Open Access
Add to BibSonomy
Abstract
The real-time online quantitative analysis instrument is highly desirable for many industrial fields. Herein, a new laser-induced breakdown spectroscopy (LIBS) setup with optimized optical route and high accuracy algorithm is designed and applied in a real industrial site. The components of total iron (TFe), silica (SiO2), aluminum oxide (Al2O3), and phosphorus (P) are quantitatively determined by the online LIBS system. The key optical part is a Maksutov-Cassegrain telescope, in which, two aspherical mirrors are specially designed and fabricated to reflect the broadband emission from ultraviolet 240 nm to infrared 890 nm with reflectivity over 90%, and pass the excited laser line of 1064 nm. The system could automatically adjust the focal length in the range of 780 mm to 940 mm. Based on the online LIBS system, the spectral pretreatment algorithm is also optimized including baseline removal and spectral normalization. The overlapped window slide (OWS) algorithm avoids the deformation of emission peaks in spectral baseline removal, in addition, two normalization steps by total back area and total spectral intensity within the sub-channel are applied to improve the spectral data stabilization. The calibration and validation are performed by utilizing the emissions that are insensitive to the detection distance. Compared with the traditional method, the prediction result shows that the root of mean square error of prediction (RMSEP) decreased from 5.091% to 1.2328%, and the mean absolute error (MAE) reduced from 4.801% to 0.9126% for TFe. Eventually, the online measurement shows good agreement with the official standard results. The high-precision online determination system based on LIBS will upgrade low frequency sampling of traditional detection to high-frequency real online determination in many industrial fields.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Customs and ports always need to quantitatively and rapidly analyze elemental content in raw iron ore, the total iron and other matrix matter affects the production process and the settlement price in steelworks. Generally, a docked cargo ship carries several hundred thousand tons of ore, and the conveyer belt connects the ship and steel mill directly for highly efficient unloading. The traditional quantitative analysis in conventional laboratories mainly involves sampling every 100 tons of ore from a belt conveyor, using traditional methods for offline post-processing and analysis, including atomic absorption spectroscopy, X-ray fluorescence, or chemical titration, which always take several hours or days [1]. Consequently, the low sampling representativeness and the significant feedback delay hinder the management of steelworks to refine production processes.
Laser-induced breakdown spectroscopy (LIBS) is a suitable online analysis method with fast and real-time quantitative capabilities. The high sampling frequency provides a significant statistical analysis result of mineral on the belt conveyer. Many researchers have successfully conducted industrial at-line analysis of various minerals based on LIBS [2–4]. However, most at-line analysis processes mainly rely on auxiliary sample processing equipment located near the main belt conveyor [5], sampling the block from the material flow and reprocessing the raw material to appropriate state for analysis. For example, in the coal industry, coal blocks are usually taken from conveyors and then crushed and granulated to a diameter less than 0.2 µm particles, and then press the coal powder into pellets under dozens of tons of pressure [6–8]. For each pellet sample, ablation hundreds of times in a certain sequence by the laser pulses. Finally, the spectral preprocessing and statistical algorithms of quantitative analysis are combined to measure the elemental content in coal. Compared to traditional analysis based on manual operation in the laboratory, the online determination save time and labor, but it is still a huge challenge to directly analyze the material flow on the moving belt. As to the iron ore in the port, unloading and transporting is synchronized, reprocess raw ore by sampling from the belt conveyer is not suit for the end-to-end transportation. Therefore, it is necessary to directly embed the LIBS analyzer online and analyze the high-speed moving raw ore on the belt. However, the complexity of industrial sites and the uncontrollability of the raw ore state have become new challenges for online LIBS solutions, such as the mixing of fly ash on material stream and iron ore with different sizes, which greatly affects the availability and stability of spectral data, thereby to reduce the precision and accuracy.
