Flotation equipment automation and intelligent froth feature extraction in flotation process: a review

2.1.1 Automatic dosing device

Flotation, as an important operational link in the production process of mineral processing plants, is closely related to the rationality of reagent addition. Reasonable automatic dosing is an important means to ensure accurate dosing of drugs.

The mining automation technology in Western countries is relatively developed. In the 1950s, countries such as Sweden, the former Soviet Union, and Finland began to develop a series of automatic dosing control devices, mainly including two types: micro metering pumps and differential pressure control valves (Ji 2016). The metering pump dosing machine is equipped with a metering pump at each dosing point, characterized by precise dosing and good corrosion resistance, but the disadvantage is that the operation is relatively cumbersome. The pressure difference control valve dosing machine uses air pressure difference to transport medicine liquid, but it also has complex operation and few industrial applications (Wu 2019). In the 1960s, China successively developed cup type automatic dosing devices and siphon type dosing devices. However, due to difficulties in achieving precise dosing, pipeline blockage, and siphon disconnection, these two types of automatic dosing devices have gradually been phased out. After the 1970s, many beneficiation plants began to use modern equipment controlled by digital computers, such as a diaphragm pump reagent adding machine at a beneficiation plant in Queensland, Australia. The flotation operation was carried out using a segmented dosing method, and the system was added reagents by real-time monitoring of product ash content changes. Currently, the widely used automatic dosing devices in the mineral processing field are electromagnetic valve type, volumetric pump type, and weighing type technologies, which have advantages such as closed-loop control, precise dosing, and convenient operation.

2.1.1.1 Electromagnetic valve type dosing technology

The electromagnetic valve type dosing device is composed of a storage device, a dosing device, a liquid level alarm, an electromagnetic valve, and a float valve. Its working principle is based on the basic principle of orifice flow and the intermittent dosing method. The structural schematic diagram is shown in Figure 2.

Figure 2:

Schematic diagram of electromagnetic valve type dosing device structure.

Currently, there is massive research and application on electromagnetic valve type dosing devices in China. For instance, the KMUST-FDCS series dosing control system developed by Kunming University of Technology has been applied to the Damajianshan beneficiation plant of Yunnan Lvchun Mining Co., Ltd., improving the efficiency of reagent utilization (Wang 2009). The BRFS type control system developed by Beijing Research Institute of Mining and Metallurgy is applied to the sorting system of the Duobaoshan Copper Molybdenum Mine in Heilongjiang Province. Eight reagents, including kerosene, yellow medicine, black medicine, Z-200, sodium sulfide, water glass, terpineol oil, and mercaptoacetic acid, are added to the process. The experimental results show that the error between the actual dosage and the designed dosage of all reagents is within ±2 %, indicating a good automatic dosing effect (Yan et al. 2016). It should be pointed out that when the dosage per unit cycle is too large, it is difficult for the float valve to maintain a constant liquid level in a timely and accurate manner, which can cause fluctuations in the reagent flow through the solenoid valve and affect the accuracy of dosing. If the dosage is too small and the selection of the solenoid valve is also small, it is easy to cause blockage of the solenoid valve.

