A sulfur flotation intelligent control system and method based on dynamic characteristics of foam grayscale

Abstract

The application discloses a kind of intelligent control system and method of sulphur flotation based on foam gray scale dynamic characteristics, it is related to mineral processing and intelligent ore dressing control technical field, including foam gray scale feature perception module, grade soft measurement prediction module and intelligent setting and controller module, the device is by setting foam gray scale feature perception module real-time acquisition flotation tank foam image and extract gray scale dynamic characteristics, grade soft measurement prediction module is based on gray scale dynamic characteristics and constructs soft measurement model and predicts sulphur concentrate grade in real time, intelligent setting and controller module according to the deviation of predicted grade and set value executes feed-back compound control strategy and automatically adjusts flotation process parameter, realizes the advance regulation and control and stable optimization of flotation process.

Description

Technical Field

[0001] This invention relates to the field of mineral processing and intelligent mineral processing control technology, specifically to an intelligent control system and method for sulfur flotation based on the dynamic characteristics of foam grayness. Background Technology

[0002] The pyrite flotation process is characterized by strong time-varying, strong coupling, and strong randomness. Its concentrate grade is affected by multiple factors, including ore properties, reagent dosage, pulp concentration, air volume, and operator subjective experience. Traditional grade detection methods mainly rely on manual periodic sampling and analysis. The long time lag means that the control strategy can only be adjusted after seeing the results, which is essentially a passive corrective control. This often leads to problems such as drastic grade fluctuations, excessive reagent addition, and frequent adjustments by operators based on experience, making it difficult to maintain stable and optimal operating conditions. Currently, some mines use foam imaging systems (focusing on foam size, speed, and agitation) or online XRF grade analyzers for detection; however, existing technologies have significant shortcomings: the physical structure of foam does not have a stable mapping relationship with grade, making it difficult to establish a reliable prediction model; XRF analyzers are expensive, complex to maintain, greatly affected by pulp concentration, and not applicable to sulfur flotation detection; the applicant has discovered through long-term industrial experiments that there is a continuous, stable, and modelable strong correlation between the gray value of flotation foam and the grade of sulfur concentrate: the "darker" the foam, the more attached pyrite particles, and the higher the concentrate grade; the "brighter" the foam, the more gangue or moisture is present, and the lower the grade; gray value features are not only related to grade, but are also easy to obtain, have low computational cost, and are easy to process image noise, making it easy to build a stable grade soft measurement model; however, there is currently no method in the existing technology to construct a real-time grade prediction model from "foam gray value features" and use it for "feedforward intelligent control," therefore, this invention has significant innovation. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent control system and method for sulfur flotation based on the dynamic characteristics of foam grayness, so as to solve the problems mentioned in the background art.

[0004] A sulfur flotation intelligent control system based on the dynamic characteristics of foam grayscale includes a foam grayscale feature sensing module, a grade soft measurement prediction module, and an intelligent setting and controller module. The foam grayscale feature sensing module is used to acquire foam images of the flotation cell in real time and extract grayscale dynamic features. The grade soft measurement prediction module constructs a soft measurement model based on the grayscale dynamic features to predict the grade of sulfur concentrate in real time. The intelligent setting and controller module executes a feedforward-feedback composite control strategy according to the deviation between the predicted grade and the set value to automatically adjust the flotation process parameters.

[0005] Furthermore, the foam grayscale feature perception module includes an image acquisition unit and an image processing unit; the image acquisition unit installs industrial grayscale cameras in the final selection tank and the sulfur coarse selection tank respectively, the cameras are equipped with an 850nm near-infrared ring light source, the camera parameters are fixed and locked, and the image acquisition frequency is 10-20 FPS; the image processing unit performs ROI region extraction, illumination equalization, median filtering and grayscale normalization processing.