Due to the irregular shape of the raw ore and changes in the height of material stream on the conveyor, most online LIBS applications need automatic focusing components. Coaxial optics schemes, such as Maksutov-Cassegrain telescope architecture with reflective optics or other remote analysis configurations rely on Galilean telescope configuration with refractive optics, are considered as the effective solution. Laserna and his workmates designed a remote sensor based on an autofocusing system to detect soil samples twenty years ago, the expander consisted of two lenses, and meanwhile, the focusing length was adjusted by those two lenses [9]. Patrick developed an online LIBS scanning system by the refraction-optics-based autofocusing system, they measured the aluminum alloy on a moving belt with a speed of 3 m/s [10,11], but the chromatism was not considered in such system. Reinhard et al. carried out an online LIBS system and installed it above a moving coal conveyer [4], the system measured the ash content with a relative small range of 1.5% ∼2.8%. Redoglio et al. combined the Galilean system and two off-axial mirrors to measure the raw coal block [12], the optical system has a large depth of field with 20 cm by which the laser fluence irradiate on the different rock materials maintain stable. A coaxial optical configuration was developed for mineral quantitative analysis in Josette’s work [5], they use the resulting data of EDS-SEM to guide the results obtained by LIBS, and the RMSE of prediction result presented below 10% for the main minerals. In addition, there are some researchers developed integrated commercial LIBS units for qualitative or quantitative analysis of raw minerals online. M. Gaft et al. developed a machine named MAYA based on LIBS technology to evaluate coal bulk on a moving belt conveyer [13], the online instrument revealed consistent results compared with prompt gamma neutron activation analyses (PGGAA) for ash analysis online, the absolute error of 2% ∼ 4% for ash measurement, but the correlation between online system and local laboratory is worse [14]. The online analyzer MAYA is also used to monitor the content of Cu, CaO, SiO2, and S for Cu-bearing minerals, and to determine the Fe and CaO content in Fe-based ore [15]. The most famous one is the LIBS analysis device made by NASA(SuperCam) that is carried on the Mars rover [16], it is a special space online device, and has many valuable technologies to learn from. SuperCam chose the Cassegrain telescope as the autofocusing system to detect remote soil material [17], the RMSEP of main composition SiO2 on Mars’ soil was 6.1% [18,19]. Until now, the reported online LIBS configurations still have lower accuracy. Especially, the optical components with both high damage threshold and broad spectral reflectivity still need to be further enhanced. In brief, a universal optical system with related high-accuracy algorithm is still missing for the real on-line detection.
Herein, considering the complexity and physicochemical state of the raw materials, we developed an online LIBS system for quantitatively determining the main components in iron ore on a fast-moving belt conveyor with the speed of 4∼4.5 m/s. On the basis of optimizing the spectral continuum fitting algorithm, two-step spectral normalization was carried out to reduce the fluctuation of spectral signals. In addition, based on the study of the variable importance of projection (VIP) in modeling, we found that certain emissions are not sensitive to detection length during autofocus. Finally, based on these insensitive emissions, the online machine was calibrated and the model was applied online. The predicted results of the online LIBS analyzer are compared with the official report, indicating that the chemical abundances of TFe, SiO2, Al2O3, and P are in good agreement with the routine laboratory.
2. Online LIBS analyzer framework
2.1. Optical system design
For high-efficiency and cost reduction, raw ore is transported directly from the cargo ship to the steel mill by the belt conveyer in the port, the running velocity of the belt is usually set at 4∼4.5 m/s. It is significantly different that the size of raw ore from the crab bucket dumped into the belt conveyer, which is the basis of the design for autofocusing range of the optical system. The online LIBS analyzer OL-FeCam is designed to quantitatively analyze iron ore online based on the autofocusing optical system. As listed in the Table. 1, the parameters demonstrate the autofocusing range, that is the working distance of the optical system, should be able to automatically adjust from 780 mm to 940 mm.
For optical components used in online LIBS systems, refractive optical lenses are often prone to involve chromatism and aberration, especially for optical systems with long focal lengths. Therefore, reflective optics (without achromatic aberration) are necessary for signal acquisition. To overcome these issues, we chose reflective optics based on the Maksutov-Cassegrain architecture as the initial design input for the optical theme. Due to the fact that the size of the focused laser spot needs to maintain stability with changes in focal length, two aspheric mirrors M1 and M2 were designed for the minimum axial aberration under different focusing lengths, as shown in Figs. 1 (a) and (b).
Fig. 1. (a) The design scheme layout of online LIBS analyzer, L1&L2: beam expander, L3: collection lens, D: dichroic with reflection wavelength 240 nm∼ 950 nm and transmission wavelength 1000 nm ∼ 1200 nm. (b) Enlarge view of two aspheric mirrors, M1: primary mirror, hyperboloid, M2: secondary mirror, paraboloid with a hole in the central. (c) The α-prototype machine in the laboratory. (d) The online LIBS analyzer installed above the belt conveyer. (e) The optical and mechanical schematic view inside the analyzer. (f) The electronic control system built into the bones of the machine.
Note that there is a large amount of emission in the ultraviolet region of the spectrum, and quantitative analysis based on the spectrum requires collecting plasma emission signals within the range of ultraviolet to visible light. Therefore, the optical components related to spectral signal collection, including two main mirrors M1, M2, and dichroism D, are HR coating relative to the ultraviolet region. In addition, the dielectric protective film layer must also be attached to the HR coating to meet the high damage threshold requirements, as listed in the Table 1. In other words, optical components correlated to reflective surfaces are typically customized to suit industrial sites.
