Intelligent flotation tank control method and system integrated with online detection function

Abstract

The application relates to the technical field of mineral processing automation, in particular to an intelligent flotation tank control method and system integrated with an online detection function, which comprises the following steps: synchronously collecting multi-dimensional original production data including a surface foam image stream, preprocessing the surface foam image stream, generating enhanced foam images eliminating high-frequency noise interference, inputting the enhanced foam images into a lightweight convolutional neural network, extracting and quantifying a dynamic visual feature vector, constructing a working condition feature tensor with space-time correlation characteristics, inputting the working condition feature tensor into a long short-term memory neural network model in a time sequence, predicting a tailing grade fluctuation value within a preset time step in the future, and calculating a cooperative control instruction in response to the discrete degree of the tailing grade fluctuation value and a preset process index. The application can capture digital fingerprints reflecting deep changes in the flotation working condition, and has higher sensitivity and robustness.

Description

Technical Field

[0001] This application relates to the field of mineral processing automation technology, and in particular to an intelligent flotation cell control method and system with integrated online detection function. Background Technology

[0002] Flotation, as a mineral processing technology that utilizes the differences in the physicochemical properties of mineral surfaces to achieve separation, plays an irreplaceable role in modern metallurgy and mineral processing. As the core equipment of the process, the stability of the flotation cell's operation directly affects the ore recovery rate and concentrate grade.

[0003] With the advancement of intelligent mine construction, how to achieve accurate perception and real-time closed-loop control of the flotation process has become a hot topic of concern in the industry. A flotation machine involving online detection and automatic control of slurry level uses mechanical floats in conjunction with electrically controlled valves to achieve single-loop regulation based on physical liquid level. In addition, a flotation reagent detection system based on machine vision analyzes the grayscale and light intensity information of foam images through neural networks to calculate the ash content of tailings.

[0004] However, existing flotation control technologies still have significant shortcomings in practical applications. First, in highly corrosive and scaling slurry environments, mechanical float-based level detection methods are prone to float jamming or transmission mechanism failure, leading to distorted detection data and extremely high equipment maintenance frequency. Second, most existing visual detection schemes are limited to simple descriptions of static image features, lacking in-depth modeling of the dynamic evolution of foam (such as merging rate and flow velocity vector), making it difficult for the system to achieve high-precision grade prediction in complex industrial environments. Finally, existing control logic often ignores the strong coupling characteristics between air intake, liquid level, and reagent addition, resulting in severe lag and overshoot in the adjustment process when facing fluctuations in feed properties, making it difficult to maintain continuous and stable production. Summary of the Invention

[0005] The objective of this application is to provide an intelligent flotation cell control method integrating online detection functionality, comprising: simultaneously acquiring multi-dimensional raw production data, including surface foam image stream, real-time slurry level height, instantaneous flow rate of the air inlet branch pipe, and slurry inlet mass percentage, through a multi-source sensing array deployed above the flotation cell and electrically connected to an industrial control electromechanical system; preprocessing the surface foam image stream using median filtering and adaptive contrast enhancement algorithms to generate an enhanced foam image that eliminates high-frequency noise interference; inputting the enhanced foam image into a pre-trained lightweight convolutional neural network based on an attention mechanism, performing spatial feature convolution and channel attention reconstruction to extract and quantify dynamic visual representation vectors, including average bubble diameter, foam layer load thickness, bubble merging rate, and surface foam flow velocity; and calculating the enhanced foam image in... The gradient co-occurrence matrices in the horizontal and vertical directions are used to obtain frequency domain texture features reflecting foam roughness and directionality. Principal component analysis is then used to fuse the dynamic visual representation vector with these frequency domain texture features in a high-dimensional manner, constructing a working condition feature tensor with spatiotemporal correlation. This working condition feature tensor is input into a long short-term memory neural network model in time series, and combined with historical working condition evolution trajectories, the tailings grade fluctuation value within a preset time step is predicted. In response to the dispersion of the tailings grade fluctuation value from preset process indicators, a collaborative control command is calculated based on an incremental proportional-integral-derivative control strategy, including the adjustment step size of the frother pumping frequency, the lifting and lowering displacement step size of the slurry outlet weir, and the adjustment step size of the aeration regulating valve opening. The collaborative control command is then output to drive the corresponding actuator, achieving closed-loop adaptive control of the flotation process in the flotation cell.

[0006] By adopting the above technical solution, and by simultaneously acquiring multi-source data such as images, liquid level, flow rate, and mass percentage, and using a deep learning model to extract dynamic visual representations, this application can capture digital fingerprints that reflect deep changes in flotation conditions, exhibiting higher sensitivity and robustness. In particular, the introduction of long short-term memory neural networks endows the system with the ability to predict the future, enabling the generation of collaborative control commands to anticipate grade fluctuations in advance, effectively solving the problem of large lag in the flotation process.

[0007] Optionally, the preprocessing of the surface foam image stream includes, but is not limited to: performing edge-preserving denoising processing on the surface foam image stream using a bilateral filtering algorithm, wherein the standard deviation of the spatial kernel function and the standard deviation of the value domain kernel function of the bilateral filtering algorithm are dynamically configured based on the local variance of the image; extracting the foam edge skeleton by calculating the second-order Laplacian operator of the denoised image; and superimposing the foam edge skeleton onto the original image according to a preset weight to generate the enhanced foam image with a high-contrast contour.

[0008] By adopting the above technical solution, and superimposing bilateral filtering with the Laplace operator, the visual contrast of the foam edge is significantly enhanced, and key bubble contour information can be preserved even in extreme cases such as uneven lighting or acid mist obscuring the flotation site.

[0009] Optionally, the process of extracting the dynamic visual representation vector includes, but is not limited to: introducing a squeezing and activation module into the feature map output layer of the lightweight convolutional neural network; wherein, the squeezing and activation module compresses spatial dimension information through global average pooling and generates importance weights for channel dimensions using a two-layer fully connected network; and reconstructs the feature map based on the importance weights to enhance the feature response of the mineralized foam region and suppress background interference.

[0010] By adopting the above technical solution and introducing the squeezing and excitation modules, the working principle of their collaboration lies in achieving automatic recalibration at the feature level by learning the important correlation between channels. This enables the model to spontaneously focus on the effective foam features rich in minerals and suppress invalid responses caused by water surface reflection or background noise. Thus, even in industrial control environments with limited computing resources, it can still maintain excellent recognition performance.

[0011] Optionally, the process of calculating the cooperative control command includes, but is not limited to: constructing a nonlinear multi-objective functional with the optimization objectives of maximizing production recovery rate and minimizing reagent consumption; using an adaptive genetic algorithm to perform global optimization on the nonlinear multi-objective functional to obtain a set of Pareto optimal solutions; and selecting from the Pareto optimal solution set the solution with the highest matching degree with the current slurry inlet mass percentage as the parameter benchmark for calculating the cooperative control command.

[0012] By adopting the above technical solution and optimizing the multi-objective functional, the optimal balance between recovery rate and economy is achieved. The original discrete and mutually disturbing control variables (drug dosage, gas volume, displacement) can be transformed into a globally optimal action sequence with consistent steps, thus completely eliminating the variable oscillation phenomenon commonly found in conventional PID control.

[0013] Optionally, it also includes: using a sensor window monitoring unit installed at the front end of the multi-source sensing array to collect window transmittance data in real time; and in response to the window transmittance data being lower than a preset cleanliness threshold, generating a pulsed high-pressure air curtain driving signal to remove mineral slurry splashes and acidic droplets attached to the sensor window.