2.1.1.2 Positive displacement pump type dosing technology

Positive displacement pump type dosing device is a type of pump that utilizes changes in the volume of the pump cylinder to transport liquid. There are two types of pumps: rotor pump and reciprocating pump. When selecting the automatic dosing device, the rotor pump uses a single screw pump for dosing, and the reciprocating pump uses a metering pump for dosing (Wei et al. 2022). The working principle of the single screw pump is that when the motor drives the pump shaft to rotate, the liquid medicine in the sealing chamber will advance by one screw pitch every time the screw rotates. With the continuous rotation of the screw, the liquid medicine will extrude the pump body from the sealing chamber. In the sulfur lead zinc flotation separation process of Maoping Lead Zinc Ore dressing plant, a single screw pump dosing device was used, which met the requirements of the dressing plant for precise dosing, solved the problem of easy blockage of the dosing system, improved the dosing operating environment, and reduced the operating cost (Ao et al. 2018). The metering pump type dosing device is similar to the electromagnetic valve type dosing device and is composed of a storage tank, dosing device, liquid level alarm, metering pump, float valve, and reagent buffer hopper. As shown in Figure 3, a metering pump is a special volumetric pump for transporting liquids (with corrosive liquids), and a more common type is a diaphragm metering pump. The diaphragm causes the movement of the ball valve, forming vacuum adsorption and squeezing phenomena, thereby achieving the purpose of transporting liquid medicine. Dongqu Coal Preparation Plant adopts a diaphragm metering pump dosing device. Compared with manually adjusting the dosage, the daily usage of kerosene was reduced by 54 kg, and the daily usage of secondary octanol was reduced by 178 kg. This saves over 840,000 yuan in reagent costs per year (Liang 2021). It should be noted that the metering pump type dosing technology was suitable for mineral processing plants with large dosage and few dosing points. However, for mineral processing plants with small dosage and many dosing points, it cannot meet production accuracy, and this type of device has the characteristics of high price and production cost.

Figure 3:

Schematic diagram of the structure of the metering pump type dosing device.

2.1.1.3 Weighing type dosing technology

The weighing type dosing device mainly consists of a dosing box, an upper weighing body, a lower weighing body, a weighing sensor, a dosing valve, and a reagent buffer bucket. Its working principle is to record the overall weight of the dosing box and the agent based on the weighing sensor. After opening the dosing valve, when the total weight reduction recorded by the weighing sensor is equal to the production design demand, the dosing valve is closed to complete the dosing. The structural schematic diagram is shown in Figure 4. According to the characteristics of the equipment, it can be seen that the sensitivity and accuracy of the weighing sensor directly affect the accuracy of dosing, and it is suitable for mineral processing plants with large dosage and few dosing points. The weighing type dosing technology was used in the iron ore flotation process of Yunnan Gejiu Chongjing Concentrator. The results showed that the dosing control accuracy reached ±0.7 % and was not affected by changes in reagent viscosity, pressure, or equipment wear. The operation effect was good and the expected goal was achieved.

Figure 4:

Structural schematic diagram of weighing type dosing device.

Production practice has shown that the flotation automatic dosing device and technology have overcome the shortcomings of relying on manual dosing, reduced reagent consumption, reduced labor intensity of workers, and improved the economic benefits of the enterprise. Then, it should be pointed out that there are still problems with the automatic dosing process of flotation, such as poor dosing environment, equipment malfunction, and operational errors. In the future, the intelligent level of the dosing process should be further improved.

2.1.2 Automatic liquid level detection device

As an important component of flotation control, flotation liquid level detection has a significant impact on product quality and economic benefits. The automatic liquid level detection device is a device for measuring liquid level and is also a major part of achieving liquid level control. In recent years, automatic liquid level detection methods have gradually undergone significant development, ranging from the simplest manual measurement methods such as glass tube method, dual color water level method, and manual gauge method to float ball method, float tube method, and sink barrel method developed based on the principle of buoyancy. With the development of modern electromagnetic technology and wave optics technology, capacitance method, resistance method, ultrasonic method, and resistance method have been successively introduced (Lin 2017). Currently, the common liquid level detection devices at home and abroad can be divided into two types: contact type and noncontact type. The noncontact detection devices are mainly ultrasonic level gauge or radar level gauge. Ultrasonic level gauge has the advantages of high measurement accuracy, convenient installation, and basically no maintenance. However, the ultrasonic speed is affected by the transmission medium, external temperature, pressure, density, and the characteristics of the froth layer on the pulp surface, resulting in too large ultrasonic reflection angle and often no measurement or measurement errors (Guo and Wang 2020; Hou 2007). The radar level gauge adopts the working mode of transmitting reflecting receiving, which is characterized by high measurement accuracy, safety and reliability, and long service life. However, the radar level gauge is not penetrating and has poor detection effect for the level in the environment containing froth (Xie and Guo 2019). The contact type liquid level detection device is mainly composed of input type or pressure type liquid level gauges, and its detection principle is to convert pressure changes into changes in liquid level. The advantage is that it has penetrability and can detect real liquid levels. The disadvantage is that it is easy to block and is greatly affected by the concentration of the slurry. Nowadays, due to the influence of slurry adhesion, corrosion, slurry mixing, and froth floating, the automatic liquid level detection device cannot accurately and effectively detect the liquid level. A liquid level control system includes a liquid level controller, a liquid level sensor, and a variable speed drive. The main difficulty in controlling the flotation liquid level lies in the highly interconnected system where each flotation cell is coupled and affects each other, requiring the use of interconnected controllers to achieve stable control of the liquid level. For instance, the liquid level stability of a multistage flotation cell is shown in Figure 5 (Cai 2021). If the liquid level of flotation cell 1 is changed while keeping the liquid levels of flotation cells 2 and 3 unchanged, it is necessary to establish a liquid level relationship model for three flotation cells and adopt a feedforward liquid level control strategy to enhance the anti-interference ability of the entire circuit.