Table 1. The optical design requirement of online ore LIBS analyzer
2.2 Prototype α-machine
As shown in Fig. 1(c), the prototype machine named α-Online Fe-Cam (hereinafter referred to as α-machine) is established in the laboratory. The beam with a diameter of 6 mm is output from a Q-switched Nd: YAG laser (Ziyu, Penny 300A, 1064 nm, China), and the laser pulse duration is 6 ns. The output laser energy is 300 mJ, and the laser fluence is approximately 1.06 J/cm2 with the beam diameter is 6 mm. The energy irradiates on the sample surface is 195 mJ. The laser beam is expanded to reduce the fluence to protect the coating layer of the mirror M1, the beam expander consisted of two quartz lenses L1(focus length f = -19 mm) and L2 (f = 50.8 mm) with a magnification of ×2. The radiation emission of plasma is coupled to a five-fiber cable (5 × 200 µm), and then transmit the light to the spectrometer (Avantes, AVS-DESKTOP-USB2 StarLine, wavelength range 179.68 nm to 893.83 nm with the resolution less than 0.1 nm of each channel). The integration time of the spectrometer is set to 30 us. The delay time between the spectrometer and the laser pulse is 0.5 µs. The autofocusing is adjusted by moving the mirror M1 along the optical axis, the working distance, i.e., the detection length from mirror M2 to the sample surface is 860${\pm} $80 mm corresponding to the separation of 95.8${\pm} $1.8 mm between the mirror M1 and M2.
2.3 Control logic design of automatic focusing system
In the laboratory, the prototype α-machine is constructed to validate the quantitative algorithm, the schematic diagram is demonstrated in Fig. 2. The autofocusing system is controlled by two microcontroller units (MCUs) C1 (Arduino UNO, ATMEGA 328P-AU, Italy) and C2 (Arduino Mega, ATMEGA 2560, Italy). The motors Z2 and Z3 control the sample stage T moving within the x-y plane, which is directed perpendicular to the paper surface, and the position of T in the z-direction is controlled by a slide below the sample stage. The motor Z1 holds the mirror M1 and drives the mirror along the optical axis, the focal length changes as Z1 moved. The distance sensor S (Omron, ZX-LD300, Japan) is installed behind the mirror M1 to measure the distance between stage T and M2. Combining signal filter (SF) and the 16-bit high-speed analog-to-digital converter (ADC) read and preprocess the distance sensor data. The control panel is shown in Fig. 1(f), which includes two MCUs, SF, ADC, and three motor drivers. (the detail of control sequence is available in Supplement 1, part IV)
Fig. 2. The control principle layout of online LIBS system, Z1∼Z3: motor controlled moving stage, a1∼a4, b1∼b4, c1∼c4, d1∼d4: I/O pin on the microcontroller unit, SF: signal filter, ADC: analog to digital converter, T: sample stage, S: distance sensor, M1/M2: aspheric mirror, L1∼L3: lenses, D: dichroic, F: fiber cable, CLC-IN: clock trigger input of laser, Q-SYNC: Q synchronization signal output of laser.
The α-machine is initialized based on the distance sensor S, and the initial position of T is set to 860 mm far from M2. As long as the detection distance changes, motor Z1 is activated. According to the Gaussian imaging formula, the working focus length of M2 is altered by the position adjustment of M1. If the position of M1 is suitable for focusing, the laser starts working through a trigger signal from pin a2. The moving stage in the plane of x-y moves 1 step after five laser pulses ablate the sample, as indicated in Fig. 2.
3 Sample and methods
3.1 Sample preparation
A total of fifty samples are chosen to calibrate the machine with an initial working distance of 860 mm, the details of the sample’s chemical compositions, i.e., TFe, SiO2, Al2O3, and P, are listed in Table S1 (see Supplement 1), and the reference abundance of those four components is analyzed by chemical titration. All samples are grained to powder with a particle diameter of less than 200 µm, then mixing potassium chloride (KCl, the binder) with iron powder in a weight ratio of 3:7 to press the powder sample into a pellet, which has a diameter of 30 mm and thickness of 5 mm under the pressure of 25 tons. The step program controls the sample stage T moving forward 1 mm every five laser pulses and yielding an ablation point grid with 7 × 6 points on the sample surface. 210 spectra are acquired for each sample.
3.2 Spectra pretreatment
The spectral data of each sample are preprocessed in following several steps. Firstly, the spectra must be standardized, which means that effective data should be selected based on the statistical criteria, as described in our previous work [6,20]. Based on the Eq. (1) ∼ Eq. (3), around 70% of spectra are remained after data selection.
Subsequently, the continuous background are removed based on the overlapped-window sliding (OWS) method, which is mentioned for the first time. Traditionally, window sliding (WS) method allows a widow with the width w sliding on the spectra data by a fixed sliding step length s, then the minimum in each window is marked, and the continuous background can be fitted by those minimums [21]. A larger window width can result in incomplete background fitting, while a smaller window width can lead to excessive spectral intensity being misjudged as the background. In general, the window width and the step sliding length are set to be equal. For example, the window width and the sliding step length are set to 64 pixels for the 8192 pixels of total spectral length. However, abnormal background fitting still appears on both sides of the peak, as shown in the Fig. 3(a) and (b). The red dashed box in Fig. 3 (a) marked an abnormal continuum fitting, the essence of the phenomenon is that the fitting background is based on insufficient data points due to the larger step width of the sliding window. The magenta dots mark the local minimum values missed due to the wider window, which lead to a concave aside the emission peak after the continuum removal, as shown in Fig. 3(b). Therefore, the overlapped widow sliding (OWS) method is proposed, and a brief principle is illustrated in Fig. 3(c). This method allows for the overlap between various windows during the sliding processes, w can be larger than s, indicating more local minima are taken into account in background fitting while maintaining a wider width of the sliding window. The OWS method removes the background as shown in Fig. 3(d). After removing the continuous background, the shape of the peak does not deform or distort, which is of great significance for quantitative analysis using peak shape related intensity as a variable in modeling [20].