[0006] Furthermore, the grayscale dynamic features include global statistical features, distribution features, texture features, and temporal features; the global statistical features include average grayscale and grayscale standard deviation; the distribution features include high grayscale proportion and low grayscale proportion; the texture features include LBP texture features and GLCM contrast; and the temporal features include grayscale change rate and its second-order difference.

[0007] Furthermore, the grade soft measurement prediction module uses a gradient boosting tree (XGBoost) to construct a soft measurement model, taking grayscale dynamic features as input vectors and outputting predicted grade values ​​through the superposition of multiple regression trees; or it uses a convolutional neural network (CNN) to construct a soft measurement model, taking a normalized ROI image as input, extracting deep texture features through convolutional layers, pooling layers, and fully connected layers, and outputting predicted grade values.

[0008] Furthermore, the grade soft measurement prediction module also includes an online correction mechanism; whenever a new test result arrives, the system stores it in a verification cache queue and lags it with the image prediction value at the corresponding time, and fine-tunes the model using recursive least squares (RLS) or incremental training; the system also has a drift detection unit, which triggers model retraining when the mean absolute error of the sliding window exceeds a threshold.

[0009] Furthermore, the intelligent setting and controller module adopts a fuzzy PID controller or a model predictive control (MPC); the feedforward-feedback composite control strategy includes: feedforward control adjusts the control quantity in advance based on the deviation between the predicted grade and the set value, feedback control performs closed-loop adjustment based on the deviation between the true test value and the set value, and finally the control quantity is weighted and synthesized by the feedforward control quantity and the feedback control quantity and output to the actuator.

[0010] Furthermore, the process parameters include the frequency of collector addition, the opening degree of the foaming agent regulating valve, the opening degree of the liquid level valve, and the air volume.

[0011] A method for intelligent control of sulfur flotation based on the dynamic characteristics of foam grayness, employing an intelligent control system for sulfur flotation based on the dynamic characteristics of foam grayness, further includes the following steps: S1: Real-time acquisition of foam images in the flotation cell; S2: Preprocess the acquired foam image, extract the ROI region and perform illumination equalization, noise reduction and grayscale normalization; S3: Extract grayscale dynamic features from the preprocessed image; S4: Input the extracted grayscale dynamic features into a pre-trained soft measurement model to predict the grade of sulfur concentrate in real time; S5: Compare the predicted grade with the set value, execute feedforward-feedback composite control, and adjust the flotation process parameters; S6: When the test results arrive, perform online calibration of the soft measurement model.

[0012] Furthermore, the grayscale dynamic features in step S3 include average grayscale, grayscale standard deviation, high grayscale ratio, low grayscale ratio, LBP texture features, GLCM contrast, and grayscale change rate.

[0013] Furthermore, the method for constructing the soft measurement model in step S4 is as follows: collect historical foam images and corresponding laboratory quality data, extract grayscale dynamic features to construct a training sample set, train it using the XGBoost algorithm or CNN algorithm, use mean squared error as the loss function, and optimize the hyperparameters through cross-validation.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. By predicting the grade of sulfur concentrate in real time through the characteristics of foam grayscale, a complete closed-loop system from image perception, soft measurement prediction to control execution was constructed, realizing the advanced regulation and stable optimization of the flotation process; 2. By adopting a feedforward-feedback composite control strategy, the amount of reagents and water is adjusted in advance based on the predicted grade trend, which significantly reduces grade fluctuations. Verified in multiple mines, grade fluctuations are reduced by 30-45%, and reagent consumption is reduced by 10-20%. 3. It can be implemented based on conventional industrial cameras and standard control interfaces, which is easy to deploy, has low maintenance costs, improves the qualified rate of concentrate grade by 3-5%, improves the recovery rate by 2-3 percentage points, and reduces the frequency of operator intervention by more than 60%. Attached Figure Description

[0015] Figure 1 This is a block diagram of the system structure of the present invention; Figure 2 This is a flowchart of grayscale feature extraction in this invention; Figure 3 This is a schematic diagram of the soft measurement model structure in this invention; Figure 4 This is a block diagram of the feedforward-feedback control strategy in this invention; Figure 5 This is a schematic diagram of the installation of the industrial grayscale camera in this invention.