Figure 5:

Liquid level stability control block diagram of multistage flotation cell. Reprinted with permission from (Cai 2021).

According to the above research, current detection devices cannot accurately measure the liquid level value, which has always been a challenge in industrial control. Therefore, developing an intelligent detection and control system for flotation liquid level has important practical significance.

2.1.3 Feed concentration device

Flotation feed concentration is also an important control parameter in flotation operation. Reasonable feed concentration can make froth carry mineral particles to the maximum extent, ensure flotation recovery, and reduce resource waste. At present, the main methods for detecting the concentration of flotation feed are manual interval sampling and instrument testing. The manual interval sampling detection method has high sampling accuracy, simple detection process, and low cost, but it is inefficient, unable to detect data in real-time, and cannot adapt to automated and intelligent production (Lv 2021). With the development of current detection technology, real-time detection technology and equipment for slurry concentration have also made significant progress, such as γ radiation detector, radioactive nuclide concentration detector, float slurry concentration detector, electromagnetic induction concentration detector, and ultrasonic concentration detector device. The γ-ray detector determines the concentration of mineral pulp by detecting the attenuation of γ radiation, and the absorption intensity of γ radiation by mineral pulp is related to the concentration of mineral pulp. The principle of a radioactive nuclide concentration detector is to capture the radiation of radioactive isotopes to a certain extent through the pulp. When the radiation generated by radioactive elements passes through the pulp, the pulp will absorb a portion of the radiation, and the absorbed pulp radiation is related to the pulp concentration. The advantage of this detection method is the noncontact detection method, which can detect data in real-time and the device and flotation process do not interfere with each other. The disadvantage is that this method is radioactive, and isotope radiation can generate a large amount of radiation, which has a health risk impact on staff. The float pulp concentration detector is based on Archimedes’ principle. When the float floats in the pulp and reaches balance, the mass of the float to discharge the liquid is the same as the mass of the float itself. The pulp concentration is obtained according to the two transformations of the inclination angle-current or voltage signal-pulp concentration changed by the float balance. The detection device has a simple structure, convenient operation, and can provide real-time feedback on slurry concentration. The disadvantage is that the slurry has a corrosive effect on the direct contact device. In addition, the viscosity of the slurry can also have an adverse impact on the accuracy of the device, requiring manual periodic cleaning of the device. The principle of an electromagnetic induction concentration detector is to convert magnetic signals into electrical signals based on the principle of electromagnetic induction and calculate the size of the slurry concentration through the integrated linkage of slurry concentration, magnetic mineral concentration, and induced current. The advantages of this detection method are simple structure and convenient detection. The disadvantage is that the detection accuracy needs to be improved. The detection principle of an ultrasonic concentration detector is based on the attenuation of sound waves, attenuation of sound velocity, and changes in sound impedance when ultrasonic waves pass through the slurry. The schematic diagram of its principle is shown in Figure 6. Under the excitation of continuous sine wave pulses, the ultrasonic emission transducer emits ultrasonic plane longitudinal waves into the slurry in a thickness vibration manner. When the ultrasonic interacts with the slurry, it attenuates and reaches the ultrasonic receiving transducer, causing energy and amplitude to decay. This degree of attenuation is closely related to the slurry concentration, and thus the slurry concentration can be calculated (Han 2014).