Fig. 3. The comparison of window slide method in spectral background removal (a)(b) Traditional method that the width of slide window is equal to the slide step ($w = s$) (c)(d) refinement method that the width of slide window is large than the slide step ($w \ge s$)
To reduce the fluctuation of spectra, normalization is applied in the third step of spectral data preprocessing. Since the complexity on the industrial site is easier to cause signal uncertainty, the normalization is necessary for online LIBS systems. Some researchers have proposed multiple normalization methods to preprocess spectral data, such as the normalization of reference elemental lines [22–24], normalization of integral background strength in subchannel [21] and whole spectrum area [25,26]. In this paper, two-step spectral normalization method are performed. Firstly, the continuous background of the spectrum in the subchannel is firstly utilized to normalize the spectral intensity, because the signal fluctuations with the continuous background are mainly induced by the bremsstrahlung radiation, which is correlated to the electronic temperature in the plasma [27]. It demonstrates that the spectral signal is normalized to a certain stable plasma state. The secondary normalization step uses the whole spectral strength in each subchannel to normalize the spectrum. The standard deviation of spectral intensity in each pixel is utilized to evaluate the stability of the signal according to formula Eq. (4),
Fig. 4. (a) The comparison of STD before and after two steps of normalization. (b) The eigenvalues of the first three components in raw data and normalized data.
From the perspective of principal components analysis (PCA), data points with poor consistency are distributed like a slender ellipsoid in the feature space, the distribution of projection points on the first component is clearly distinguished from that in other component directions, which means that the size of the first eigenvalue is significantly larger than other eigenvalues [28]. However, data points with high consistency have a distribution closer to a spherical or fat ellipsoidal in high-dimension space, the distribution of projection points cluster on each component [29,30]. As shown in Fig. 4(b), the smaller fluctuation of one dataset in variable space means that the closer eigenvalues of different feature vectors in the feature space. Before normalization, the eigenvalues of the first three components are 5113, 913, and 575, respectively. Correspondingly, the eigenvalues of the first three components after spectral normalization are 1462, 912, and 591, respectively. The first principal component dominates the explanation in the spectral dataset, and the second and third eigenvalues are numerically close before and after spectral normalized. However, the eigenvalue of the first component is closer to the others in the normalized spectral dataset. Geometrically, the variance of data projection is significant in the direction of the first eigenvector compared with others. The closer eigenvalues of the first three principal components after spectral normalization illustrate the variances of data projection in each direction of eigenvectors are approximate, which implies the reduction of variation information in data. Therefore, the principal component analysis demonstrates the data consistency is enhanced after the spectral normalization.
4 Online quantitative determination
4.1 Calibration of the autofocusing system
The spectral data of 50 samples are split into calibration dataset and test dataset by a ratio of 7:3 after data pretreatment. Thus, the calibration and test dataset are consisted of 38 and 12 sets of samples, respectively. Then the averaged spectra of samples are calculated and delivered to the modeling algorithm based on the principal component analysis based partial least squares method (PCA-PLS), which is a multiple variable regression method based on PCA dimension reduction and features extraction [31,32], and the hyperparameter is determined by the 5-fold cross-validation. The instrument is calibrated with the working distance of 860 mm, the calibration result is shown in Fig. 5(a). The coefficient of determination R2 and root of mean square error (RMSE) of TFe are 0.99, 0.3610%, and 0.95, 0.5987% in the training and test dataset, respectively.
Fig. 5. The calibration results of four compositions in the ore: (a)TFe, (b) SiO2, (c) Al2O3, and (d) P.
The modeling results of SiO2, Al2O3, and P are shown in Fig. 5(b)∼(d) and Table 2.
Table 2. The modeling results in training and testing dataset
Since the collection solid angle of plasma emission constantly changes along with the focal length automatic adjusting, which will lead to the spectral intensity of the same sample varied (see Supplement 1, Fig. S1). For our setup, when the focal length change is less than 10 mm, the spectral intensity does not change much, indicating that the focal depth of the optical system is ±10 mm. Therefore, if the detection focal length changes larger than 10 mm, the calibration curve established in the initial position, i.e., the working distance of 860 mm, is invalid to predict the industrial parameters of the ore in other working distances. Thus, to calibrate the instrument in different focal points every 20 mm can overcome the influence from autofocusing. However, it is always laborious and abandoned in the industrial site. From another perspective, it is crucial that some emissions are not sensitive to working distances, which means that the instrument can be calibrated at its initial position through these special emissions and the calibration model can be applied to other different working distances. Here, a totally new simple calibration method is proposed to calibrate the online-LIBS instrument quickly and conveniently based on certain spectral emission lines.