[0014] By adopting the above technical solution, the negative impact of the industrial environment on the lifespan of the sensor is solved. Through the linkage between sensor window monitoring and air curtain cleaning drive signal, the system has online self-maintenance capability, ensuring that the system can operate stably for a long time without manual intervention and reducing the false alarm rate caused by sensor contamination.

[0015] Optionally, the process of identifying the surface foam flow velocity includes, but is not limited to: extracting scale-invariant feature transform operators in two adjacent frames of the enhanced foam images; performing feature matching on the scale-invariant feature transform operators using a random sampling consensus algorithm to eliminate mismatched point pairs; calculating the Euclidean distance between successfully matched feature point pairs, and calculating the instantaneous motion displacement vector of the foam layer in conjunction with the camera intrinsic and extrinsic parameter array.

[0016] Optionally, the process of quantifying the bubble merging rate includes, but is not limited to: segmenting the enhanced bubble image using a watershed algorithm to identify the closed contours of individual bubbles; counting the number of closed contours of individual bubbles that disappear per unit time; wherein the number of disappearing bubbles is positively correlated with the average brightness change rate of the corresponding region.

[0017] By adopting the above technical solution, the foam characteristics are refined from two dimensions: spatial displacement and time evolution. The quantification of flow velocity vector and bubble merging rate provides a direct basis for judging the dynamic intensity in the tank.

[0018] Optionally, the construction process of the long short-term memory neural network model includes, but is not limited to: dividing the continuous working condition feature tensor into multiple groups of time series samples using the sliding window technique; introducing forget gate logic in the neuron nodes to discard long-term redundant information in the historical sequence that is less than a preset value in relation to the current tailings grade fluctuation; and using the error backpropagation algorithm combined with an adaptive learning rate optimizer to iteratively update the model parameters.

[0019] By adopting the above technical solution, the time series modeling is optimized through forget gate logic. Combined with the above working condition feature tensor, it can automatically filter out the historical information fragments that are most valuable for the current prediction, which significantly improves the model's adaptability when faced with sudden changes in ore hardness or grade.

[0020] Optionally, the process of outputting the coordinated control command includes, but is not limited to: determining whether the current mechanical displacement of each actuator has reached the physical limit position; if it has, then performing truncation processing on the corresponding adjustment step size and reallocating the compensation weights of the remaining control variables to prevent the actuator from being overloaded or damaged.

[0021] By adopting the above technical solution, limit position verification and weight redistribution logic are introduced, which can ensure that even if individual actuators reach the upper limit of their stroke during complex closed-loop adjustment, the system can still maintain the achievement of the quality target by adjusting other redundant variables, thus avoiding forced damage to the mechanical structure.

[0022] The second objective of this application is to provide an intelligent flotation cell control system with integrated online detection function, including a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed as described above for the intelligent flotation cell control method with integrated online detection function. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the intelligent flotation cell control method integrating online detection function in this application. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0025] This application provides a smart flotation cell control method and system with integrated online detection function to address the technical bottlenecks existing in the flotation process. Specifically, in the operation of flotation cells in the mineral processing industry, due to the complex source of ore and frequent fluctuations in feed properties, strong nonlinear coupling relationships exist between multiple key process parameters such as slurry level, foam state, reagent reaction, and aeration rate. Traditional methods based on single mechanical float level detection or manual observation of foam are difficult to achieve accurate and real-time closed-loop control. Specifically, in highly corrosive and high-humidity slurry environments, mechanical float level sensors are prone to jamming due to slurry scaling or acid corrosion, leading to distorted level detection data and consequently causing gate control lag or malfunction. Relying on manual visual inspection of foam status at regular intervals results in significant subjective judgment variations and slow response times. It is impossible to quantify and capture dynamic evolution characteristics such as foam layer thickness, bubble merging rate, and surface flow velocity, making it difficult to predict tailings grade fluctuations in advance. Existing automatic control logic mostly adopts single-loop proportional-integral-derivative control strategies, which cannot coordinate and handle the mutual interference between frother dosage, slurry outlet weir height, and aeration regulating valve opening. When dealing with sudden changes in ore grade, the system adjustment process experiences severe overshoot and oscillation, leading to a decrease in concentrate recovery rate and waste of flotation reagents. In addition, slurry splashes and acidic droplets in industrial sites easily contaminate the observation windows of optical sensors, causing the visual inspection system to fail rapidly without manual cleaning intervention, resulting in insufficient long-term reliability of the system.

[0026] like Figure 1 As shown, in order to solve the above-mentioned technical problems, this specific embodiment discloses an intelligent flotation cell control method with integrated online detection function, including the following steps.

[0027] S01: Through a multi-source sensing array deployed above the flotation cell and electrically connected to the industrial control electromechanical system, multi-dimensional raw production data including surface foam image stream, real-time slurry level height, instantaneous flow rate of the air inlet branch pipe, and slurry inlet mass percentage are collected simultaneously.

[0028] Understandably, the deployment method of a multi-source sensing array may specifically include: fixing a protective enclosure made of corrosion-resistant alloy to a rigid bracket on top of the flotation cell; mounting a high-resolution industrial area array camera as a vision sensor on the front of the protective enclosure; the camera being equipped with a specific model of complementary metal-oxide-semiconductor sensor with 5 million effective pixels, a maximum resolution of 2592 pixels by 1944 pixels, and a working frame rate set to 30 frames per second to meet the needs of capturing dynamic changes in foam; and using a 12mm fixed-focus industrial lens for the camera lens, along with a ring... The LED supplemental lighting has a color temperature of 5600 Kelvin and a color rendering index higher than 90, providing a stable and uniform lighting environment. The overall protection level of the protective enclosure meets the international protection standard level 67, capable of withstanding ambient temperature fluctuations within a range of -20 degrees Celsius to 70 degrees Celsius and the erosion caused by slurry splashes. A non-contact ultrasonic level gauge is installed on the side wall of the flotation cell, 1.2 meters above the bottom of the cell. Its measuring range is 0 to 2.5 meters, the output is a 4 to 20 mA analog signal, and the measuring accuracy is ±1 mm. This ultrasonic level gauge is connected to the cell wall via a flange. The probe surface is coated with a polytetrafluoroethylene anti-adhesion coating to prevent slurry scaling from affecting acoustic wave transmission and reception. A thermal gas mass flow meter is installed at the connection between the main air inlet pipe and each branch pipe of the flotation cell. Its flow rate measurement range is set from 0 to 100 standard cubic meters per hour, based on the flotation cell's designed aeration rate, and its output signal is also 4 to 20 mA. This gas mass flow meter has a self-cleaning function, automatically removing minor dust accumulation on the probe in the airflow. An online slurry concentration meter and a grade analyzer are installed on the slurry inlet pipe of the flotation cell. The online slurry concentration meter can use radiometric density... The flotation cell can be equipped with an ultrasonic concentration meter or an online X-ray fluorescence analyzer, with a sampling cycle set to acquire the mass percentage data of the target metal element in the slurry every five minutes. This data is transmitted to the industrial control computer via industrial Ethernet. The power supply and signal cables of all the above sensors are connected to the industrial control computer located in the control cabinet of the flotation cell through armored cables. The industrial control computer adopts an edge general-purpose controller with an integrated high-performance processor, large-capacity memory and dedicated graphics processing module, providing a computing power of up to 275 trillion operations per second for running subsequent image processing and deep learning algorithms.