[0016] In the picture: 1. High-definition camera; 2. LED constant voltage ring light source; 3. Flotation cell. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1-5 The present invention provides a technical solution: a sulfur flotation intelligent control system based on the dynamic characteristics of foam grayness, including a foam grayness characteristic sensing module, a grade soft measurement and prediction module, and an intelligent setting and controller module.

[0019] I. Hardware Configuration and Installation Please see Figure 5 The foam grayscale feature perception module includes an image acquisition unit and an image processing unit.

[0020] Image acquisition unit: Industrial grayscale cameras (1) are installed in the final cleaning tank and the sulfur roughing tank respectively. The cameras use monochrome CMOS sensors with a resolution ≥1.3MP, a frame rate of 10-20FPS, and a gigabit network interface. The cameras are equipped with an 850nm near-infrared ring light source (2), the brightness of the light source is adjustable by PWM, and a diffuser plate is used to reduce the reflection of the foam surface. The distance between the camera and the foam surface is 0.8-1.2m, the downward angle is 15-30°, and the light source and camera are coaxially installed to ensure uniform illumination of the foam layer surface. The cameras are equipped with IP67 protective housing, heating and demisting device and anti-corrosion pipeline to adapt to the high humidity and corrosive gas environment of the flotation workshop. After the initial installation, grayscale-brightness-ROI three-step calibration is performed, and an automatic recalibration function is set every 7-14 days to ensure long-term operational stability.

[0021] Image processing unit: An edge computing box (4) is used as the computing power carrier. An x86 or ARM architecture industrial control computer is selected and running the Linux operating system. It is responsible for image processing, feature extraction and model inference tasks. The edge computing box communicates with the PLC controller (5) through OPC DA or Modbus TCP protocol. The control cycle is adjustable from 1Hz to 0.2Hz. The system is equipped with a whitelist and TLS security mechanism. When the connection is lost, it automatically falls back to the safe operating condition.

[0022] II. Image Preprocessing and Feature Extraction Please see Figure 2 The image processing unit performs the following preprocessing steps on the acquired foam images: (1) Automatic ROI extraction: Based on edge detection and geometric templates, the foam layer intervals are automatically identified, and the tank walls and non-foam areas are removed to reduce background interference.

[0023] (2) Illumination equalization: Homomorphic filtering is used to perform high-pass filtering in the log domain to suppress local illumination unevenness and enhance details in dark areas.

[0024] (3) Noise reduction: 3×3 median filtering is used to remove random noise and reflective points of bubbles, and to maintain the integrity of the bubble edge.

[0025] (4) Gray-level normalization: linear normalization G_norm=(G-G_min) / (G_max-G_min) or adaptive histogram equalization CLAHE is used to improve contrast, so that the gray levels of images under different lighting conditions are comparable.

[0026] (5) Temporal smoothing: 1-3 point exponential smoothing is used to reduce the impact of inter-frame jitter on feature stability.

[0027] After preprocessing, grayscale dynamic features are extracted, specifically including: Global statistical characteristics: mean gray level G_mean = (1 / N)∑G_i, standard deviation of gray level G_std = √[(1 / N)∑(G_i-G_mean)] 2 ]; Distribution characteristics: Low grayscale proportion P_low = #(G<τ_l) / N, high grayscale proportion P_high = #(G>τ_h) / N, threshold τ is determined adaptively by OTSU or quantile; Texture features: LBP(8,1) uniform mode histogram, GLCM contrast, energy, and entropy; Temporal characteristics: the grayscale change rate ΔG_t = G_mean(t)-G_mean(t-Δt), and its second-order difference Δ 2 G_t = ΔG_t-ΔG_{t-Δt}.

[0028] In addition, the system also calculates an image health index (Laplace variance) to monitor camera focus drift or lens smudges, and automatically alarms when the health index falls below a threshold.