In order to investigate which emission lines that are not sensitive to working distance, sample # 47 was selected for testing at 10 different working distances, namely 780 mm, 800 mm, 820 mm, 840 mm, 860 mm, 880 mm, 900 mm, 920 mm, 940 mm, and 960 mm. The testing was repeated three times at each detection location. We selected spectra obtained through three different detection positions to show the difference of emission intensity, with the farthest detection position being 960 mm, the secondary collection solid angle being 0.034 sr, the calibration position being 860 mm, the angle being 0.169 sr, the closest position being 780 mm, and the maximum solid angle being 0.205 sr. As shown in Fig. 6 (a), there are significant differences in the original spectral intensity at different working distances, including peak intensity and spectral background. In the wavelength range of 240 nm to 450 nm, the spectral intensity significantly decreases as the working distance increases from 780 mm to 960 mm. Figure 6 (b) shows an enlarged spectrogram within the range of 355 nm to 372 nm, showing detailed emission intensity differences. After spectral normalization, the emission intensity differences of each spectrum become indistinct, especially in the wavelength range of 240 nm to 450 nm, as shown in Fig. 6 (c). Compared to Fig. 6 (b), the emissions intensity at three different working distances are consistent after normalization, as shown in Fig. 6(d). However, the spectral intensity varies within the wavelength range of 450 nm to 800 nm, indicating that these emissions are sensitive to working distance.
Fig. 6. The comparison of spectra obtained in three different working distances (WD = 780 mm, 860 mm, and 960 mm) before and after normalization refinement (a)(b) the original spectra and the corresponding local enlarged drawing (c)(d) the normalized spectra and the corresponding local enlarged drawing
In order to determine which emissions are suitable for calibration at the initial position, thirty sets of spectral data of sample # 47 are classified using the Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) algorithm [33,34]. According to the different detection distances, three sets of tests at the same distance are regarded as the same group, with ten groups corresponding to ten different working distances. The OPLS-DA method based on PCA considers the maximum variance in the variable score space and captures the variance between each group. Similar to the regression modeling process, spectral data is used to train classification models after spectral preprocessing, and the 5-fold cross-validation determines the optimal model. As shown in Fig. 7(a), the score plot based on the OPLS-DA demonstrates the spectral data obtained under the same detection distance cluster together, the digits around the pentagon are the identifier of tests in the same detection position. The modeling result demonstrates that the accuracy of classification is 90%, the detailed classification result is evaluated in the confusion matrix (see Supplement 1, part II). The importance of variables in the projection (VIP) [31] in the classification model is utilized to determine which emission peaks influence the class model to a great extent. The emission peaks with VIP greater than one in the classification model ($VI{P_C} > 1$) are marked by the blue square in Fig. 7(b).
Fig. 7. (a) The OPLS-DA based score plot has five components and two orthogonal components used in the modeling. (b) The emission peak-marked plot of $VI{P_R} > 1$ (VIP in the TFe-regression model greater than 1) by the red star sign and the emission peaks of $VI{P_C} > 1$ (VIP in the classification model greater than 1) marked by the blue square.
Correspondingly, in the prediction model of TFe calibrated under the working distance of 860 mm, the emission peaks with VIP greater than 1 ($VI{P_R} > 1$) are marked with the red stars in the figure. Note that most of the peaks that have a significant impact on both models are separated, with only a few emission peaks repeating in the classification and regression models. The essential emissions in the regression model are all in the wavelength range of 240 nm to 450 nm, with a small portion of emissions contributing to the classification model. According to Fig. 6(c), there are no emission peaks are correlated to the regression model in the wavelength range of 450 nm to 600 nm, as their intensity is sensitive to working distances. Consequently, reconstructing the prediction model of TFe only using the emissions with $VI{P_R} > 1$ (excluding the peaks with $VI{P_C} > 1$), the model robustness will be improved as the working distance changes (the VIP information about other composition’s model is available in Supplement 1, Fig. S2). With the autofocusing system working, 10 samples are randomly selected to validate the performance of the original and the reconstructed models, and the samples under six different working distances of 780 mm, 820 mm, 860 mm, 900 mm, 940 mm, and 960 mm are tested, respectively, as shown in Fig. 8(a). Before model reconstruction, the RMSEP of the original model (OM-RMSEP) remains minimum at the calibration point of 860 mm and gradually increases as the detection distance moves away from both sides of the calibration position. The model performance becomes poor and invalid at the extreme working distance of 960 mm and 780 mm, the maximum OM-RMSEP = 5.0908%, and OM-ARE = 4.8006% at the farthest working distance of 960 mm. On the contrary, after reconstructing models using distance insensitive spectral emissions, the RMSEP of the reconstructed model (RCM-RMSEP) and RCM-ARE remained around 1%.