[0029] Specifically, the synchronous acquisition process of the surface foam image stream is as follows: The industrial control computer continuously reads the raw foam images captured by the industrial camera at a rate of 30 frames per second through the gigabit Ethernet interface. Each frame is in red-green-blue color format with a resolution of 2592 x 1944 pixels. At the same time, the industrial control computer synchronously acquires the analog signal of liquid level height from the ultrasonic level gauge and the analog signal of instantaneous flow rate from the thermal gas mass flow meter through its analog input module at a sampling frequency of 100 Hz. It also reads the latest slurry inlet mass percentage value from the online grade analyzer through a specific industrial communication protocol. The industrial control computer runs a data synchronization thread, which uses the camera frame trigger signal as the time reference to align and package the image data, liquid level data, flow rate data, and the most recently updated grade data at the same timestamp, forming a data packet containing multi-dimensional raw production data. The time synchronization error of this data packet is less than 10 milliseconds.

[0030] S02: The surface foam image stream is preprocessed using median filtering and adaptive contrast enhancement algorithms to generate an enhanced foam image that eliminates high-frequency noise interference.

[0031] In this embodiment, the preprocessing of the surface foam image stream using median filtering and adaptive contrast enhancement algorithms includes, but is not limited to: after receiving the original foam image, the industrial control computer first converts it from the red-green-blue color space to a grayscale image to simplify subsequent processing; median filtering is performed on the grayscale image, with the filtering window size set to 5 pixels by 5 pixels. This operation can effectively filter out high-frequency salt-and-pepper noise caused by slurry splashing or camera sensor thermal noise, while better preserving the edge information of the foam; after denoising, adaptive contrast enhancement is performed. The specific algorithm can be a limited contrast adaptive histogram equalization algorithm, which divides the image into multiple local regions of 8 pixels by 8 pixels, calculates its grayscale histogram in each region, and performs cropping restrictions on the histogram. The cropping limit coefficient is set to 0.2 to prevent excessive enhancement of local contrast and the generation of unnatural artificial traces. Then, histogram equalization is performed on each region, and boundary artifacts between regions are eliminated by bilinear interpolation. Finally, an enhanced foam image with uniform global contrast and prominent local details is generated, and the grayscale difference between the foam and the background in the enhanced foam image is significantly improved.

[0032] S03: Input the enhanced foam image into a pre-trained lightweight convolutional neural network based on an attention mechanism, perform spatial feature convolution and channel attention reconstruction, and extract and quantify dynamic visual representation vectors including average bubble diameter, foam layer load thickness, bubble merging rate, and surface foam flow velocity.

[0033] Understandably, the process of inputting an enhanced foam image into a pre-trained attention-based lightweight convolutional neural network includes, but is not limited to: the lightweight convolutional neural network employs a specific lightweight network architecture as its base network and adaptively modifies it to suit the task of extracting features from floating foam; the improved network input layer receives an enhanced foam grayscale image with a size adjusted to 224 pixels by 224 pixels, and the network front end includes convolutional layers with a kernel size of 3 by 3 and a stride of 2 for initial extraction of low-level features; subsequently, a series of inverse residual structures are connected, each structure employing depthwise separable convolutions to reduce the number of computational parameters, and a squeezing and activation attention module is introduced after the depthwise convolutions; the squeezing and activation module has specific... The core operation involves global average pooling of the feature map, compressing the spatial information of each channel into a scalar. Then, two fully connected layers are used to learn the correlation between channels. The first fully connected layer compresses the number of channels to one-sixteenth of the original number and uses the rectified linear unit activation function. The second fully connected layer restores the number of channels to the original number and uses the sigmoid growth curve activation function to generate an importance weight vector for each channel. Finally, this weight vector is multiplied with the original feature map channel by channel to complete feature recalibration. This allows the network to automatically focus on the features of mineralized foam regions in the image and suppress interference from water reflections or shadow backgrounds. The network ultimately outputs a 512-dimensional feature vector, which encodes the global semantic information of the image.

[0034] Understandably, the process of extracting and quantifying dynamic visual representation vectors, including average bubble diameter, foam layer load thickness, bubble merging rate, and surface foam flow velocity, includes, but is not limited to: on a 512-dimensional feature vector extracted by a lightweight convolutional neural network, multiple dedicated regression head network branches are connected in parallel to quantify and calculate each dynamic visual indicator; for the average bubble diameter, a regression branch with two fully connected layers is designed, the first fully connected layer having 256 neurons and the second fully connected layer having one neuron. The input to this branch is the 512-dimensional feature vector extracted by the network, and the output is... The scalar value, physically representing the arithmetic mean of the equivalent circular diameters of all bubbles within the image field of view, is measured in millimeters, typically ranging from 0.5 mm to 5 mm. For the foam layer load thickness, another regression branch is designed, similar in structure to the bubble diameter branch, but its output scalar represents the vertical distance from the foam layer surface to the slurry surface. This load thickness is calculated using image perspective geometry combined with camera calibration parameters, also measured in millimeters, typically ranging from 50 mm to 300 mm. For the bubble merging rate, a time-series regression branch is designed, with its input... The algorithm includes not only the 512-dimensional feature vector of the current frame but also the feature vectors of the previous five consecutive frames. These feature vectors are stacked in the time dimension and then input into a small recurrent neural network (RNN) containing long short-term memory (LSM) units. This RNN includes an LSM layer with 64 hidden units, followed by a fully connected layer. The final output is a scalar representing the percentage of bubbles that disappear due to merging within the current field of view per unit time (typically one second) out of the total number of bubbles. A typical value ranges from 0.1% to 5% per second. An optical flow calculation regression branch is designed for the surface foam flow velocity. This branch takes the enhanced foam images of the current frame and the previous frame as input, and uses the Lucas-Cannard sparse optical flow algorithm to calculate the displacement vectors of 100 pre-defined feature points in the image. The feature points are selected in the foam image through a specific corner detection algorithm. Then, the average magnitude of the displacement vectors of all feature points is calculated, and combined with the camera frame rate and spatial calibration coefficient (millimeters per pixel), it is converted into the average motion speed of the foam surface, measured in millimeters per second, with a typical value range of 10 millimeters per second to 200 millimeters per second. Finally, the output values ​​of the above four regression branches are combined into a four-dimensional dynamic visual representation vector.

[0035] S04: By calculating the gradient co-occurrence matrix of the enhanced foam image in the horizontal and vertical directions, the frequency domain texture features reflecting the roughness and directionality of the foam are obtained. The principal component analysis method is used to fuse the dynamic visual representation vector and the frequency domain texture features in a high dimension to construct a working condition feature tensor with spatiotemporal correlation characteristics.

[0036] Understandably, the process of obtaining frequency domain texture features reflecting foam roughness and directionality by calculating the gradient co-occurrence matrices of the enhanced foam image in the horizontal and vertical directions includes, but is not limited to: calculating the first-order gradients of the enhanced foam grayscale image in the horizontal and vertical directions respectively, using the Sobel operator with a convolution kernel size of 3 x 3; and constructing a gradient co-occurrence matrix based on the calculated horizontal and vertical gradient images, where the gradient co-occurrence matrix has a dimension of 256 x 256, and the row index of the matrix represents the gray level of the horizontal gradient image, and the column index represents the gray level of the vertical gradient image. The gradient co-occurrence matrix is ​​a matrix where each element represents the probability that a pair of pixels satisfying a specific horizontal and vertical gradient value will appear simultaneously in the image under a set spatial relationship (e.g., pixel spacing of 1 and direction of 0 degrees). A set of statistics is extracted from the constructed gradient co-occurrence matrix as frequency domain texture features, including energy, contrast, correlation, and homogeneity. Energy reflects the uniformity of the gradient distribution in the image, contrast reflects the sharpness of the gradient image, correlation reflects the linear dependence between the horizontal and vertical gradients, and homogeneity reflects the uniformity of local changes in the gradient image. Finally, a four-dimensional frequency domain texture feature vector is obtained.

