Near-infrared (NIR) on-line analysis for coarse-grained raw materials

The near-infrared (NIR) region of the electromagnetic spectrum extends from 780 nm to 2500 nm (wavelength) or 12 800 cm^-1 to 4000 cm^-1 if measuring in wave numbers (the number of waves per cm). NIR spectroscopy is concerned with absorptions of NIR energy by molecules within a sample. ­Absorptions are caused by different mechanisms: fundamental vibrations, overtones of fundamental vibrations, combinations of fundamental vibrations, as well as electronic absorptions [1-3].
Overtones are approximate multiples of the fundamental vibrations. The vast majority of fundamental vibrations occur in the mid-infrared region (4000-400 cm^-1). A combination band occurs when a photon of NIR energy is shared between two (or more) vibrations which would be individually observed as fundamentals in the mid-infrared region. Electronic absorptions are caused by the movement of electrons from one orbit to a higher-energy orbit [4]. Furthermore hydrogen bonding occurs. Whenever water is present in a sample there will be a complex, dynamic interaction between water and the sample, which may be unique for that sample [4]. The shape and strength of water absorption bands depends on the mineral composition of the investigated raw material [5,6].

The NIR absorption spectrum of a raw material is a property of its mineral content. For quantitative analysis a chemometric model maps the infrared spectrum to the chemical composition. Such a model is derived by reference to accurate elemental analysis, accounting for moisture content. The feature of an infrared absorption spectrum depends not only on the mineral content but also on the particle size of the analyzed raw material especially if the particle size distribution is not well constrained [6].

For the first NIR analyzer at the Raysut cement plant in Oman (Fig. 1) the question arose as to whether the particle size effect can be compensated to permit the use of chemically representative fine-grained materials for calibration instead of taking samples off the conveyor belt. Such chemically representative fine-grained material were the drill hole samples from the quarry. This paper describes the steps and the methodology to clarify that subject.

The investigations were performed on 19 limestone samples out of a crusher. The samples had been collected by the end user and their chemical composition covers the expected concentration range of the four oxides of key interest CaO, SiO₂, Al₂O₃ and Fe₂O₃. A part of the samples contained a marl component. Table 1 shows the chemical composition of the investigated samples. The samples were analysed by energy-dispersive x-ray fluorescence spectrometry (ED-XRF) on fused beads.

The original raw material was obtained directly from the conveyor belt. After analyzing the original, coarse-grained raw material with NIR-spectroscopy the samples were dried and crushed. The gap width of the crusher was set so that 90 % of the batch possesses an equivalent particle diameter smaller than 2 mm. The fine-grained samples were investigated with NIR-spectroscopy again. Figure 2 illustrates the coarse-grained original raw material and the same sample after crushing as fine-grained material. Obviously the coarse-grained raw material inherently contains some fine-grained material. Figures 3 and 4 show the particle size distributions of the coarse- and the fine-grained material for the two samples from Figure 2. All the other samples had similar particle size distributions.

3 Acquisition of NIR spectra

The NIR-spectra were acquired using an ABB Spectraflow system. Since the incorporation of moisture into the calibration is crucial, the acquisition of the NIR-spectra was performed for each sample at varying moisture levels. The methodology of determining variable moisture contents in the samples is based on ISO 11465 [7]. First the moisture of the original samples was determined. Therefore the samples were weighed and dried during 24 hours at a temperature of 105 °C in a laboratory-type drying cabinet. The dry net weight was determined immediately after having taken the samples out of the drying cabinet.

Then the samples were allowed to cool down. Within this cooling period a sample starts to capture moisture from the ambient air. The weight of the sample was determined again just before the acquisition of the first spectra. For each moisture level a defined amount of water was sprayed across the sample and then the sample was weighed (Fig. 5). The number of moisture levels has been evaluated experimentally in advance. After having added the water, a total of five NIR-spectra were acquired. Before each scan the sample was mixed in order to take into account the variable position of the coarser or finer particles and the possible heterogeneity of the sample containing more or less moisture. Each NIR-spectrum was recorded by placing the sample on a turntable to approximate the situation on the conveyor belt (Fig. 6). After having recorded five NIR-spectra, the sample’s weight was determined again in order to take into account possible evaporation effects during scanning.