Figure 6:

Schematic diagram of ultrasonic measurement. Reprinted with permission from (Han 2014).

This method is also noncontact detection for radioactivity, which can provide real-time feedback on changes in slurry concentration. It has high detection accuracy, low maintenance costs, and does not generate ionizing radiation. However, the disadvantage is that when the slurry concentration is high, the attenuation of ultrasound is not linearly related to the concentration. In addition, factors such as the absorption of froth to ultrasonic waves and the influence of slurry temperature change on ultrasonic attenuation will also interfere with the test results.

Overall, when measuring the concentration of mineral slurry, it is a three-phase mixture of solid, liquid, and gas, which is a strong acid or alkali environment and also has high viscosity. These factors will have an impact on the corrosion resistance, wear resistance, and measurement accuracy of the testing device, and the above methods cannot detect concentration data in real-time for a long time and with high accuracy. Therefore, developing noncontact high-precision, safe and efficient detection devices will be a key development direction in the future. The classification of flotation automation equipment and corresponding advantages and disadvantages in the application process is shown in Table 1.

2.1.4 Automatic adjustment device for feed flow rate

It is necessary to accurately and truly measure the flow value in a beneficiation plant such as tailings leakage line detection, heavy medium cyclone and hydrocyclone detection, and flotation process detection. Traditional methods such as electromagnetic flow meters, ultrasonic flow meters, vortex flow meters, wedge flow meters, etc. pose a huge challenge in accurately measuring flow due to scaling on pipe walls, changes in material properties, calibration deviations, and entrainment of bubbles. Currently, the widely used technology internationally is a new method for flow testing based on array sensors. The measurement principle of this method is based on an array algorithm of passive sonar sensors to detect, track, and measure any disturbance velocity of pipeline axial movement. These disturbances can be divided into fluid transmission, acoustic transmission, and vibration transmission on the inner wall of the pipeline. Each type of disturbance has a different speed, and the type of disturbance and measurement of disturbance rate can be accurately distinguished through the difference in disturbance speed. The measurement principle based on arrays has proven to have significant advantages in volume flow measurement and gas porosity measurement for various mineral processing applications, especially in situations such as bubble entrainment, pipeline fouling, high wear and corrosion of pipelines, and the presence of magnetic minerals. Currently, this technology is being applied in monitoring over 700 mineral processing processes in 22 countries (VO’Keefe et al. 2010).

2.2.1 Extraction of physical and dynamic characteristics of froth

In the process of mineral flotation, the physical and dynamic characteristics (color, texture, size, moving speed) of froth are closely related to the flotation production indicators, working conditions, and operating variables, which can be used as an important basis for judging the effect of mineral separation operations. Because froth images are formed by the accumulation of mineralized bubbles with different numbers, sizes, shapes, and colors, the boundaries between bubbles are not clear, and bubbles are piled and squeezed with each other. Bubbles are broken and merged seriously, so it is difficult to obtain effective results with conventional froth image processing methods. The physical and dynamic characteristics of froth can be accurately and quickly extracted through intelligent algorithms, machine vision technology, and image processing technology. For instance, some scholars and teams have harnessed a machine vision system to cull froth’s visual characteristics, encompassing attributes such as bubble dimensions, uniformity, hue, and texture. These attributes are subsequently subject to scrutiny via assorted intelligent algorithms, thereby enabling the modeling of the nuanced associations between froth characteristics and metallurgical parameters. The eventual consequence is the formulation of predictive models (Jahedsaravani et al. 2014; Massinaei et al. 2020; Mehrabi et al. 2014; Tan et al. 2016).