[0037] Understandably, the process of using principal component analysis to fuse dynamic visual representation vectors and frequency domain texture features in a high dimension to construct a working condition feature tensor with spatiotemporal correlation characteristics includes, but is not limited to: concatenating the four-dimensional dynamic visual representation vector with the four-dimensional frequency domain texture feature vector to form an initial eight-dimensional fused feature vector; to eliminate differences in the dimensions and numerical ranges between different feature dimensions, standardizing the eight-dimensional fused feature vector by subtracting the mean of each dimension obtained from historical data statistics and dividing by the standard deviation; and preparing a historical dataset including 1000 consecutive production batches from the past, where each sample is standardized. Eight-dimensional fusion feature vectors are generated. Principal component analysis is performed on the historical dataset to calculate its covariance matrix and perform eigenvalue decomposition. The top 5 principal component directions with a cumulative contribution rate exceeding 95% are selected. Each new standardized eight-dimensional fusion feature vector is projected onto these 5 principal component directions to obtain the dimensionality-reduced five-dimensional principal component feature vectors. To introduce time dimension information, the five-dimensional principal component feature vectors of 20 consecutive time steps (corresponding to 20 frames of images, with a time span of approximately 0.6 seconds) are stacked in chronological order to form a 20 x 5 two-dimensional matrix. This two-dimensional matrix is ​​the working condition feature tensor with spatiotemporal correlation characteristics.

[0038] S05: Input the working condition feature tensor into the long short-term memory neural network model in time series, and combine it with the historical working condition evolution trajectory to predict the tailings grade fluctuation value within the future preset time step.

[0039] Understandably, the Long Short-Term Memory (LSTM) neural network model includes an input layer, two stacked LSM layers, and a fully connected output layer. The input layer receives a 20x5 tensor of working condition features, treating it as a time series of length 20, with a feature dimension of 5 for each time step. The first LSM layer contains 128 hidden units, including an input gate, a forget gate, an output gate, and cell states. The forget gate uses an S-shaped growth curve activation function, with an output value between 0 and 1, controlling the retention of historical cell state information. When the correlation between a feature at a historical time step and the current prediction target (tailings grade fluctuation) is lower than a preset threshold (e.g., 0.3), the forget gate outputs a value close to zero, discarding the redundant information. The second LSM layer contains 64 hidden units for further extraction of high-level temporal features. A randomized control mechanism is used between the two LSM layers. Machine dropout regularization is applied with a dropout rate of 0.2 to prevent overfitting. The output of the last time step passes through a fully connected layer with 32 neurons, and finally through a linear output layer to generate a scalar prediction value. This scalar prediction value is the predicted tailings grade fluctuation value at the 5th time step (approximately 0.17 seconds later). The tailings grade fluctuation value is defined as the difference between the current tailings grade and the target grade, measured as a percentage by mass. The training data for the model comes from historical production processes, collecting sample pairs including working condition feature tensors as input and the actual tailings grade fluctuation value at the 5th time step after the corresponding time point as labels, with a total of no less than 100,000 samples. The backpropagation algorithm combined with an adaptive moment estimation optimizer is used to iteratively update the model parameters. The initial learning rate is set to 0.001, the batch size is 64, the training epochs are 100, and the loss function is mean squared error.

[0040] S06: In response to the dispersion of tailings grade fluctuations from preset process parameters, the system calculates coordinated control commands based on an incremental proportional-integral-derivative control strategy, including the step size for adjusting the frother pumping frequency, the step size for the lifting and lowering displacement of the slurry outlet weir, and the step size for adjusting the opening of the aeration regulating valve.

[0041] Understandably, the preset process parameters include the target tailings grade value, which is set according to the ore type and production plan. For example, for the flotation of a copper ore, the target tailings grade can be set to 0.15% copper. The dispersion is calculated as the absolute value of the predicted tailings grade fluctuation. Based on this dispersion, an incremental proportional-integral-derivative (PID) control strategy is used to calculate the adjustment step size of the three control variables. For the frother pumping frequency adjustment step size, the calculation formula is: the frequency adjustment step size equals the proportional coefficient multiplied by the error value at the current moment, plus the integral coefficient multiplied by the cumulative error, plus the derivative coefficient multiplied by the difference between the current error and the error at the previous moment. The unit of the frequency adjustment step size is Hertz, the error value at the current moment is the tailings grade fluctuation value, and the proportional coefficient, integral coefficient, and derivative coefficient are obtained through on-site tuning. For example, the proportional coefficient... The integral coefficient is 0.1 Hz per 100% and the derivative coefficient is 0.05 Hz per 100% per second. For the lifting displacement step of the slurry outlet weir, the same control law is used, but a different set of tuning coefficients is employed, for example, a proportional coefficient of 0.2 mm per 100% and an integral coefficient of 0.05 mm per 100% per second, with a derivative coefficient of 0.01 mm per 100% per second. For the opening adjustment step of the air-filling regulating valve, the same structure is used, with coefficients such as a proportional coefficient of 0.3 opening percentage per 100% and an integral coefficient of 0.08 opening percentage per 100% per second, with a derivative coefficient of 0.02 opening percentage per 100% per second. All three control loops have an anti-integral saturation limit for their integral terms; when the actuator reaches its physical limit, the integral accumulation stops.

[0042] S07: Outputs coordinated control commands to drive the corresponding actuators, realizing closed-loop adaptive control of the flotation process in the flotation cell.

[0043] Understandably, the industrial control computer sends the calculated coordinated control commands to the corresponding actuators through its digital and analog output modules; the frother pump frequency adjustment step size is output to the frequency converter via an analog signal of 4 to 20 mA, driving the motor of the frother dosing pump to change the reagent addition rate; the slurry outlet weir plate lifting displacement step size is output to the stepper motor driver via a pulse signal, driving the stepper motor of the weir plate lifting mechanism, with each pulse corresponding to a displacement of 0.01 mm; the air regulating valve opening adjustment step size is output to the electric positioner via another analog signal of 4 to 20 mA, driving the valve core position of the pneumatic diaphragm regulating valve to change the flow area of ​​the air inlet branch pipe; the actual action feedback signals of all the above actuators (such as the actual motor speed, the actual weir plate position, and the actual valve opening) are returned to the industrial control computer through sensors, forming a closed-loop control loop; the industrial control computer repeatedly executes the entire process from data acquisition, feature extraction, grade prediction to control command calculation and output at a frequency of ten times per second, thereby realizing real-time closed-loop adaptive control of the flotation process in the flotation cell.

[0044] In other embodiments of this application, the preprocessing of the surface foam image stream may further include: performing edge-preserving denoising processing on the surface foam image stream using a bilateral filtering algorithm, wherein the standard deviation of the spatial kernel function and the standard deviation of the value domain kernel function of the bilateral filtering algorithm are dynamically configured based on the local variance of the image; extracting the foam edge skeleton by calculating the second-order Laplacian operator of the denoised image; and superimposing the foam edge skeleton onto the original image according to a preset weight to generate an enhanced foam image with high-contrast contours.

