, the water supply valve opening decreases or closes.

[0088] Through this collaborative control, the system achieves fine-tuning and stable control of the medium density, ensuring that the working density quickly converges to the target set value, with density overshoot less than 5% and stabilization time less than 3 minutes during dynamic adjustment.

[0089] The sorting and product collection unit includes heavy medium cyclone separators and concentrate and tailings collection devices.

[0090] The heavy medium cyclone adopts a standard DenseMediumCyclone structure with a cylinder diameter of 500mm and a cone angle of 20°.

[0091] The feed pressure of the hydrocyclone is kept constant by a variable frequency drive feed pump with a rated pressure of 150 kPa.

[0092] The feed pump motor has a power of 55KW, and the frequency converter receives readings from the pressure sensor at the hydrocyclone inlet.

[0093] The pressure sensor has a range of 0 to 500 kPa. The control system constructs a pressure closed loop, with a set pressure value of 150 kPa.

[0094] When the measured pressure deviates from the set value by more than ±5 kPa, the inverter output frequency is adjusted according to the proportional-integral algorithm to change the pump speed and bring the pressure back to the set range.

[0095] Under controlled medium density conditions, the ore is separated in the hydrocyclone.

[0096] Ore particles with a density higher than that of the medium are discharged from the bottom outlet as sediment, forming concentrate products.

[0097] Particles with a density less than that of the medium are discharged from the overflow outlet as floating debris, forming tailings.

[0098] Please refer to the attached document. Figure 5The concentrate and tailings collection devices are each equipped with an online grade analyzer. The online grade analyzer installed at the concentrate collection sluice uses X-ray fluorescence technology to continuously monitor the grade of calcium fluoride in the concentrate product.

[0099] The analyzer performs a measurement every 30 seconds, with the measurement beam covering the entire cross-section of the chute to ensure representative sampling.

[0100] The calcium fluoride grade measurement range is 0 to 100%, with an accuracy of ±2%.

[0101] Online grade analyzers are also installed at the tailings collection pipelines, using the same technology to monitor the loss of fluorite in the tailings, i.e., the calcium fluoride content in the tailings. This grade data is fed back to the intelligent decision-making center in real time via an industrial wireless network. The feedback data packet includes a timestamp, grade value, and data quality identifier.

[0102] The data fusion and preprocessing module of the intelligent decision-making center receives this data and associates it with the corresponding ore feed characteristic data and medium density setpoints, storing it in the historical database as long-term learning data for optimizing the parameters of the sorting density prediction model.

[0103] The update mechanism of the sorting density prediction model in the intelligent decision-making center is as follows: After the system processes each batch of ore, approximately every 500 tons of raw ore, a model update check is triggered.

[0104] The update process first collects the sorting performance data for this batch, including the actual average grade of the concentrate and the cumulative recovery rate.

[0105] The recovery rate was calculated based on the feed rate, concentrate rate, and their respective grades. These actual results were then compared with the expected sorting results predicted by the model.

[0106] The expected separation effect is calculated based on the ore characteristics and the set medium density using a simplified theoretical separation model.

[0107] Performance deviation is calculated as the absolute difference between the actual recovery rate and the expected recovery rate, and the absolute difference between the actual concentrate grade and 95% of the target grade. If the performance deviation exceeds a preset threshold, such as a recovery rate deviation greater than 3% or a grade deviation greater than 1.5%, model fine-tuning is initiated.

[0108] The fine-tuning process utilizes the key feature vectors of the feed characteristics corresponding to this batch and the calculated performance deviations to make minor adjustments to the network weights of the sorting density prediction model through the backpropagation algorithm.

[0109] The learning rate is set to 1 / 10 of the initial training learning rate, or 0.0001, during fine-tuning. The fine-tuning process uses only the most recent 1000 batches of data as the validation set to prevent overfitting. This mechanism enables the model to adapt and continuously optimize, allowing the system to adapt to the long-term, slow changes in ore properties during mining operations.

[0110] The system operates on a multi-level control architecture.

[0111] Please refer to the attached document. Figure 6 The architecture comprises a strategic planning layer, a tactical scheduling layer, and a real-time control layer. The strategic planning layer operates on a weekly or monthly timescale on the mine's central server.

[0112] Based on the mine's medium- and long-term mining plan, and taking into account market calcium fluoride prices and operating costs, this layer sets the overall sorting targets for the system, such as a concentrate grade target of 95% to 97% and a recovery rate target of 88% to 92%. These targets are distributed to the tactical scheduling layer in the form of parameter configuration files.