2.2.1.2 Texture feature extraction of froth image

Froth texture is the comprehensive performance of froth roughness, contrast, and viscosity. As another key feature characterizing the statistical distribution of froth images, it can be used to describe the change of froth state caused by the change of operating conditions and mineral properties in the flotation process (Gui et al. 2013).

Abundant research work has been done on texture feature extraction of froth images at home and abroad, and the feature extraction methods in the literature of running images are divided into four categories, namely, texture feature extraction methods based on statistical families, texture feature extraction methods based on model families, texture feature extraction methods based on structure families, and texture feature extraction methods based on signal processing families. Among them, texture feature extraction methods based on statistical families mainly focus on statistical data analysis of texture image features, including grayscale co-occurrence matrix method (GLCM), grayscale shape statistics method, grayscale difference statistics method, cross diagonal matrix method, grayscale gradient statistics method, local grayscale statistics method, autocorrelation function method, semi variogram method, and texture spectrum statistics method. Most statistical based methods were discovered by Julez (Julesz 1975; Julesz and Caelli 1979). The texture feature extraction method based on model families assumes that the texture is distributed according to a certain model and then analyzes this model. Common model methods include Markov random field (MRF) model method (Woods 1972), Gibbs random field model method (Sivakumar and Goutsias 1999), Wold model (Liu and Picard 1994), Fractai model (Tang et al. 2002), complex network model, mosaic model, etc. The key to feature extraction by model method is to select appropriate models and parameter values. However, because of the complex texture of froth, it is difficult to accurately describe it through a single model, and the calculation is heavy. The texture feature extraction method based on structural family is a texture feature extraction method based on texture primitive theory, which emphasizes the regularity of textures. However, due to the lack of regularity in most textures in nature, this method limits its research depth in texture feature extraction. The texture feature extraction method based on signal processing family is to convert texture images from spatial domain to other transform domains through certain methods, which can also be called filtering method. Including Radon transform, ring and wedge filtering, discrete cosine transform, local Fourier transform, local Walsh transform and Hadamard transform, Gabor wavelet transform, binary wavelet, multi band wavelet, pyramid wavelet decomposition, ridge wavelet decomposition, curved wave decomposition, Laws texture measurement, feature filter, orthogonal mirror filtering, optimized FIR filter, etc. The Fourier transform method (Stromberg and Farr 1986), Gabor transform method (Bovik et al. 1990), and wavelet transform method (Pun 2003) are commonly used in the above methods.

Nowadays, the research on texture feature extraction of flotation froth has made some progress, and the commonly used method is to characterize the texture feature of froth surface through the energy, entropy, and moment of inertia of the field gray correlation matrix, spatial gray correlation matrix, or gray difference matrix. Gui et al. (2012, 2013) used grayscale co-occurrence matrix (GLCM) to extract full texture features of images. Based on the second order combined conditional probability density function of the estimated image, the gray correlation between the pixels in the image that are related in distance and direction is calculated and counted. This correlation is mainly reflected in the comprehensive information of the image in direction, adjacent interval, change amplitude, speed, and so on. The physical meaning of froth’s visual texture feature parameters (energy [thickness], moment of inertia, parameter entropy) is analyzed, and the correlation between froth feature parameters and froth texture is pointed out. The flow of froth image texture feature extraction is shown in Figure 7. Bartolacci et al. summarized the commonly used froth texture feature extraction methods, studied three commonly used froth texture analysis methods through Multivariate Image Analysis (Liu et al. 1990), GLCM (Hu et al. 2006), and wavelet texture analysis (Tang et al. 2011), extracted froth texture features respectively, discussed the classification of flotation state based on froth texture features, and established a concentrate grade prediction model with PLS (Partial Least Squares) (Wold et al. 2001) for flotation control, but no good conclusive results were obtained (Bartolacci et al. 2006). To sum up, although froth texture feature extraction technology has made great progress, due to the complex working conditions of the flotation process, the micro heterogeneity and complexity of froth texture, as well as the robustness of conceptual uncertainty and light changes, the precise extraction of froth texture features still need further exploration.