[0045] Understandably, the specific process of performing edge-preserving denoising on a surface foam image stream using a bilateral filtering algorithm includes, but is not limited to: the bilateral filtering algorithm simultaneously considers the spatial proximity and gray-level similarity between pixels; its spatial kernel function adopts a Gaussian function form, and its range kernel function also adopts a Gaussian function form; the standard deviation of the spatial kernel function is dynamically adjusted according to the noise level of the local region of the image, and is calculated as follows: first, the image is divided into multiple 16-pixel multiplied 16-pixel blocks, and the variance of the pixel values ​​within each block is calculated; then the standard deviation corresponding to that block is equal to 0.5 plus 0.1. Multiplying by (one plus the natural logarithm of the variance), its value range is limited to between 0.5 pixels and 3 pixels; the standard deviation of the range kernel function is dynamically set according to the overall contrast of the image, and the calculation formula is that the standard deviation is equal to 0.1 multiplied by (the maximum gray value of the image minus the minimum gray value of the image), and its typical value range is between 5 and 30 gray levels; using a sliding window of size 7 pixels by 7 pixels, bilateral filtering is performed on each pixel according to the above dynamically calculated standard deviation. This processing can smooth the noise areas in the image while sharply preserving the edge between the foam and the background.

[0046] Understandably, the process of extracting the foam edge skeleton by calculating the second-order Laplacian operator of the denoised image includes, but is not limited to: applying the Laplacian operator to the bilaterally filtered image, which uses a 3-pixel multiplied 3-pixel convolution kernel with a center value of -4, a value of 1 at the four adjacent positions (top, bottom, left, and right), and a value of 0 at the four diagonal positions; performing a convolution operation between the kernel and the image to obtain the Laplacian response image, where the zero-crossing points correspond to the edges of the original image; binarizing the Laplacian response image with a threshold set to 1.5 times the mean of the absolute values ​​of the response to obtain a preliminary set of edge points; and thinning the binary edge image using a Zhang-Sun parallel thinning algorithm, iteratively deleting boundary pixels that meet specific template conditions until the edge line width becomes a single pixel, ultimately obtaining the skeleton image of the foam edge.

[0047] Understandably, the process of superimposing the foam edge skeleton onto the original image with a preset weight to generate an enhanced foam image with a high-contrast contour includes, but is not limited to: setting the preset weight to 0.7; superimposing the foam edge skeleton image (a binary image with edge point values ​​of 255 and background values ​​of 0) onto the original grayscale image, with the superposition formula being: the enhanced image equals the original image plus the preset weight multiplied by the foam edge skeleton image; to ensure that the grayscale values ​​of the superimposed image do not overflow (not exceeding 255), the original image is linearly stretched before superposition, and its grayscale range is adjusted to 0 to 200; after superposition, the enhanced image is again subjected to contrast-limited adaptive histogram equalization to further improve the overall contrast. In the final enhanced foam image, the contour of each bubble is significantly brightened, forming a sharp contrast with the dark background.

[0048] In the embodiments of this application, the process of extracting dynamic visual representation vectors includes, but is not limited to: introducing a squeezing and activation module into the feature map output layer of a lightweight convolutional neural network; wherein, the squeezing and activation module compresses spatial dimension information through global average pooling and generates importance weights for channel dimensions using a two-layer fully connected network; and reconstructs the feature map based on the importance weights to enhance the feature response of the mineralized foam region and suppress background interference.

[0049] Specifically, assuming the final feature map output by the lightweight convolutional neural network is sized as height multiplied by width multiplied by the number of channels, where height and width are spatial dimensions, and the number of channels is the number of channels in the feature map (typically 7 x 7 x 512), the compression operation is implemented through global average pooling. This involves averaging the height multiplied by the width of each channel's feature values ​​to obtain a vector representing the number of channels. This operation compresses global spatial information into the descriptor of each channel. The activation operation is implemented through two fully connected layers. The first fully connected layer compresses the channel-dimensional vector to the original number of channels divided by 16, with a compression ratio of 16. This layer uses rectified linear units (RCUs) as the activation function. The second fully connected layer restores the dimension to the original number of channels and uses an S-curve activation function to output the desired feature map. The importance weight vector is a channel-number dimension, where each element has a value between 0 and 1, representing the importance of the corresponding channel. The weighted reconstruction operation multiplies this channel-number dimension weight vector with the original feature map channel by channel, that is, multiplying all spatial location feature values ​​of each channel by the corresponding weight coefficient. After this operation, the network automatically learns that feature channels that contribute significantly to the identification of mineralized foam (usually manifested as specific colors and textures) (e.g., channels encoding a specific yellow tone) will receive larger weights close to 1 and be activated, while channels that respond to background interference (such as water surface highlights) will receive smaller weights close to 0 and be suppressed. This allows the model to adaptively focus on key regions in the image without the need for manual feature design, significantly improving the robustness and specificity of feature extraction.

[0050] In the embodiments of this application, the process of calculating the cooperative control command includes, but is not limited to: constructing a nonlinear multi-objective functional with the optimization objectives of maximizing production recovery rate and minimizing reagent consumption; using an adaptive genetic algorithm to perform global optimization on the nonlinear multi-objective functional to obtain a set of Pareto optimal solutions; and selecting the solution with the highest matching degree with the current slurry inlet mass percentage from the Pareto optimal solution set as the parameter benchmark for calculating the cooperative control command.

[0051] Specifically, a decision variable vector is defined, comprising nine control parameters of three incremental proportional-integral-derivative (PID) control loops: the proportional coefficient, integral coefficient, and derivative coefficient of each loop. The first objective function is the production recovery rate, calculated using a support vector machine (SVM) regression model trained on historical data. This SVM model takes the decision variable vector and the current operating condition feature tensor as input and outputs the predicted recovery rate. The recovery rate is defined as the ratio of the mass of useful metal in the concentrate to the mass of useful metal in the feed, expressed as a percentage. The objective is to maximize this objective function. The second objective function is reagent consumption, calculated using another neural network model. This neural network model predicts the amount of frother consumed per unit time, in kilograms per hour, using the same input. The objective is to minimize this objective function. Therefore, the multi-objective optimization problem is formalized as: maximizing the first objective function and minimizing the second objective function. Simultaneously, each component of the decision variable vector must fall within a preset lower and upper limit. For example, the lower limit for the proportional coefficient is 0.01, and the upper limit is 2; the lower limit for the integral coefficient is 0.001, and the upper limit is 5; and the lower limit for the derivative coefficient is 0, and the upper limit is 0.1.

[0052] The genetic algorithm population size can be set to 100, and individual encoding uses real numbers, with each individual being a nine-dimensional decision variable vector. The initial population is randomly generated within the upper and lower limits of the parameters. Fitness evaluation uses non-dominated sorting and crowding distance calculation. First, individuals are sorted according to their performance on the two objective functions using the Pareto front, with individuals belonging to the first non-dominated front receiving the highest rank, and so on. Within the same non-dominated layer, the crowding distance of individuals is calculated to maintain solution diversity. The selection operation uses a tournament selection method, randomly selecting five individuals from the population each time, choosing the one with the highest non-dominated rank and the highest crowding distance. The individual with the largest distance enters the next generation; the crossover operation uses simulated binary crossover, with a crossover probability set to 0.9 and a distribution exponent set to 20; the mutation operation uses multinomial mutation, with a mutation probability set to one-ninth (i.e., each individual mutates one dimension on average), and a distribution exponent also set to 20; the crossover and mutation probabilities of the genetic algorithm are adaptively adjusted according to the diversity index during the population iteration process. When the average crowding distance of individuals in the population is less than the threshold, the mutation probability is increased to increase diversity; the algorithm terminates after 200 generations, outputting all non-dominated solutions in the final population, forming a Pareto optimal solution set.