Fig. 8. The model performance comparison by RMSEP and ARE before (OM-RMSEP and OM-ARE) and after (RCM-RMSEP and RCM-ARE) model reconstructed: (a) the model of TFe, (b) the model of SiO2, (c) the model of Al2O3, and (d) the model of P.
The prediction results of SiO2, Al2O3, and P are shown in Fig. 8(b)∼(d), no matter what the working distance is, the OM-RMSEP and OM-ARE of the original SiO2 model are greater than 2% and 1.5%, respectively. With model reconstruction, all predictions of key components have been improved. For SiO2, both RCM-RMSEP and RCM-ARE were reduced to around 1% by the reconstructed model. For Al2O3, the RMSEP and ARE of the original model and reconstructed model remain at 0.22% to 0.3% at the working distance of 860 mm, showing no significant improvement before and after the reconstruction. Nevertheless, the OM-RMSEP and OM-ARE dramatically increase with the working distance increasing and decreasing. In particular, RCM-RMSEP and RCM-ARE maintain small levels of 0.3% to 0.6%, indicating that the reconstructed model mainly relies on peaks that are not sensitive to changes in working distance. The result of P is similar to that of Al2O3. The prediction results of four components at the working distance of 780 mm are listed in Table 3 and other results at different working distances are available in Table S2 (see Supplement 1). It can be found that the prediction errors for different components are all lower enough, indicating the refinement algorithm can be well used in any industrial online site.
Table 3. The prediction results of the original model and reconstructed model at the WD = 780 mm
4.2 Calibration and online validation
For online validation, the optimal focal distance of the online LIBS analyzer OL-FeCam is firstly determined through a distance sensor S that is installed approximately two meters ahead, as shown in Fig. 1(d) and (e). The detection point of the distance sensor is collinear with the laser point, ensuring that the excited point is the point detected by the distance sensor. Therefore, the system has enough time to adjust M1 to the proper position. The sampling frequency is set to 1 Hz, and quantitative results are output once per second. The output result represents the average level of every 1.8-ton iron ore. Compared to the traditional quantitative sampling method for every 500 tons of iron ore, the online LIBS analyzer has increased the sampling frequency by nearly 265 times.
Based on the above-mentioned spectral pretreatment algorithm, the quantitative model of the online LIBS instrument is constructed under the working distance of 860 mm. The calibration procedure for the online system OL-FeCam use the same 52 samples. Note that the models for the four chemical components TFe, SiO2, Al2O3, and P are only calibrated by the emissions that are not sensitive to working distance.
For the measurement of ore blocks on the belt conveyor, the automatic focusing system is activated with the help of distance detection sensors. Continuously collect the spectra of the mineral block online at a laser repetition rate of 1 Hz for 90 minutes. Therefore, approximately 104 -ton ore block is transported to the steel factory within ninety minutes. For different types of ores, the total ion content in the ore varies from 50% to 70%, and the width of the belt and the thickness of the ore flow on the belt average 0.5 m and 0.2 m. In order to validate the predictive performance of the online LIBS system, the results of online configuration OL-FeCam are compared with those of traditional chemical titration in routine laboratories. As shown in Fig. 9(a), the TFe prediction results of three kinds of ore A, ore B, and ore C by OL-FeCam are marked by gray dots, and the colored line is the smoothed result of the gray dots per minute. The black dashed line in each subfigure shows the average value obtained by chemical titration in the routine laboratory, and the average value from online detection is always fluctuate along the line. The prediction results and comparison of SiO2, Al2O3, and P are shown in the Fig. 9(b)∼(d), respectively (as we show in Dataset 1 [35]). As conclusion in Table 4, the close value of average online prediction results (Mean.pre %) and laboratory results indicating the high degree of consistency between them.
Fig. 9. The result comparison between OL-FeCam and laboratory for the three different kinds of ore (Ore A, Ore B, and Ore C) on the belt conveyer: (a) the result of TFe, (b) the result of SiO2, (c) the result of Al2O3, and (d) the result of P. The colored lines marked the result by averaging online measurement points in each minute, and the colored arrow marked the total average value in 90 minutes of online measurement. The black dashed lines and arrows marked the result given by the routine laboratory.
Table 4. Statistical evaluation of online prediction results of ore A, B, and C in 90 minutes
The application of LIBS technology in industrial sites requires careful consideration of the impact of environmental factors on various instruments. Equipment sealing and electromagnetic protection should be done well to ensure robust and high-precision measurement over a long period. Additionally, the physical state of different kinds of ore on the belt is different, the Ore A, B, and C used in the manuscript both are dry blocks. The methods mentioned in this work are conducted for the dry ore blocks which means the moisture content is less than 5%. If the moisture content exceeds 5% or even 10%, the laser energy 300 mJ used in this work is not suitable, the clammy ore blocks need to be dried in some other ways before LIBS measurement, it will be another research project that is worthing to further investigate.