Figure 7:

Flow of froth image texture feature extraction. Reprinted with permission from (Gui et al. 2012, 2013).

2.2.1.3 Bubble size feature extraction

The size of flotation froth is closely related to pulp pH, pulp concentration, and material fineness and can reflect the production indicators such as flotation product grade and recovery rate. The size feature extraction technology of froth images often depends on the image segmentation effect. However, due to the mutual extrusion, different shapes, and uneven distribution of froth, it is difficult for conventional segmentation technology to achieve accurate segmentation effect. Scholars have carried out massive research on the size characteristics of flotation froth, mainly focusing on watershed segmentation method (Bonifazi et al. 2001), reconstruction morphology method (Sadrkazemi and Cilliers 1997), homogeneous gradient method (Botha et al. 1999), valley bottom edge method (Wang et al. 2003), improved reconstruction transformation and watershed method (Yang et al. 2009a,b; Zhou et al. 2009), morphology technology (processing steps are shown in Figure 8) (Wang and Tang 2002; Zhou et al. 2010), clustering preclassification and distance high-low precision reconstruction method (Yang et al. 2008), the image segmentation technology based on the seed region boundary growth method (Mou and Zhang 2009) solves the difficult segmentation problems of froth image, such as bubble adhesion, and boundary blur. Among the above methods, the valley edge method, watershed segmentation method, and morphological technology are the three commonly used segmentation methods at home and abroad. However, each of these three methods has its own limitations. Although the edge detection algorithm has a relatively simple template and is easy to operate, it also has problems such as inaccurate edge positioning, nonsingle pixel wide edges, and poor operator noise resistance (Ren 2008; Wang 2003). The commonly used methods for watershed segmentation are distance-based, gradient-based, and marker-based watershed image segmentation methods. But the most common problem is severe undersegmentation and oversegmentation (Li 2011; Zhang 2013). In addition, the morphological parameters in morphological algorithms should be selected based on the characteristics of the object, and in practical industrial applications, the flotation conditions are unpredictable, which can easily lead to the loss of robustness of fixed morphological parameter patterns and cause analysis errors.

Figure 8:

Processing steps of froth image segmentation method based on classification and morphology. Reprinted with permission from (Zhou et al. 2009).

In recent years, deep learning has become more and more well-known. Horn et al. (2017) found that convolutional neural networks (CNNs) performed better than traditional methods in feature extraction. Montes-Atenas et al. (2016) employed deep neural networks (DNNs) to predict bubble size and bubble rate using data obtained from validated computational fluid dynamics (CFD) computations. Pure water and slurry (in conditions similar to those employed in mineral froth flotation) case studies are evaluated. It is found that the DNN can predict the CFD results accurately when using four hidden layers, describing discontinuities in the bubbly flow regime. The relative errors computed between the CFD data and the prediction obtained by the DNN is as low as 8.8 % and 1.8 % for bubble size and bubble rate, respectively. Furthermore, Fu and Aldrich (2018) found that better results can be obtained with deeper CNN. Zhang et al. (2022) developed a generative adversarial network (GAN)-based setpoint computation model, and its effectiveness is verified in experiments on zinc flotation data. It should be noted that these methods require a large amount of offline data to train a reliable model, which includes various sources of data such as flotation froth images, process variables, and manipulated variables. The image acquisition equipment in the industrial flotation process is usually composed of hardware equipment and software systems for image acquisition, display, storage, online analysis, etc.

2.2.1.4 Extraction of froth dynamic characteristics (moving speed)

The change of air recovery rate and froth size determined by the dynamic characteristics (moving speed) of flotation froth directly affects the flotation production efficiency and flotation separation indicators. In view of the flow distortion such as froth collapse, fragmentation, overlap, and merger in the flotation process, the traditional image processing algorithm is difficult to identify and track froth and cannot achieve accurate extraction of froth dynamic characteristics. Focusing on the accurate extraction of froth dynamic characteristics, many scholars have carried out more research on froth dynamic characteristics.