[0053] In historical data, for different slurry inlet mass percentage ranges, such as less than 0.5%, 0.5% to 1%, 1% to 1.5%, and greater than 1.5%, the above-mentioned genetic algorithm optimization is run separately, and a representative Pareto optimal solution set is saved for each range. During online operation, the system obtains the current slurry inlet mass percentage value in real time and determines its range. From the Pareto optimal solution set of the corresponding range, the predicted values ​​of each solution for the two objective functions under the current operating conditions are calculated. According to the production scheduling priority, if maximizing the recovery rate is the primary objective, the solution with the largest predicted value of the first objective function is selected; if saving reagents is the primary objective, the solution with the smallest predicted value of the second objective function is selected. The nine-dimensional control parameter vector corresponding to the selected solution is used as the parameter benchmark of the current incremental proportional-integral-derivative controller to calculate real-time cooperative control commands, thereby realizing the adaptive adjustment of control parameters to the feed properties.

[0054] In this embodiment of the application, the method further includes using a sensor window monitoring unit installed at the front end of the multi-source sensing array to collect window transmittance data in real time; in response to the window transmittance data being lower than a preset cleanliness threshold, a pulsed high-pressure air curtain driving signal is generated to remove mineral slurry splashes and acidic droplets attached to the sensor window.

[0055] Understandably, a small transmittance detection device consisting of a light-emitting diode (LED) and a photodiode is installed on the outside of the protective glass in front of the industrial camera lens. The LED emits a red parallel beam with a wavelength of 650 nanometers, which is incident perpendicularly on the outer surface of the protective glass. The photodiode receives the intensity of the beam after it passes through the glass. The transmittance data is calculated by dividing the received light intensity by the initial transmitted light intensity under clean glass conditions and then multiplying by 100%. This transmittance data is collected and transmitted to the industrial control computer at a frequency of ten times per second.

[0056] The preset cleanliness threshold can be set to 80%. When the real-time light transmittance is below 80%, the sensor window is considered contaminated. The industrial control computer immediately triggers a pulse signal through its digital output port, which drives the high-speed solenoid valve to open. The high-speed solenoid valve is connected to the factory's compressed air network, with an inlet pressure of 0.6 MPa. After the solenoid valve opens, high-pressure air is ejected through an annular nozzle surrounding the camera lens, forming a pulsed high-pressure air curtain with a duration of 0.1 seconds. The airflow speed of the air curtain exceeds 50 meters per second, which can effectively blow away mineral slurry droplets, dust particles, and acidic condensate droplets adhering to the protective glass. After the blowing is completed, the system checks the light transmittance again after a two-second delay. If it is still below the threshold, the blowing process is repeated, but the number of consecutive blowings does not exceed three to prevent excessive consumption of compressed air. This self-cleaning mechanism ensures that the vision sensor can maintain a clear field of view for a long time in harsh industrial environments, guaranteeing the stability of image data quality.

[0057] In this embodiment, the process of identifying the surface foam flow velocity includes, but is not limited to: extracting scale-invariant feature transform operators from two adjacent enhanced foam images; performing feature matching on the scale-invariant feature transform operators using a random sampling consensus algorithm to eliminate mismatched point pairs; calculating the Euclidean distance between successfully matched feature point pairs, and calculating the instantaneous motion displacement vector of the foam layer in conjunction with the camera intrinsic and extrinsic parameter array.

[0058] Understandably, the scale-invariant feature transform operator is a local feature descriptor that remains invariant to image scaling, rotation, and brightness changes. For the current frame and the previous frame's enhanced bubble image, the following steps are performed: First, a Gaussian difference scale space is constructed, consisting of five groups, each with four layers of images. A Gaussian difference image is obtained by subtracting adjacent scale Gaussian blurred images. Extrema are detected in the Gaussian difference scale space, i.e., each pixel is compared with all 26 neighboring points in the scale space and image space. If it is the maximum or minimum value, it is used as a candidate location for keypoints. Candidate keypoints are precisely located by fitting a three-dimensional quadratic function to correct their position and scale, and edge response points with low contrast or instability are removed. An orientation is assigned to each keypoint, and the main orientation is selected based on the gradient orientation histogram of the keypoint's neighboring pixels. Finally, the scale space neighborhood of each keypoint is divided into 4x4 sub-regions, and the gradient orientation histograms of eight directions are calculated for each sub-region, forming a 4x4x8 equal 128-dimensional feature descriptor vector, i.e., the scale-invariant feature transform operator descriptor.

[0059] For the two sets of scale-invariant feature transform operator descriptors extracted from the two frames of images, a nearest neighbor matching strategy is adopted. That is, for each descriptor in the first set, the nearest and second nearest descriptors in the second set are searched for by Euclidean distance. If the ratio of the nearest distance to the second nearest distance is less than a preset threshold, such as 0.7, the matching pair is accepted; otherwise, it is considered a fuzzy match and rejected. The point pairs obtained from the initial matching usually include a large number of mismatches caused by noise or similar textures. A random sampling consensus algorithm is used to remove these mismatches. The steps of this algorithm are: randomly select four pairs from the matching point pairs as the interior point set and calculate the basic matrix model; use this model to test all matching point pairs, mark the point pairs that conform to the model (i.e., the projection error is less than 1 pixel) as interior points, and mark the ones that do not conform as exterior points; repeat the above process (e.g., iterate 1000 times), and finally select the model with the most interior points, and retain all interior points as correct feature matching pairs, while exterior points are removed.

[0060] For the correctly matched point pairs retained after the random sampling consensus algorithm, let the coordinates of the point in the first frame image be the first coordinate, and the coordinates of the corresponding point in the second frame image be the second coordinate. Calculate the pixel displacement component of each point pair on the image plane. Since the foam movement is basically in a two-dimensional plane, and the camera has been calibrated to obtain its intrinsic and extrinsic parameters, including the camera focal length (in pixels), principal point coordinates, and camera mounting height (i.e., the vertical distance from the lens optical center to the foam surface, in millimeters), use the pinhole camera model to transform the image pixel coordinates to the world coordinates of the foam surface. Therefore, the world coordinate displacement component of the foam surface point is the pixel displacement component multiplied by the mounting height and then divided by the focal length. Average the displacement vectors calculated for all matched point pairs to obtain the average instantaneous motion displacement vector of the foam layer. Divide this by the time interval between two frames (e.g., one-thirtieth of a second) to obtain the instantaneous motion velocity vector of the foam layer. Its magnitude is the surface foam flow velocity, and its direction reflects the mainstream direction of the foam in the channel.

[0061] It is understandable that the process of quantifying the bubble merging rate includes, but is not limited to: segmenting the enhanced bubble image using the watershed algorithm to identify the closed contour of a single bubble; counting the number of closed contours of a single bubble disappearing per unit time; wherein the number of disappearing bubbles is positively correlated with the average brightness change rate of the corresponding region.

