Cloud Hosting

Mineral-intensive technologies such as renewable energy and electric vehicles will be in high demand as climate change is addressed and a sustainable energy future is transitioned. Despite this, the mining sector, especially in copper, is facing considerable difficulties due to the growing demand1,2. The depletion of high-grade ore and the rise of high-arsenic copper resources are more prominent issues in this field. Arsenic in low-grade copper ores not only makes mineral processing more difficult, but also causes environmental and health concerns due to the presence of arsenic in wastewater and exhaust gas3,4. The correlation between arsenic exposure and a range of health problems highlights the pressing requirement for ecologically viable methods in the mining sector. The incorporation of modern technologies such as hyperspectral imaging into sensor-based ore sorting systems has significant potential in this situation5. Ore sorting systems can successfully separate high-arsenic ores from valuable material by utilizing the precise and accurate analysis of mineral composition provided by hyperspectral imaging. This not only mitigates environmental dangers but also reduces processing costs. This strategy not only tackles the difficulties presented by ores containing high levels of arsenic, but also aids in the advancement of a mining industry that is environmentally friendly, in line with the objectives of the Paris Agreement.

Sensor-based ore sorting has become a crucial technology in mineral processing, providing numerous advantages that transform conventional mining methods. With sensor-based ore sorting systems, valuable minerals can be selectively extracted from ore streams according to their unique physical and chemical properties based on advanced sensor technologies. This process of selective extraction maximizes the efficient use of resources by effectively separating valuable ore from nonvaluable materials (or gangue minerals). In the field of mineral processing, Sensor-based ore sorting is a vital component as it enhances ore grades and minimizes the amount of waste material that is processed6. Evidence demonstrates that it effectively decreases the usage of energy, water, and reagents, while also minimizing the formation of fine waste, by disposing of trash before undergoing additional processing7,8. To successfully apply sensor-based sorting, it is crucial to select a sensing approach that can efficiently distinguish between ore and waste9. Sensor fusion, the integration of data from several sensing systems, has the potential to enhance the characterization of the detected material and improve sorting capability4. Microwave imaging (MWI) is a promising technique that can penetrate deeply into rock particles and serve as an additional approach for analyzing ores with significant differences in electromagnetic characteristics10. The efficacy of MWI in ore sorting has been validated by simulations and tests, affirming its capability to segregate valuable minerals or metals from unproductive particles. The utilization of sensor-based ore sorting presents substantial advantages in terms of reducing costs and enhancing efficiency in mineral processing.

Ore sorting techniques can be significantly enhanced by leveraging hyperspectral imaging technology, which offers unparalleled capabilities for mineral characterization and classification. Hyperspectral imaging allows ore sorting systems to analyze the distinct spectral fingerprints of minerals over a wide range of wavelengths, unlike traditional sorting methods that only consider physical attributes like size, shape, and density. This enables the identification and differentiation of minerals by analyzing their unique chemical compositions and optical features. Hyperspectral imaging is used in sensor-based ore sorting to analyze ore streams in real-time without causing damage5. This technique offers important details on the mineralogy and quality of the material being processed. By using hyperspectral imaging technology into sorting systems, mining companies can enhance their efficiency, precision, and selectivity in segregating valuable minerals from waste material. As a result, mineral processing enterprises have higher rates of recovery, lower costs of processing, and increased profitability.