5 Conclusion
In order to determine the chemical components in the raw ore online, the machine based on LIBS technology was developed and installed on the industrial site. The chemical content of total iron TFe, silica SiO2, aluminum oxide Al2O3, and phosphorus P in the ore were measured online. The optical system has been optimized to meet the requirement of height variation of the material stream with the working distance range of 860 ± 80 mm and a focus depth of ±10 mm. The customized reflectors cover the spectral range of 240 nm to 890 nm, meeting the requirements of online quantitative analysis with rich emission lines. To further ensure accuracy and precision, the refinement algorithm is further optimized for the extreme spectral fluctuation from online detection. Firstly, a novel baseline fitting method was proposed, which combined with two-step spectral normalization to overcome fluctuations in spectral intensity. Interestingly, during the autofocus process, some special emission lines that are not sensitive to the working distance were captured, and only these emission lines were subsequently used to calibrate and validate the models of various main components. Therefore, by utilizing these emissions, the RMSEP and MAE of the four compositions were reduced by more than 50%. Finally, spectral data was obtained continuously through an online system for 90 minutes, indicating a high degree of consistency between laboratory results and OL FeCam prediction results. The online machine truly realizes online analysis of ore in the industry sites with high sampling rate and accuracy.
Funding
National Key Research and Development Program of China (2018YFC2001100).
Disclosures
The authors declare no conflicts of interest.
Data availability
All data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request. Some data can be found in Dataset 1 [35].
Supplemental document
See Supplement 1 for supporting content.
References
1. R. Noll, C. Fricke-Begemann, S. Connemann, et al., “LIBS analyses for industrial applications – an overview of developments from 2014 to 2018,” J. Anal. At. Spectrom. 33(6), 945–956 (2018). [CrossRef]
2. M. Dong, J. Lu, S. Yao, et al., “Application of LIBS for direct determination of volatile matter content in coal,” J. Anal. At. Spectrom. 26(11), 2183 (2011).
3. Y. G. Soskind, C. Olson, B. Noharet, et al., “A compact LIBS system for industrial applications,” in Photonic Instrumentation Engineering II (2015).
4. R. Noll, C. Fricke-Begemann, M. Brunk, et al., “Laser-induced breakdown spectroscopy expands into industrial applications,” Spectrochim. Acta, Part B 93, 41–51 (2014). [CrossRef]
5. J. El Haddad, E. S. de Lima Filho, F. Vanier, et al., “Multiphase mineral identification and quantification by laser-induced breakdown spectroscopy,” Miner. Eng. 134, 281–290 (2019). [CrossRef]
6. L. An, X. Zhang, X. Wang, et al., “High-accuracy quantitative analysis of coal by small sample modelling algorithm based Laser Induced Breakdown Spectroscopy,” J. Anal. At. Spectrom. 37(10), 2022–2032 (2022). [CrossRef]
7. A. Metzinger, D. J. Palásti, É. Kovács-Széles, et al., “Qualitative Discrimination Analysis of Coals Based on Their Laser-Induced Breakdown Spectra,” Energy Fuels 30(12), 10306–10313 (2016). [CrossRef]
8. T. Zhang, C. Yan, J. Qi, et al., “Classification and discrimination of coal ash by laser-induced breakdown spectroscopy (LIBS) coupled with advanced chemometric methods,” J. Anal. At. Spectrom. 32(10), 1960–1965 (2017). [CrossRef]
9. S. Palanco and J. Laserna, “Remote sensing instrument for solid samples based on open-path atomic emission spectrometry,” Rev. Sci. Instrum. 75(6), 2068–2074 (2004). [CrossRef]
10. P. Werheit, C. Fricke-Begemann, M. Gesing, et al., “Fast single piece identification with a 3D scanning LIBS for aluminimu cast and wrought alloys recycling,” J. Anal. At. Spectrom. 26(11), 2166–2174 (2011). [CrossRef]
11. P. Werheit, R. Noll, and C. Fricke-Begemann, “Fully automated LIBS sorting system for single particle analysis in metal recycling” (2012).