Mou (2012) compared and analyzed the matching effect of froth images based on optical flow, macroblock tracking method, and phase correlation method. The results showed that macroblock tracking technology combining phase correlation and grayscale template matching was more suitable for estimating the motion characteristics of froth images. Ventura-A-Iedina and Cilliers (2000) found that under certain conditions, froth flow speed can be used as a quantitative feature of flotation performance. Neethling (2008) discussed the influence of froth movement speed characteristics on recovery. By analyzing froth structure parameters and froth surface area flow rate, froth texture spectrum characteristics are used to estimate froth size changes. Zhang and Liu et al. (2016) proposed to improve the matching conditions of SIFT algorithm according to the size and direction of froth speed, eliminate mismatches through random sampling consistent algorithm, and extract dynamic features according to the matching results. Fu and Aldrich (2018) used convolutional neural network (CNN) to develop flotation froth image sensor and identified froth status through a large number of image sample training. Feng (2017) achieved accurate mapping of the velocity field through block matching, pairwise matching, PROSAC screening matching, and other strategies for adjacent frame froth images and also achieved quantitative description of froth stability. The ORB-based improvement of the measurement process of flotation froth velocity field is shown in Figure 9. Holtham and Nguyen 2002 proposed pixel tracking technology to obtain froth velocity characteristics. Kaartinen et al. (2002) used froth cross-correlation peak value to describe froth speed and applied it to zinc flotation process. Brown et al. (2001) used the flow rate of froth to evaluate the stability of grade and realized the monitoring of the grade of gold ore rough concentrate.

Figure 9:

Measurement process of flotation froth velocity field improved based on ORB. Reprinted with permission from (Feng 2017).

In conclusion, the real-time online measurement of froth flow speed can provide guidance for flotation production, realize the optimized and stable operation of mineral flotation, and improve the flotation production efficiency.

2.2.2 Key indicator prediction

Real-time detection is the foundation of mineral flotation process monitoring. Accurately extracting flotation information helps to accurately judge the flotation process and provide reasonable suggestions. During the flotation process, the grade, recovery rate, ash content in the concentrate, and tailings are key indicators for process detection. Realizing real-time detection of key indicators in the flotation process is of great significance for improving flotation efficiency. Therefore, how to detect these key indicators of flotation production online has attracted the attention of scholars at home and abroad (Nakhaei et al. 2023). At present, key indicator prediction methods mainly include traditional data-driven models and image feature-based prediction methods.