The processing of hyperspectral data is more challenging than that of other types of data due to the sheer volume of information collected, which may be affected by issues with its dimensions. High-dimensional spectral bands in hyperspectral images are often highly similar, which makes them susceptible to the "curse of dimensionality," a phenomenon that affects many traditional algorithms11. Within the domain of hyperspectral ore sorting systems, the notion of wavelength selection arises as a crucial strategy for enhancing sorting efficiency and precision. Wavelength selection is the process of strategically identifying and using wavelengths of electromagnetic radiation (light) that provide the most useful information for differentiating between various minerals or compounds in an ore stream. Through the analysis of distinct spectral patterns displayed by minerals at different wavelengths, the process of wavelength selection allows ore sorting systems to concentrate on the specific spectral bands that are most effective in distinguishing the desired minerals. By employing this focused method, the precision, effectiveness, and dependability of mineral identification and separation procedures are enhanced, resulting in better utilization of resources and increased operational performance in mineral processing. The process of choosing the right wavelength is also extremely important to reduce the likelihood of incorrect positive and negative results, maximize the rate at which valuable minerals are recovered, and to reduce the waste stream losses of potentially valuable materials. The significance of ore sorting lies in its ability to facilitate efficient and precise separation of valuable ore from waste or gangue materials. Based on their unique reflectance or absorption properties, sensors can effectively distinguish ore from gangue by using specific wavelengths, such as visible or mid-infrared ones. This enhances the system's ability to choose and efficiently sort materials, especially when working with intricate ores or comparable substances. Utilizing wavelength selection can improve the ability of photometric sensors to distinguish between different substances and simplify the creation of new sensors for the purpose of sorting ores and characterizing minerals12. A variety of advanced techniques are used to analyze multidimensional spectrum data and extract relevant features from hyperspectral data, such as spectral features extraction and machine learning algorithms i.e. linear regression, K-means clustering, neural network13,14,15.

The intricate nature and extensive dimensions of multi-spectral data require the application of sophisticated data classification techniques such as Neighborhood Component Analysis (NCA). Advanced data classification techniques like NCA are needed to handle multi-spectral data due to several reasons. To begin with, hyperspectral data typically includes a substantial number of spectral bands, which might provide computing difficulties during processing and analysis16. The issue can be addressed by using NCA, which involves lowering the dimensionality of the data. This would lead to improved processing and classification efficiency17. Additionally, it is essential to note that conventional classification methods designed for multispectral data may not be appropriate for hyperspectral data, as the latter offers more intricate and comprehensive spectral information18. The NCA method can effectively handle hyperspectral data with a high number of dimensions. It achieves improved classification accuracies by taking into account both spectral and spatial information19. Additionally, NCA offers advantages such as low computational requirements and shorter classification times20. Therefore, advanced techniques like NCA are essential for accurately classifying hyperspectral data while overcoming the challenges associated with high dimensionality and detailed spectral information.

In this study, Neighborhood Component Analysis (NCA) was applied as a preprocessing step to reduce the dimension of Hyperspectral (HS) data of arsenic-bearing minerals by identifying several wavelength bands important for mineral identification. Then the identified wavelength bands were used as inputs to train machine learning algorithms for identifying Arsenic (As) minerals concentration in simulated ore materials. Multispectral (MS) cameras are more cost-effective and provide faster data collecting and processing compared to HS cameras; hence, they are projected to enable mineral identification utilizing data from a few wavelength bands. The HS data of arsenic-bearing minerals (enargite) were used NCA a machine learning method, as a band selector, and identified several wavelength bands important for mineral identification. Then, the data containing only the minimum number of wavelengths were analyzed for identification of mineral contents/ratios using machine learning algorithms. These will improve the selectivity of wavelengths, considering the ore characteristics produced by each mine. In addition, the application of the herein proposed machine learning algorithm for HS images analysis is expected to improve the efficiency of ore selectivity, i.e. improve the speed of the ore sorting process.

To develop environmentally sustainable resources, its essential to develop advanced metal recovery technology for these high-grade arsenic ores, and Sensor-Based Ore Sorting (SBOS) can achieve this. SBOS, when implemented as a presorting process before the normal beneficiation process, can reduce the amount of ore that must be processed to produce a certain amount of value-added metal, which has a significant impact on the economics of the mine and the plant as a whole21,22,23. It can also reduce the environmental impact by reducing the tailings produced in the subsequent beneficiation process. Non-destructively classified tailings are geotechnically stable and can be easily stored due to their low moisture content24. Robben and Wotruba highlighted that the introduction of SBOS would have an impact on both the environmental and economic aspects of the mineral processing process25. However, the authors pointed out that SBOS is still in the market entry stage of the mineral industry and further technological development is required.

Mineral analysis requires knowledge of crystallography as well as chemical analysis26. However, the methods commonly used for mineral analysis, such as Electron Probe Micro Analyzer (EPMA), X-ray diffraction (XRD) and Scanning Electron Microscope (SEM), are relatively time-consuming and depend on experience27, they are not realistic in terms of identification speed, convenience, and economy when used in actual mineral processing operations.