12. D. Redoglio, E. Golinelli, S. Musazzi, et al., “A large depth of field LIBS measuring system for elemental analysis of moving samples of raw coal,” Spectrochim. Acta, Part B 116, 46–50 (2016). [CrossRef]
13. M. Gaft, I. Sapir-Sofer, H. Modiano, et al., “Laser induced breakdown spectroscopy for bulk minerals online analyses,” Spectrochim. Acta, Part B 62(12), 1496–1503 (2007). [CrossRef]
14. M. Gaft, E. Dvir, H. Modiano, et al., “Laser Induced Breakdown Spectroscopy machine for online ash analyses in coal,” Spectrochim. Acta, Part B 63(10), 1177–1182 (2008). [CrossRef]
15. M. Gaft, L. Nagli, Y. Groisman, et al., “Industrial online raw materials analyzer based on laser-induced breakdown spectroscopy,” Appl. Spectrosc. 68(9), 1004–1015 (2014). [CrossRef]
16. S. Maurice, R. C. Wiens, P. Bernardi, et al., “The SuperCam Instrument Suite on the Mars 2020 Rover: Science Objectives and Mast-Unit Description,” Space Sci. Rev. 217(3), 47 (2021). [CrossRef]
17. R. C. Wiens, S. Maurice, S. H. Robinson, et al., “The SuperCam Instrument Suite on the NASA Mars 2020 Rover: Body Unit and Combined System Tests,” Space Sci. Rev. 217(1), 4 (2021). [CrossRef]
18. R. C. Wiens, S. Maurice, J. Lasue, et al., “Pre-flight calibration and initial data processing for the ChemCam laser-induced breakdown spectroscopy instrument on the Mars Science Laboratory rover,” Spectrochim. Acta, Part B 82, 1–27 (2013). [CrossRef]
19. R. B. Anderson, O. Forni, A. Cousin, et al., “Post-landing major element quantification using SuperCam laser induced breakdown spectroscopy,” Spectrochim. Acta, Part B 188, 106347 (2022). [CrossRef]
20. A. Li, X. Zhang, Y. Yin, et al., “High-accuracy measurement on heat of detonation with good robustness by laser induced breakdown spectroscopy of Energetic Materials,” J. Anal. At. Spectrom. 38(4), 810–817 (2023). [CrossRef]
21. X. Xu, A. Li, X. Wang, et al., “The high-accuracy prediction of carbon content in semi-coke by laser-induced breakdown spectroscopy,” J. Anal. At. Spectrom. 35(5), 984–992 (2020). [CrossRef]
22. N. B. Zorov, A. A. Gorbatenko, T. A. Labutin, et al., “A review of normalization techniques in analytical atomic spectrometry with laser sampling: From single to multivariate correction,” Spectrochim. Acta, Part B 65(8), 642–657 (2010). [CrossRef]
23. A. R. Boyain-Goitia, D. C. Beddows, B. C. Griffiths, et al., “Single-pollen analysis by laser-induced breakdown spectroscopy and Raman microscopy,” Appl. Opt. 42(30), 6119–6132 (2003). [CrossRef]
24. L. Zhang, L. Dong, H. Dou, et al., “Laser-induced breakdown spectroscopy for determination of the organic oxygen content in anthracite coal under atmospheric conditions,” Appl. Spectrosc. 62(4), 458–463 (2008). [CrossRef]
25. S. Y. Oh, F. Y. Yueh, J. P. Singh, et al., “Preliminary evaluation of laser induced breakdown spectroscopy for slurry samples,” Spectrochim. Acta, Part B 64(1), 113–118 (2009). [CrossRef]
26. F. Bredice, H. Sobral, M. Villagran-Muniz, et al., “Real time measurement of the electron density of a laser generated plasma using a RC circuit,” Spectrochim. Acta, Part B 62(8), 836–840 (2007). [CrossRef]
27. S. Zhang, X. Wang, M. He, et al., “Laser-induced plasma temperature,” Spectrochim. Acta, Part B 97, 13–33 (2014). [CrossRef]
28. K. H. Esbensen and P. Geladi, “Principal Component Analysis: Concept, Geometrical Interpretation, Mathematical Background, Algorithms, History, Practice,” Comprehensive Chemometrics (Elsevier, 2009).
29. J. It and Jorge Cadima, “Principal component analysis a review and recent developments,” Phil. Trans. R. Soc. A. 374(2065), 20150202 (2016). [CrossRef]
30. Z. Gniazdowski, “New Interpretation of Principal Components Analysis,” Zeszyty Naukowe WWSI 11, 43–65 (2017).
31. R. Rosipal and N. Krämer, “Overview and Recent Advances in Partial Least Squares,” in Subspace, Latent Structure and Feature Selection (2006), pp. 34–51.
32. I. S. Helland, “Partial Least Squares Regression and Statistical Models,” Scandinavian J. Statistics 17, 97–114 (1990).
33. J. A. Westerhuis, E. J. van Velzen, H. C. Hoefsloot, et al., “Multivariate paired data analysis: multilevel PLSDA versus OPLSDA,” Metabolomics 6(1), 119–128 (2010). [CrossRef]
34. B. Worley and R. Powers, “PCA as a practical indicator of OPLS-DA model reliability,” Curr. Metabolomics 4(2), 97–103 (2016). [CrossRef]
35. A. Li, X. Zhang, X. Liu, et al., “Datasheet of the online measurement results by LIBS,” figshare (2023).https://doi.org/10.6084/m9.figshare.24290395