[0113] The tactical dispatch layer operates on the central control room computer of the ore dressing plant on a shift or day time scale.

[0114] This layer adjusts the system operating parameters based on strategic objectives and the expected nature of incoming minerals.

[0115] Pre-adjustment operations include setting the initial medium density value, the feed pump reference frequency, and the thickener initial rake speed.

[0116] Meanwhile, the system maintenance cycle is planned for this layer, for example, the media purification system is backwashed every 720 hours of operation, and the online detection unit is optically calibrated every 2000 hours of operation.

[0117] The real-time control layer, measured in seconds or minutes, is deployed in the programmable logic controllers and industrial computers of each unit to execute the closed-loop control process of online ore detection, intelligent decision-making, dynamic control of medium density, and sorting.

[0118] Data is exchanged between different levels through the Manufacturing Execution System, ensuring that instructions flow from top to bottom and production data converges from bottom to top, forming a complete integrated control system.

[0119] This embodiment provides an alternative implementation scheme for an intelligent sorting system for fluorite quality improvement. Its core features are the different construction strategies of the sorting density prediction model adopted by the intelligent decision center and the differentiated configuration of the actuators in the medium density dynamic control unit.

[0120] The overall system architecture remains consistent, including the serial and closed-loop connection of the online ore detection unit, intelligent decision center, dynamic control unit for medium density, and sorting execution and product collection unit.

[0121] However, optimizations and adjustments have been made to the model algorithm and some execution details to adapt to application scenarios with higher frequency fluctuations in ore characteristics or special requirements for model interpretability.

[0122] In the intelligent decision center section, the sorting density prediction model does not use a deep neural network, but is built based on the gradient boosting decision tree algorithm.

[0123] The model consists of an ensemble of one hundred decision trees, with each tree having a maximum depth limit of six layers.

[0124] The model's input consists of three key feature vectors extracted by the data fusion preprocessing module. The training process employs a gradient boosting framework with a learning rate of 0.1 and a subsampling ratio of 0.8. The loss function is the absolute error.

[0125] Compared to deep neural network models, gradient boosting decision tree models often exhibit better generalization performance with relatively small training data, such as fewer than 5,000 samples. Furthermore, the model's prediction results are interpretable to a certain extent, and the contribution of different features such as hyperspectral, granularity, and density to the final density decision can be quantified by ranking the importance of features.

[0126] When sending control commands, the decision output module attaches a confidence index, which is calculated based on the distribution density of the predicted density value in the neighborhood of the training samples.

[0127] When the confidence level is less than 90%, the system will maintain the previous high-confidence density setting value and issue an early warning signal to check the operating status of the online ore detection unit, thereby enhancing the robustness of the system's decision-making.

[0128] The dynamic control unit for medium density has been enhanced in terms of the control strategy of the adjustable thickener.

[0129] In addition to adjusting the overall concentration through the underflow discharge rate, this embodiment adds an automatic polymer flocculant dosing system to the top of the thickener. This system includes a flocculant preparation unit and two metering pumps.

[0130] The flocculant preparation unit prepares a 0.1% solution of dry powder flocculant.

[0131] Two metering pumps are used to add flocculant to the thickener feed well and the settling zone, respectively.

[0132] The dosage is feedforward compensated based on the rate of change of the thickener underflow density setpoint.

[0133] When the rate of change of the set value is greater than 0.01 g / cm 3 When the flow rate is 1 / min, the feedforward controller will increase the amount of flocculant added proportionally to accelerate the settling of media particles, thereby speeding up the underflow density response.

[0134] The metering pump has a flow control accuracy of 1% and a maximum dosing rate of 10 L / min.

[0135] This improvement is particularly suitable for treating media suspensions with high fine mud content, improving the settling effect of thickeners and enhancing the dynamic response performance of density control.

[0136] The control algorithm for the density control valve assembly incorporates fuzzy logic compensation based on proportional-integral-derivative (PID) algorithms. The system defines the medium density deviation and the rate of change of deviation as fuzzy input variables.

[0137] The fuzzy set of density deviation is defined as negative large, negative small, zero, positive small, and positive large. The fuzzy set of deviation change rate is defined as negative fast, negative slow, zero, positive slow, and positive fast.

[0138] The output variable is the correction amount for the opening of the water supply valve and the medium supply valve.