[0062] Specifically, the gradient magnitude is first calculated for the enhanced foam grayscale image using the Sobel operator. A distance transformation is then performed on the gradient image, calculating the distance from each foreground pixel (foam region) to the nearest background pixel. The distance transformation result is used as the terrain surface, with its local maxima considered as seed points for the bubbles. Starting from these seed points, a flooding process is simulated. When the water surfaces rising from different seed points meet, a watershed boundary is formed, which serves as the dividing line between the bubbles. Since the original image may contain noise leading to oversegmentation, a marker-controlled watershed is required before the watershed transformation. Specifically, a morphological opening operation (erosion followed by dilation, with a 3-pixel multiplied disk as the structuring element) is performed on the image to separate slightly adhered bubbles. Then, a distance transformation is performed, and maxima are found. These maxima are used as deterministic markers input into the watershed algorithm. After segmentation, each connected region surrounded by the watershed boundary is identified as a single bubble, its outline represented by a closed polygon composed of a series of pixel coordinates.

[0063] At a certain moment, a set of bubble contours is obtained from the current frame image using the watershed algorithm; at the next moment, a similar set of bubble contours is obtained. Bubble disappearance is defined as follows: a bubble contour that existed in the previous moment no longer encloses the image region in the next moment, or the region largely overlaps with multiple new contours in the next frame (meaning the original bubble has burst or merged into other bubbles). To quantify this, the average brightness change rate of each contour from the previous moment in the corresponding region of the next frame is calculated. Specifically, in the previous frame, the average grayscale value of the pixels within the contour is calculated, and in the next frame, the average grayscale value of the pixels within the contour is calculated. Calculate the average gray value of the region slightly larger than the original outline at the same spatial location (allowing for slight displacement). If the decrease in the average gray value of the next frame compared to the average gray value of the previous frame exceeds a preset threshold (e.g., 20%), and the region is crossed by a new watershed boundary in the next frame, then the bubble is determined to have disappeared within the time interval. Count the number of bubble outlines that meet the above disappearance conditions between two consecutive frames. The bubble merging rate is defined as the number of disappeared bubbles divided by the total number of bubbles and then divided by the time interval, with the unit being the proportion of the number of bubbles disappearing per second to the total number of bubbles. Its typical value range is 0.1% to 5% per second.

[0064] The mechanism by which the number of bubbles disappearing is positively correlated with the average brightness change rate of the corresponding area is that when two or more bubbles merge, the liquid film at the merging point ruptures, the foam structure collapses, and the specular reflection characteristics of the original bubbles change, resulting in a significant decrease in the average brightness of the area in the image. At the same time, the merging process generates larger bubble outlines, and the original small outlines disappear accordingly. By establishing the correspondence between the brightness change rate and the outline disappearance event, the dynamic process of bubble merging can be quantified more reliably from the image sequence. This dynamic parameter is an important indicator reflecting the stability and load of mineralized foam in the flotation cell.

[0065] It is understandable that the construction process of the long short-term memory neural network model includes, but is not limited to: dividing the continuous working condition feature tensor into multiple groups of time series samples using the sliding window technique; introducing forget gate logic into the neuron nodes to discard long-term redundant information in the historical sequence that is less than the preset value in relation to the current tailings grade fluctuation; and using the error backpropagation algorithm combined with an adaptive learning rate optimizer to iteratively update the model parameters.

[0066] Specifically, the operating condition feature tensors are generated continuously, with each tensor corresponding to a time step and including feature information from the past 20 time steps. To construct the samples required for supervised learning, a fixed historical window length of 100 time steps and a fixed prediction step size of 5 time steps are defined. A sliding window method is used to extract samples from the continuous tensor sequence. Specifically, starting from the first time step, 100 consecutive operating condition feature tensors from the initial time step are taken as the input feature sequence, with a shape of 100 time steps multiplied by 5-dimensional features. The actual tailings grade fluctuation value corresponding to the 5th time step after this input feature sequence is taken as the label. Then, the window is slid forward by one time step, and the extraction is repeated until all available data has been traversed. Finally, tens of thousands of sample pairs are generated, of which 80% are used for model training, 10% for validation, and 10% for testing.

[0067] In a Long Short-Term Memory (LSTM) neural network, each cell unit includes a forget gate. The forget gate output is calculated as follows: the S-shaped growth curve function applied to the weight matrix, the previous hidden state, and the current input, plus a bias term. The forget gate output value is between 0 and 1, controlling how much information from the previous cell state is retained in the current cell state. To automatically discard redundant information, a correlation-guided constraint is applied to the forget gate activation value during training. Specifically, in the training samples, the mutual information or Pearson correlation coefficient between the features at each time step in the historical sequence and the final prediction target (tailings grade fluctuation) is calculated to obtain a correlation score. During training, this correlation score is then applied to the forget gate's activation value. The relevance score serves as a reference signal, encouraging the model to learn a positive correlation between the forget gate output and the relevance score (limited to between 0 and 1). That is, for historical time steps with low relevance (relevance score less than a preset threshold, such as 0.3), the model should tend to output a lower forget gate value, thereby forgetting more of the corresponding old cell state information. This can be achieved by adding a regularization term to the loss function, for example, by adding a regularization coefficient multiplied by the sum of the squares of the difference between the forget gate output value and (the pruned relevance score). With the regularization coefficient set to 0.01, the model can adaptively construct the memory most relevant to the current task, improving the efficiency and accuracy of temporal modeling.

[0068] The loss function is defined as the mean squared error between the predicted tailings grade fluctuation value and the actual value. The backpropagation algorithm is used to calculate the gradient of the loss function with respect to all model weight parameters (including the weight matrices and bias terms of each gate, and the weights of the fully connected layers) over time. An adaptive learning rate optimizer is used, with the following parameters: an initial learning rate of 0.001, an exponential decay rate of 0.9 for the first moment estimate, an exponential decay rate of 0.999 for the second moment estimate, and a small constant of 10 to the power of -8 for numerical stability. Training is performed in batches with a batch size of 64. After each training round, the model performance is evaluated on the validation set. If the validation loss does not decrease for 10 consecutive rounds, the learning rate is multiplied by 0.5 for decay. The entire training process continues until the preset maximum number of training rounds (e.g., 200 rounds) is reached or the validation loss decreases by less than one-thousandth within 20 consecutive rounds, at which point training is stopped early. After training, the model parameters with the best performance on the validation set are saved for online prediction.

[0069] It is understandable that the process of outputting coordinated control commands includes, but is not limited to: determining whether the current mechanical displacement of each actuator has reached the physical limit position; if it has, then performing truncation processing on the corresponding adjustment step size and reallocating the compensation weights of the remaining control variables to prevent the actuator from overloading or being damaged.

[0070] Specifically, each actuator is equipped with a position feedback sensor; for the variable frequency motor of the foaming agent dosing pump, the actual operating frequency is fed back through an encoder, and its physical limit position corresponds to the upper and lower limits of the frequency. The lower limit is 10 Hz, below which the pump may stop, and the upper limit is 50 Hz, which is the rated maximum frequency of the motor; for the stepper motor for lifting the slurry outlet weir, its actual height position is fed back through a grating ruler, and its physical limit position corresponds to the upper and lower limits of the weir travel. The lower limit is 0 mm, i.e., fully closed, and the upper limit is 300 mm, i.e., fully open; for the electric positioner of the air-filling regulating valve, its actual opening degree is fed back through a potentiometer, and its physical limit position corresponds to the mechanical limit of the valve. The lower limit is 0 percent, i.e., fully closed, and the upper limit is 100 percent, i.e., fully open; the industrial control computer reads these feedback values ​​in real time and compares them with the preset limit values.