[0029] III. Grade Soft Measurement Prediction Module Please see Figure 3 The grade soft measurement prediction module adopts two optional model structures: Implementation Method A: XGBoost Model Input feature vector X = [G_mean, G_std, P_low, P_high, LBP_hist, GLCM_contrast, ΔG, Δ 2G, optional process variables (slurry concentration, air volume)]. Model structure is as follows: , f_k∈F, where f_k is the k-th regression tree, output To predict grade values.

[0030] Training configuration: Loss function is mean squared error (MSE), objective function... Regular term Ω(f) = γT + ½λ∑_j w_j 2 The complexity of the tree and the weights of the leaf nodes are controlled by γ and λ to prevent overfitting. Tree depth is set to 3-6, learning rate to 0.03-0.15, and number of trees to 100-500. Feature selection and interpretability display are based on Gain or SHAP values ​​to help operators understand the mapping relationship between grayscale features and quality.

[0031] Implementation Method B: CNN Model When the image texture features are complex and the illumination varies greatly, a convolutional neural network model can be used. The input is a 128×128 normalized ROI image. The network structure is as follows: Conv2D(3×3, 32) + ReLU MaxPool(2×2) Conv2D(3×3, 64) + ReLU MaxPool(2×2) Flatten Dense(128) + ReLU + Dropout(0.3) Dense(1) Training configuration: Loss function MSE, optimizer Adam, initial learning rate 1e-3, cosine annealing reduced to 1e-4, Dropout rate 0.2-0.5. Data augmentation strategies include random brightness ±10%, small-angle rotation ±5°, and affine shearing ±3% to improve the model's robustness to changes in illumination and foam morphology. The CNN model can directly learn deep texture features from images and is more sensitive to the micro-texture of pyrite particles.

[0032] Online correction mechanism Please see Figure 1 The system has an online calibration mechanism. Because test results may have a 2-4 hour delay, the system first... The optimal time delay was obtained through cross-correlation analysis with Y_lab. Map the test value Y_lab(t) to The predicted image values ​​at each time point are aligned. The aligned data is stored in a verification cache queue, and the linear bias term of the model is fine-tuned using recursive least squares (RLS), or incremental training is performed using mini-batch replay and distillation loss. The system also includes a drift detection unit: when the mean absolute error (MAE) of the sliding window exceeds a threshold, or when concept drift is detected by the PH test or ADWIN algorithm, full or batch retraining is triggered to ensure that the model maintains high accuracy as the ore properties change.

[0033] IV. Intelligent Setting and Controller Module Please see Figure 4 The intelligent setting and controller module adopts a fuzzy PID controller or model predictive control MPC, and communicates with the PLC controller (5) through OPC DA / Modbus TCP.

[0034] Feedforward control: Prediction window h = 5-15 min, feedforward control variable Where K_f is the feedforward gain coefficient and sat is the saturation function, including rate and amplitude limits. When the predicted grade is lower than the set value, the system increases the amount of collector and frother added in advance; when the predicted grade is higher than the set value, the amount of reagents added is appropriately reduced to achieve proactive control.

[0035] Feedback control (PID): The feedback control quantity u_PID(t) = K_p·e(t) + K_i·∫e(t)dt + K_d·de(t) / dt, where e = Y_lab - G_set, which includes anti-integral saturation and filters to prevent overshoot caused by integral saturation.

[0036] Feedback control (MPC optional): For operating conditions with strong multivariate coupling, model predictive control with soft constraints and moving penalties can be used, with a control cycle of 30-60s, while simultaneously optimizing multiple objectives such as grade and recovery rate.

[0037] Output synthesis: The final control quantity u(t) = w_f·Δu_f + w_b·u_PID / MPC, where the weights w_f and w_b are adaptively adjusted according to the prediction confidence and the test deviation. The feedforward weight is increased when the prediction confidence is high, and the feedback weight is increased when the test deviation is large.