The traditional key indicator prediction model takes process parameters such as pulp concentration, feed rate, feed particle size, feed grade, flotation reagents, ventilation rate, liquid level, etc. as input variables to establish a data-driven model. Gonzalez et al. (2003) compared the application of several commonly used prediction models, such as ARMAX (Autoregressive Moving Average Model), neural network, fuzzy combination model, and PLS, to predict the grade of copper flotation concentrate using raw copper ore grade, feed rate, slurry level, and concentration as input variables. They concluded that PLS can obtain the most accurate results, but the PLS model used is linear. Dayal (1997) used a recursive exponential partial least squares model to predict the concentrate grade and recovery rate of lead and copper using the input variables of air flow rate, pulp pH value, and feed rate. The experimental results showed that the effect was superior to the least squares and linear PLS methods. Nakhaei et al. (2010, 2023) used the thickness of froth, the amount of water added, and the feeding rate as the input variables and used the artificial neural network to establish the prediction model of copper concentrate grade, which provided guidance for the adjustment of production process parameters. Wang et al. (2022) proposed a flotation circuit automatic operation adjustment framework based on generative confrontation imitative learning (GAIL), which is a new method of learning operation adjustment strategy (FlotGAIL). In FlotGAIL, learning flotation operation from expert demonstration is expressed as the imitative learning problem in the Markov Decision Process. The proposed model uses the signal provided by the discriminator as a reward function to guide the strategy generator to complete the training process. The engineering application process has verified that the proposed operation adjustment method has good effects in controlling the concentrate grade and achieving recovery rate standards. Zhang et al. (2021) studied the method of predicting zinc tailings grade using an encoder decoder model. Sampling through froth video and X-ray fluorescence (XRF) analyzer, the first rough feature time series is automatically extracted and fed to the encoder to generate the context vector, and then the context vector and the previously measured level are sent to the decoder to predict the current tailings level. The validity of the model in froth flotation is verified by experiments. Compared with traditional recurrent neural network (RNN)-based models, this encoder decoder model has significant advantages. Wen et al. (2023) used the element content and ash content of coal samples as input parameters and used multiple polynomial regression, random forest regression (RFR), extreme gradient boosting (XGBoost), deep neural network regression (DNN), and other supervised regression learning algorithms to build XRF data sets. The data sets were divided into training sets and test sets by 8:2. During model training, random search was used to optimize hyperparameter. The experimental results indicate that learning models and XRF data are good choices for predicting ash content. Sun (2021) used BP neural network and intelligent optimization algorithm as the main technical means to build a flotation concentrate grade prediction model based on the improved atomic search algorithm ASO optimized BP neural network. Performance tests showed that the improved ASO algorithm has better computational accuracy and convergence performance. The industrial test results show that the ASOINU-BP algorithm has good predictive ability and reliability in the study of flotation concentrate grades.

The image feature-based prediction method is to extract image feature data that can reflect the flotation process for modeling and to achieve effective prediction of key indicators (Szmigiel et al. 2024; Wang et al. 2023). At present, the machine vision technology is widely used to extract froth features and analyze quality indicators. Yang et al. (2022) proposed a convolution attention parallel network (CAPNet) hybrid model to quickly and accurately determine the ash content of flotation coal by analyzing froth images. The results indicate that CAPNet outperforms other methods in terms of accuracy and stability, as shown in Figure 10. The application of CAPNet in actual production will significantly improve the automation and intelligence level of coal flotation and can also increase economic benefits. Moolman et al. 1994 first pointed out that the appearance characteristics of flotation froth contain a lot of information related to operating parameters and process technical indicators, which is an important basis for judging flotation effect. Liu et al. (2003) established a classification recognition model for flotation state based on the texture features extracted from the froth image of slime concentrate. Different froth corresponds to different concentrate grades and different reagent systems, thus realizing machine recognition of flotation conditions. Aldrich et al. (2010, 2022) searched for the relationship with concentrate grade by obtaining the characteristics of froth color, crushing rate, and size of zinc roughing, established a prediction model to estimate concentrate grade, and gave the expert control strategy for production operation. Hargrave and Hall (1997) established the relationship model between froth color parameters and flotation performance indicators (including concentrate grade, pulp flow, etc.) by combining conventional statistical methods and neural network models. Kaartinen et al. (2006) extracted the characteristics of froth flow rate, stability, bearing rate, etc. and found that it had strong correlation with concentrate grade. They also developed a multi camera flotation monitoring system for roughing cells, cleaning cells, and scavenging cells, which improved the accuracy of flotation froth classification. Moolman et al. (1996) applied froth image processing technology to molybdenum flotation process, extracted texture features such as energy, entropy, inertia, and correlation from the image, and established neural network prediction model to predict concentrate grade and recovery.

Figure 10:

Convolutional attention parallel network (CAPNet) hybrid model. Reprinted with permission from (Yang et al. 2022).

In summary, although traditional data-driven models and image feature-based prediction methods have made significant progress in predicting key flotation indicators, the flotation process is too complex, including dozens of flotation cells. The input variables of each flotation cell are interrelated, and these factors collectively determine the value of flotation key indicators. How to analyze the relationship between input variables and key indicators, establish effective prediction models, and achieve real-time and accurate automatic prediction of key indicators requires further systematic research.