[0071] Assuming the calculated coordinated control command has three adjustment steps, a feedforward limit check is first performed: predicting the new position after executing the command. If the new position is greater than the upper limit, the adjustment step is truncated to the difference between the upper limit and the current position, and the excess control quantity is recorded. Similarly, the same check and truncation process is performed on other adjustment steps. If one or more control variables are truncated, i.e., the excess is not all zero, the control quantities need to be redistributed to compensate for the control actions that could not be executed due to the limit. The redistribution principle is based on the sensitivity coefficient of each control variable's influence on the tailings grade. These sensitivity coefficients are obtained through historical data analysis or process experiments. For example, the sensitivity coefficient of the frother frequency is 0.05% per Hz, the sensitivity coefficient of the weir height is 0.1% per millimeter, and the sensitivity coefficient of the valve opening is 0.08%. The percentage of each opening degree is considered. Assuming only the frequency adjustment step size is truncated, and its excess is a certain value, then the equivalent control effect of multiplying the excess by the corresponding sensitivity coefficient needs to be allocated to the control variables that have not reached the limit. The allocation weight is determined based on the available adjustment space and sensitivity of the remaining control variables. Specifically, the calculation is as follows: First, calculate the remaining adjustment space of the remaining control variables, which is equal to (upper limit minus current position) multiplied by the corresponding sensitivity coefficient; then allocate proportionally and calculate the increment that each remaining control variable should compensate; finally, the output cooperative control command is corrected to the truncated step size plus the compensation increment, and ensures that the corrected command will not cause new limit exceedances. This ensures that when individual actuators reach their limits, the system can still approach the control target as close as possible by adjusting other controllable variables, thereby maintaining the stability of the flotation process and effectively protecting the actuators from mechanical damage.

[0072] This application also discloses an intelligent flotation cell control system with integrated online detection function, including a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed as described above for the intelligent flotation cell control method with integrated online detection function.

[0073] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0074] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0075] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0076] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0077] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0078] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for controlling an intelligent flotation cell with integrated online detection function, characterized in that, include: By deploying a multi-source sensing array above the flotation cell and electrically connected to the industrial control system, multi-dimensional raw production data, including surface foam image stream, real-time slurry level height, instantaneous flow rate of the air inlet branch pipe, and slurry inlet mass percentage, are collected simultaneously. The surface foam image stream is preprocessed using median filtering and an adaptive contrast enhancement algorithm to generate an enhanced foam image that eliminates high-frequency noise interference. The enhanced foam image is input into a pre-trained lightweight convolutional neural network based on an attention mechanism, and spatial feature convolution and channel attention reconstruction are performed to extract and quantify dynamic visual representation vectors including average bubble diameter, foam layer load thickness, bubble merging rate and surface foam flow velocity. By calculating the gradient co-occurrence matrix of the enhanced foam image in the horizontal and vertical directions, the frequency domain texture features reflecting the roughness and directionality of the foam are obtained. The dynamic visual representation vector and the frequency domain texture features are then fused in a high dimension using principal component analysis to construct a working condition feature tensor with spatiotemporal correlation characteristics. The working condition feature tensor is input into the long short-term memory neural network model in time series, and the tailings grade fluctuation value is predicted in the future preset time step by combining the historical working condition evolution trajectory. In response to the dispersion of the tailings grade fluctuation value and the preset process index, the coordinated control commands, including the adjustment step size of the frother pumping frequency, the lifting and lowering displacement step size of the slurry outlet weir plate, and the adjustment step size of the air regulating valve opening, are calculated based on the incremental proportional-integral-derivative control strategy. The output of the coordinated control command drives the corresponding actuator to achieve closed-loop adaptive control of the flotation process in the flotation cell.

2. The intelligent flotation cell control method integrating online detection function according to claim 1, characterized in that, The preprocessing of the surface foam image stream also includes, but is not limited to: performing edge-preserving denoising processing on the surface foam image stream using a bilateral filtering algorithm, wherein the standard deviation of the spatial kernel function and the standard deviation of the value domain kernel function of the bilateral filtering algorithm are dynamically configured based on the local variance of the image; extracting the foam edge skeleton by calculating the second-order Laplacian operator of the denoised image; and superimposing the foam edge skeleton onto the original image according to a preset weight to generate the enhanced foam image with a high-contrast contour.

3. The intelligent flotation cell control method integrating online detection function according to claim 1, characterized in that, The process of extracting the dynamic visual representation vector includes, but is not limited to: introducing a squeezing and activation module into the feature map output layer of the lightweight convolutional neural network; wherein, the squeezing and activation module compresses spatial dimension information through global average pooling and generates importance weights for channel dimensions using a two-layer fully connected network; and reconstructs the feature map based on the importance weights to enhance the feature response of the mineralized foam region and suppress background interference.

4. The intelligent flotation cell control method integrating online detection function according to claim 1, characterized in that, The process of calculating the cooperative control command includes, but is not limited to: constructing a nonlinear multi-objective functional with the optimization objectives of maximizing production recovery rate and minimizing reagent consumption; using an adaptive genetic algorithm to perform global optimization on the nonlinear multi-objective functional to obtain a set of Pareto optimal solutions; and selecting the solution with the highest matching degree with the current slurry inlet mass percentage from the Pareto optimal solution set as the parameter benchmark for calculating the cooperative control command.

5. The intelligent flotation cell control method integrating online detection function according to claim 1, characterized in that, Also includes: The window transmittance data is collected in real time using the sensor window monitoring unit installed at the front end of the multi-source sensing array; In response to the window transmittance data being lower than a preset cleanliness threshold, a pulsed high-pressure air curtain drive signal is generated to remove mineral slurry splashes and acidic droplets adhering to the sensor window.

6. The intelligent flotation cell control method integrating online detection function according to claim 1, characterized in that, The process of identifying the surface foam flow velocity includes, but is not limited to: extracting scale-invariant feature transform operators from two adjacent frames of the enhanced foam images; performing feature matching on the scale-invariant feature transform operators using a random sampling consensus algorithm to eliminate mismatched point pairs; calculating the Euclidean distance between successfully matched feature point pairs, and calculating the instantaneous motion displacement vector of the foam layer in conjunction with the camera intrinsic and extrinsic parameter array.

7. The intelligent flotation cell control method integrating online detection function according to claim 1, characterized in that, The process of quantifying the bubble merging rate includes, but is not limited to: segmenting the enhanced bubble image using a watershed algorithm to identify the closed contours of individual bubbles; The number of times the closed outline of a single bubble disappears per unit time is statistically analyzed; the number of disappearing bubbles is positively correlated with the average brightness change rate of the corresponding area.

8. The intelligent flotation cell control method integrating online detection function according to claim 1, characterized in that, The construction process of the long short-term memory neural network model includes, but is not limited to: dividing the continuous working condition feature tensor into multiple groups of time series samples using the sliding window technique; introducing forget gate logic into the neuron nodes to discard long-term redundant information in the historical sequence that is less than a preset value in relation to the current tailings grade fluctuation; and using the error backpropagation algorithm combined with an adaptive learning rate optimizer to iteratively update the model parameters.

9. The intelligent flotation cell control method integrating online detection function according to claim 1, characterized in that, The process of outputting the coordinated control command includes, but is not limited to: determining whether the current mechanical displacement of each actuator has reached the physical limit position; if it has, then performing truncation processing on the corresponding adjustment step size and reallocating the compensation weights of the remaining control variables to prevent the actuator from being overloaded or damaged.

10. A smart flotation cell control system integrating online detection function, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and execute the intelligent flotation cell control method with integrated online detection function as described in any one of claims 1 to 9.

Images