[0038] The actuators include: a collector addition device (6), a foaming agent regulating valve (7), a liquid level regulating valve (8), and a blower (9). The system is equipped with a safety protection mechanism: when the limit is exceeded, mismatch occurs, or the sensor malfunctions, it will trigger a return to a safe operating condition, and all action records are traceable.

[0039] V. Field Application Results (Best Practice) This system was deployed in the final cleaning tank of a pyrite beneficiation plant. Offline modeling was performed for two weeks during the data collection period, using XGBoost as the primary model and CNN as a backup model. The initial validation set MAE was 0.65-0.85%, and R... 2 It is 0.78-0.86.

[0040] After automatic operation, the prediction step size is set to 10 minutes. When the system predicts... When the target grade is 42%, reduce the amount of collector and foaming agent added in advance, and appropriately reduce the blower volume. Results from three consecutive months of operation show: The pass rate for quality products increased from 85% to 88% (an increase of 3 percentage points); The collector consumption per ton of ore decreased from 120g / t to 105g / t (a reduction of 12.5%). Grade volatility (3σ range) decreased by approximately 35%; The frequency of manual intervention has decreased by more than 60%.

[0041] VI. Other Optional Embodiments Example C (Multi-point camera fusion): Industrial grayscale cameras are installed simultaneously in the coarse and fine selection tanks to extract grayscale dynamic features. A stacked model fusion strategy is adopted, with XGBoost as the first-level learner to output coarse and fine selection grade predictions respectively, and linear regression or shallow neural network as the second-level fusioner to output the final grade prediction, further improving prediction accuracy and robustness.

[0042] Example D (Multi-objective joint optimization): Grade and recovery rate are both used as optimization objectives. The recovery rate soft constraint and reagent cost weight are introduced into the MPC objective function to achieve multi-objective joint optimization control, taking into account both economic benefits and product quality.

[0043] Example E (Edge-Cloud Collaboration): The industrial control computer at the edge performs image processing and model inference in real time, while the cloud is responsible for model retraining and policy distribution. It maintains independent local operation when offline and automatically synchronizes and updates model parameters upon reconnection, enabling continuous optimization.

[0044] Example F (Robustness Enhancement): Under strong light disturbance or water mist conditions, a polarized light source or a dual-spectrum (visible light + near-infrared) backup image channel is enabled to ensure stable image quality and guarantee the reliability of grayscale feature extraction.

[0045] VII. Calibration and Maintenance The system has a comprehensive calibration and maintenance mechanism: Initial calibration: Camera-light source-ROI triple calibration to ensure consistency of acquisition parameters; Periodic verification: Illumination / grayscale drift verification is performed every 7-14 days to automatically compensate for environmental changes; Self-test function: Real-time monitoring of cleanliness, focus status, and light source aging, prompting maintenance when abnormalities occur; Software management: Supports version rollback, model versioning, and grayscale-grade curve auditing; Security audit: permission level, operation log recording, and abnormal alarm notification.

[0046] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A sulfur flotation intelligent control system based on the dynamic characteristics of foam grayness, comprising a foam grayness characteristic sensing module, a grade soft measurement and prediction module, and an intelligent setting and controller module, characterized in that: The foam grayscale feature sensing module is used to acquire flotation cell foam images in real time and extract grayscale dynamic features; the grade soft measurement prediction module constructs a soft measurement model based on the grayscale dynamic features to predict the grade of sulfur concentrate in real time; the intelligent setting and controller module executes a feedforward-feedback composite control strategy to automatically adjust the flotation process parameters according to the deviation between the predicted grade and the set value.

2. The intelligent control system for sulfur flotation based on the dynamic characteristics of foam grayscale according to claim 1, characterized in that: The foam grayscale feature perception module includes an image acquisition unit and an image processing unit. The image acquisition unit installs industrial grayscale cameras in the final selection tank and the sulfur coarse selection tank, respectively. The cameras are equipped with an 850nm near-infrared ring light source, the camera parameters are fixed and locked, and the image acquisition frequency is 10-20 FPS. The image processing unit performs ROI region extraction, illumination equalization, median filtering, and grayscale normalization processing.