[0139] The fuzzy rule base contains 25 rules. For example, if the density deviation is positive and the rate of change of the deviation is positive and fast, then the correction amount for the water supply valve opening is positive and the correction amount for the medium supply valve opening is negative.

[0140] The Mamdani method is used for fuzzy reasoning, and the centroid method is used for defuzzification.

[0141] The fuzzy compensator runs once every 0.5 seconds, and its output is superimposed with the output of the proportional-integral-derivative controller to form the final valve position command.

[0142] This composite control strategy has a better control effect on the nonlinear and large hysteresis characteristics in the process of medium density regulation, and can further reduce steady-state error and suppress overshoot.

[0143] In the sorting and product collection unit, the constant feed pressure control of the heavy medium cyclone adds coupling logic between the pressure setpoint and the medium density setpoint.

[0144] The control system incorporates a pressure-density coupling model, which dynamically fine-tunes the feed pressure setpoint based on the current medium density setpoint. The coupling relationship is described by an empirical formula; when the medium density setpoint is 2.6 g / cm³... 3 Up to 2.9 g / cm 3 Within the specified range, the feed pressure setting is calculated by adding the difference between the density setting and 2.75 to the base value of 150 kPa, multiplied by a coupling coefficient of 50 kPa / g / cm³. 3 To determine.

[0145] For example, when the medium density is set to 2.8 g / cm³ 3 At that time, the calculated feed pressure setting value was 152.5 kPa.

[0146] This dynamic coupling ensures that the internal flow field of the hydrocyclone always maintains optimal separation conditions under different medium densities, which helps to further improve the separation efficiency.

[0147] The utilization of feedback data from the online grade analyzer has also been enhanced. In addition to its use in long-term model learning, this embodiment adds a real-time fine-tuning loop based on grade feedback. This loop is independent of the main prediction model in the intelligent decision center.

[0148] When the online concentrate grade analyzer detects that the concentrate grade is continuously less than the target value by more than 2% for three consecutive times within 1.5 minutes, the real-time fine-tuning loop will trigger a small medium density reduction command with a reduction margin of 0.5%.

[0149] Conversely, if the grade remains more than 2% above the target value, a minor density increase instruction will be triggered.

[0150] This fine-tuning instruction is weighted and combined with the set value output by the main model, with the main model accounting for 90% and the fine-tuning loop accounting for 10%.

[0151] This design enables the system to quickly compensate for transient changes in ore properties that the model fails to capture in time, thereby further stabilizing the quality of concentrate products.

[0152] The system's multi-level control architecture adds a function for pre-adjusting operating parameters based on weather forecast data to the tactical scheduling layer.

[0153] The computer at the tactical dispatch layer obtains environmental temperature and humidity forecast data for the next 24 hours through an application programming interface.

[0154] Since ambient temperature and humidity affect the viscosity and sedimentation characteristics of the medium suspension, the system compensates for the initial medium density setpoint based on the forecast data.

[0155] The compensation logic is that when the forecast ambient temperature is higher than 25℃ or the relative humidity is less than 40%, the initial density setting value is increased by 0.5% above the base value.

[0156] When the forecast ambient temperature is less than 10℃ or the relative humidity is higher than 80%, the initial density setting value is adjusted up or down by 0.5% from the base value.

[0157] This feature reflects the system's comprehensive consideration of a wider range of influencing factors and enhances the system's adaptability under different environmental conditions.