Therefore, SBOS has been developed as a form of mineral identification suitable for beneficiation. In recent years, more and more equipment has been installed that can withstand larger production scales25. SBOS methods have utilized a range of sensing technologies, including X-ray transmission, X-ray fluorescence, optical sensing, and inductive sensing9,10,28. Furthermore, the utilization of area-scan cameras and predictive tracking systems that rely on machine learning approaches have demonstrated potential in minimizing characterization and separation errors29. Researchers have also studied the combination of data from several sensing approaches to improve the sorting ability of these systems6. While different SBOS methods have been developed and introduced particularly focusing on SWIR data, there are few studies or methods on mineral identification/sorting using VNIR short-wavelength HS data. However, there is growing interest in visible to near-infrared (VNIR) spectroscopy for mineral identification, and in some recent studies VNIR wavelengths have been used to classify rocks and minerals30,31.

Sensor-based ore sorting methods and technologies have the potential to significantly improve ore grades and reduce waste in mineral processing6. These methods, which rely on the electromagnetic spectrum, can be classified based on their characteristics and limitations28. An example of a successful method for sorting complicated ores is the utilization of hyperspectral short-wave infrared sensors in conjunction with machine learning, as demonstrated by Tusa32. Sensor-based ore sorting can be applied at various stages in the process flow diagram, making it a versatile and valuable tool in the mining industry33.

In the field of remote sensing, mineral identification in the near-infrared region has been widely used34,35 and they have shown excellent performance in ore classification. On the other hand, the high cost of HS cameras and the time required for data acquisition have been barriers to their application in actual operations, where immediate classification is required. In a previous study24,36, HS data of minerals were analyzed by deep learning to identify minerals. The use of deep learning allows the creation of more versatile and simplified learning models compared to conventional machine learning or identification methods that combine multiple machine learning models. However, since HS images consist of several hundred spectral bands, there is a high correlation between proximity spectra, and data analysis without preprocessing is highly redundant and computationally intensive. Therefore, dimensionality reduction is necessary as a preprocessing step for a large amount of data to be generated.

Dimensionality reduction methods for HS commonly fall into two categories: band extraction and wavelength selection. The band extraction methods map a high-dimensional feature space to a low-dimensional space; therefore, cannot preserve the original physical interpretation of the image and is not applicable as a dimensionality reduction method37. While the wavelength selection method can maintain the original physical interpretation of the images. According to a review by Sun and Du38, wavelength selection methods can be categorized into six groups: ranking-based methods, searching-based methods, clustering-based methods, sparsity-based methods, embedding learning-based methods, and hybrid scheme-based methods.

A variety of studies have explored different techniques for wavelength selection and spectral data classification in mineral processing. Ghosh39 introduced an infrared thermography-based method for sorting alumina-rich iron ores, while Kern40 suggested utilizing short-wavelength infrared and dual-energy X-ray transmission sensors for the Hammerlein SnInZn deposit. Phiri41 investigated the potential of near-infrared sensors for separating carbonate-rich gangue from copper-bearing particles in a coppergold ore sample. Tusa32 advanced the field by evaluating hyperspectral short-wave infrared sensors, combined with machine learning methods, for pre-sorting complex ores. These studies collectively illustrate the potential of various wavelength selection techniques for enhancing the efficiency and effectiveness of ore sorting systems.

Furthermore, numerous research endeavors have investigated the implementation of machine learning algorithms to automate the task of wavelength selection and spectral data classification. Passos42 introduced an automated deep learning pipeline to optimize neural architecture and hyperparameters for spectral classification. Duan43 proposed a template matching approach achieving high accuracy without training, while Wang44 developed a multifunctional optical spectrum analysis technique utilizing support vector machines for optimal accuracy and speed. Baskir45 presented a MATLAB toolbox for user-friendly wavelength selection and automated spectral region selection. These investigations collectively underscore the potential of machine learning in automating and enhancing the process of wavelength selection and spectral data classification.

In addition to this, Advancements in hyperspectral imaging technology have significantly expanded the potential applications of this technology46. However, the complexity of hyperspectral data, including its high dimensionality and size, requires innovative methodologies for effective processing and analysis47. These challenges have led to the development of a range of image processing and machine learning analysis pipelines46. Notably, hyperspectral imaging finds application in microscopy, enabling the capture and identification of different spectral signatures in a single-pass evaluation48.