3. The intelligent control system for sulfur flotation based on the dynamic characteristics of foam grayness according to claim 2, characterized in that: The grayscale dynamic features include global statistical features, distribution features, texture features, and temporal features; the global statistical features include average grayscale and grayscale standard deviation; the distribution features include high grayscale proportion and low grayscale proportion; the texture features include LBP texture features and GLCM contrast; and the temporal features include grayscale change rate and its second-order difference.

4. The intelligent control system for sulfur flotation based on the dynamic characteristics of foam grayness according to claim 3, characterized in that: The grade soft measurement prediction module uses a gradient boosting tree (XGBoost) to construct a soft measurement model, taking gray-scale dynamic features as input vectors and outputting predicted grade values ​​through the superposition of multiple regression trees; or it uses a convolutional neural network (CNN) to construct a soft measurement model, taking a normalized ROI image as input, extracting deep texture features through convolutional layers, pooling layers, and fully connected layers, and outputting predicted grade values.

5. The intelligent control system for sulfur flotation based on the dynamic characteristics of foam grayness according to claim 4, characterized in that: The grade soft measurement prediction module also includes an online correction mechanism; whenever a new test result arrives, the system stores it in the verification cache queue and lags it with the image prediction value at the corresponding time, and fine-tunes the model using recursive least squares (RLS) or incremental training; the system also has a drift detection unit, which triggers model retraining when the mean absolute error of the sliding window exceeds the threshold.

6. The intelligent control system for sulfur flotation based on the dynamic characteristics of foam grayness according to claim 1, characterized in that: The intelligent setting and controller module adopts a fuzzy PID controller or model predictive control (MPC); the feedforward-feedback composite control strategy includes: feedforward control adjusts the control quantity in advance based on the deviation between the predicted grade and the set value, feedback control performs closed-loop adjustment based on the deviation between the true test value and the set value, and the final control quantity is output to the actuator by weighted synthesis of the feedforward control quantity and the feedback control quantity.

7. The intelligent control system for sulfur flotation based on the dynamic characteristics of foam grayness according to claim 6, characterized in that: The process parameters include the frequency of collector addition, the opening degree of the foaming agent regulating valve, the opening degree of the liquid level valve, and the air volume.

8. A method for intelligent control of sulfur flotation based on the dynamic characteristics of foam grayness, comprising an intelligent control system for sulfur flotation based on the dynamic characteristics of foam grayness, characterized in that, Includes the following steps: S1: Real-time acquisition of foam images in the flotation cell; S2: Preprocess the acquired foam image, extract the ROI region and perform illumination equalization, noise reduction and grayscale normalization; S3: Extract grayscale dynamic features from the preprocessed image; S4: Input the extracted grayscale dynamic features into a pre-trained soft measurement model to predict the grade of sulfur concentrate in real time; S5: Compare the predicted grade with the set value, execute feedforward-feedback composite control, and adjust the flotation process parameters; S6: When the test results arrive, perform online calibration of the soft measurement model.

9. The intelligent control method for sulfur flotation based on the dynamic characteristics of foam grayness according to claim 8, characterized in that: The grayscale dynamic features in step S3 include average grayscale, grayscale standard deviation, high grayscale ratio, low grayscale ratio, LBP texture features, GLCM contrast, and grayscale change rate.

10. The intelligent control method for sulfur flotation based on the dynamic characteristics of foam grayscale according to claim 8, characterized in that: The method for constructing the soft measurement model in step S4 is as follows: collect historical foam images and corresponding laboratory grade data, extract grayscale dynamic features to construct a training sample set, train it using the XGBoost algorithm or CNN algorithm, use mean squared error as the loss function, and optimize the hyperparameters through cross-validation.

Images