[0158] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0159] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent sorting system for fluorite upgrading, characterized in that, include: The online ore detection unit is deployed downstream of the feed inlet of the sorting system to collect multi-dimensional physical property data of the fluorite ore to be sorted in real time. The intelligent decision center communicates with the online ore detection unit to receive and process multi-dimensional physical property data, and calculates the current optimal density setting value of the sorting medium through the built-in intelligent analysis model. The media density dynamic control unit is connected to the intelligent decision center to receive the optimal sorting media density setpoint and drive the actuator to make precise and rapid adjustments to the density of the circulating media in the sorting system. The sorting and product collection unit includes a heavy medium hydrocyclone sorting device and a concentrate and tailings collection device, which is used to sort the ore under controlled medium density conditions and collect high-grade fluorite concentrate and low-grade tailings respectively. The online ore detection unit includes a hyperspectral imaging module, an online particle size analysis module, and a real-time density detection module; The hyperspectral imaging module continuously scans the surface of the feed ore at a preset scanning frequency to acquire spectral reflectance characteristic data. The online particle size analysis module uses the principle of laser diffraction to measure the particle size distribution of ore particles in real time; The real-time density detection module uses gamma-ray transmission technology to non-contactly measure the overall average density of the feed ore. The intelligent decision-making center includes a data fusion preprocessing module, a sorting density prediction model, and a decision output module. The data fusion preprocessing module normalizes and extracts features from the hyperspectral data, particle size distribution data and average density data from the online ore detection unit, eliminating dimensional differences and extracting key feature vectors for decision-making. The sorting density prediction model is a regression model built on a deep neural network. The input is a preprocessed key feature vector, and the output is the predicted optimal sorting medium density value. The decision output module sends the density value calculated by the sorting density prediction model as a control command to the medium density dynamic control unit. The dynamic control unit for medium density includes a medium density sensor, an adjustable thickener, and a density control valve assembly. The medium density sensor monitors the instantaneous density value of heavy media in the circulating pipeline in real time; The adjustable thickener changes the solid concentration of the media suspension by adjusting its underflow discharge rate based on the optimal sorting media density setting value received from the intelligent decision center. The density control valve assembly includes a water supply valve and a medium supply valve. It receives feedback signals from the medium density sensor and setpoint instructions from the intelligent decision center. Through the coordinated action of proportional-integral-derivative control algorithms, it accurately adds water or high-density medium powder to the circulation system. The data fusion preprocessing module needs to perform normalization processing. The normalization processing adopts the maximum and minimum value method to linearly transform the value of each feature dimension to between 0 and 1. In the feature extraction stage, three principal components are extracted from the seven-dimensional original feature vector through principal component analysis algorithm; The three principal components collectively contribute more than 95% of the variance information of the original data, forming the key feature vectors used for decision-making.

2. The intelligent sorting system for fluorite upgrading according to claim 1, characterized in that, The step of training data for the sorting density prediction model is to obtain the training data from the ore feeding characteristics, the actual density of the medium used, and the corresponding sorting efficiency index recorded in the historical sorting process. The sorting density prediction model adopts a deep neural network structure, including an input layer, three hidden layers, and an output layer.

3. The intelligent sorting system for fluorite upgrading according to claim 1, characterized in that, The feed pressure of the heavy medium cyclone in the sorting and product collection unit is maintained constant by a frequency-controlled feed pump. The speed of the feed pump is adjusted in a closed loop based on the pressure sensor reading at the hydrocyclone inlet; The concentrate and tailings collection devices are each equipped with an online grade analyzer to continuously monitor the grade of calcium fluoride in the concentrate product and the loss of fluorite in the tailings.

4. The intelligent sorting system for fluorite upgrading according to claim 3, characterized in that, The online grade analyzer uses X-ray fluorescence technology; Grade data is fed back to the intelligent decision-making center in real time via an industrial wireless network, serving as long-term learning data for optimizing the parameters of the sorting density prediction model.

5. The intelligent sorting system for fluorite upgrading according to claim 1, characterized in that, The sorting density prediction model of the intelligent decision-making center has an update mechanism; After processing each batch of ore, the system collects the sorting effect data for that batch, including the actual grade and recovery rate of the concentrate. The actual results are compared with the expected sorting results predicted by the model, and the performance deviation is calculated. By utilizing performance deviations and corresponding ore feeding characteristic data, the network weights of the sorting density prediction model are fine-tuned using the backpropagation algorithm.

6. The intelligent sorting system for fluorite upgrading according to claim 1, characterized in that, The system operates in a multi-level control architecture, including a strategic planning layer, a tactical scheduling layer, and a real-time control layer; At the strategic planning level, the overall sorting grade and recovery rate targets of the system are set on a weekly or monthly time scale, based on the medium- and long-term mining plans of the mine. The tactical dispatch layer uses shifts or days as time scales, and adjusts system operating parameters and plans maintenance cycles based on strategic objectives and the expected nature of incoming minerals. The real-time control layer executes a closed-loop control process, using seconds or minutes as the time scale, for online ore detection, intelligent decision-making, dynamic control of medium density, and sorting.

7. The intelligent sorting system for fluorite upgrading according to claim 6, characterized in that, The tactical dispatch layer adds a function for pre-adjusting operating parameters based on weather forecast data; Obtain future ambient temperature and humidity forecast data through application programming interfaces; The initial medium density setpoint is compensated based on the forecast data.

8. A smart sorting method for fluorite upgrading, characterized in that, Intelligent sorting is achieved using the intelligent sorting system for fluorite upgrading as described in any one of claims 1-7.

Images