The effectiveness of machine learning algorithms, particularly Neighborhood Component Analysis (NCA), for multi-spectral data classification in mineral processing has been highlighted in recent research. Jahoda49 and Sinaice50 both emphasize the advantages of combining spectroscopic methods with machine learning for mineral identification. Jahoda49specifically highlights the superiority of machine learning methods in this context, while Sinaice50 proposes a system integrating hyperspectral imaging, NCA, and machine learning for rock and mineral classification. These findings are further supported by Carey51, who stresses the importance of spectrum preprocessing and a weighted-neighbors classifier for optimal mineral spectrum matching performance.

In their previous study Okada et al.24, developed a basic technology of SBOS, using hyperspectral (HS) imaging and deep learning as an expert system for mineral identification. HS is promising as SBOS to avoid As-containing copper minerals technic instead. In that study, HS imaging was used as a sensor to collect the intensity of wavelength data, which was then used to train deep learning algorithms for mineral identification. The HS image is cube-shaped data with dimensions in the wavelength and spatial directions, with wavelength data from the visible to near-infrared regions (400~1000nm, 204 bands). Minerals (hematite, chalcopyrite, galena) identification was performed by analyzing detailed wavelength data of 204 bands in the shorter wavelength range of 4001000nm (from the visible light region to a part of the near-infrared region) using deep learning. However, the HS data used in that study consisted of 204 high-dimensional data, which required heavy computational resources. In addition, the HS camera itself is expensive, which was a barrier to its introduction/implementation in the operating site (mineral processing plant).

Yokoya and Iwasaki52 reported that, since each pixel provides continuous electromagnetic spectral characteristics, its possible to obtain detailed information about the target object. Owing to the high spectral resolution HS imaging is applied in fields such as remote sensing and quality control of food deep and pharmaceuticals. Robben et al.53 pointed out that, minerals show specific absorption characteristics in the near-infrared region from 1300 to 2550nm due to vibrations of the bonding molecules contained in each mineral. A skilled expert can identify some minerals visually (Color), and the continuous electromagnetic spectrum in the short wavelength region is considered to contain optical data with mineral-specific physical properties.

Sorting machines that use MS images with a reduced number of dimensions are now technically feasible. They sort by acquiring specific wavelength information predetermined for each ore type. However, even for the same type of mineral, there are subtle differences in the formation and shape of each mine that affect the spectra. Additionally, the light environment inside each plant varies, which also affects the spectrum. Based on these factors, it is suggested that ore selectivity could be improved by defining the wavelength to be acquired for each ore type. To achieve this, we propose a framework that allows for the selection of spectral bands based on the characteristics of the ore produced. This framework will greatly support the tuning of the sorting process. As a case study, we will use a mineral sample containing arsenic.

The literature review highlights various gaps in current mineral processing practices, emphasizing the need for innovative approaches to improve efficiency and sustainability. While Sensor-Based Ore Sorting (SBOS) offers promise for environmentally friendly metal recovery, further technological development is required to enhance its effectiveness and practicality in operational settings. Traditional mineral analysis methods are time-consuming and impractical for real-time processing, prompting the exploration of faster and more economical techniques. Additionally, the application of machine learning algorithms and hyperspectral imaging for mineral identification presents computational challenges and limitations in the practical implementation due to the high dimensionality of data.

In response to these challenges, the proposed framework integrates Neighborhood Component Analysis (NCA) and machine learning algorithms to address the complexities of mineral identification and sorting using multi-spectral data. By reducing data dimensionality and identifying crucial wavelength bands, the framework enables efficient mineral identification while considering the unique characteristics of each ore type. Furthermore, by utilizing multi-spectral cameras with reduced dimensions, the framework enhances sorting efficiency and selectivity, paving the way for more sustainable mining practices and improved operational outcomes in mineral processing plants. In this study, a clustering-based method, Neighborhood Components Analysis (NCA), was used to perform dimensionality reduction and wavelength selection on HS data. After the selection, the selected bands were learned by machine learning algorithms to experiment with mineral identification. It is expected that mineral identification using fewer wavelengths than HS data will enable data acquisition with less expensive MS cameras and increase the efficiency of mineral identification.

The rest is here:
Optimizing multi-spectral ore sorting incorporating wavelength selection utilizing neighborhood component analysis for ... - Nature.com