For every moisture level the corresponding moisture content of the material was determined. On the basis of the moisture content in g/100g the effective percentage of every chemical component present in the samples can be calculated by:

CaO_eff  = CaO_dry · (1 - moisture/100)

4 Results

The method of partial least squares (PLS) [8] was used to create the multivariate calibrations. For each sample the training spectra were acquired at 8 different moisture levels, covering a range from approximately 0 % up to a maximum of about 9 % moisture content. The five distinct spectra per moisture level were averaged arithmetically. Hence each sample is represented by eight averaged spectra. Thus, the training set contains a total number of 152 samples.

The coarse-grained material was also measured at varying moisture levels with a moisture content ranging between 0 % and a maximum of 6.2 %. This was the maximum moisture level, which the coarse-grained material could absorb. The coarse-grained material samples were not included in the calibration training set. For the specific fine-grained material based multivariate calibrations only those spectral bands were chosen, of which we knew from past experience that they significantly contribute to the calibration instead of the complete spectrum. Cross validation was used leaving out a block of eight samples.

Once the model for the specific constituent was determined, the selected model was applied to the NIR spectra of the coarse-grained raw material and the predictions were compared to the corresponding reference values. The scatter plots of Figs. 7–11 summarize prediction results when each multivariate calibration was applied to the coarse-grained samples. The abscissa corresponds to the ‘true’ content of CaO, SiO₂, etc. while the ordinate corresponds to the predicted values. The black line corresponds to a least squares fit to the prediction data and the pink line indicates the location of an error-free prediction.

These results show that the percentage oxide weight predictions of CaO, SiO₂, Al₂O₃ and Fe₂O₃ are primarily affected by offset effects when the particle size range of the samples is substantially altered. In the case of moisture, the offset is small but the slope is higher than unity. The fine-grained material has a larger overall surface than coarse-grained material. Therefore the fine-grained material can absorb more moisture than coarse-grained material. So the moisture calibration used to predict the moisture in the coarse-grained material tends to predict higher than the actual values.

The above results were obtained for a single set of multivariate calibrations. The results of correlation methods can not always be generalized. In order to ensure that the above results are not fortuitous, it was decided to verify the above results by a multitude of multivariate calibrations per constituent.

For this purpose we employed a simple combinatorial strategy. Different spectral bands were combined with a variety of preprocessing techniques that are commonly applied in NIR-spectroscopy and which are included in standard commercially available calibration software.

Within the present investigation, three spectral intervals were used. Their limits varied as indicated in Table 2. Thus, a total of 23 716 different spectral regions each consisting of three non-overlapping spectral bands were tested. These spectral regions were combined with the following signal preprocessing techniques: Standard Normal Variate transformation [9], with or without de-trending, with either no derivative or a ­Savitzky-Golay numerical approximation [10] of the first or second derivative using 17, 23, 31 or 35 smoothing points plus mean centering or variance scaling.

All these methods are included in commercial calibration software. Our reference was Grams. Combining all variants resulted in a total of 23 716 · 3 · (1+2 · 4) · 2 = 1 280 664 different versions that were tested in the course of the investigation.

The PLS-1 algorithm, the different preprocessing algorithms combined with the possibility of varying the spectral bands, and modules for performing cross-validation and predicting new unknown spectra were implemented in Matlab. During its implementation caution was exercised to maintain numerical compatibility with the commercial software package PLSplusIQ
(Grams/AI, Thermo Scientific). A Matlab computational result was usually equal to 6- to 7-digits with the corresponding Grams computational result.

Each multivariate calibration, which was done with the NIR spectra of the fine-grained material samples, used the NIR spectra of the coarse-grained material to predict the concentration of the constituents of interest. These predictions were compared with the reference concentrations of these samples. The linear trend of the predictions, i.e. the slope and the offset value of their linear least squares fit was determined for each of the 1 280 664 cases. The subsequent histograms show the slopes of the linear fits and the resulting offset of the linear fit of the 4 key oxides, CaO, SiO₂, Al₂O₃ and Fe₂O₃.

The histograms (Figs. 12–15) show that for all four oxides the great majority of the slopes lies within a band of +/-0.1 around the value 1. This confirms the initial result that the effect of using a calibration based on fine-grained material on material as it really comes out of the crusher is primarily an offset.

In the case of moisture content, the slope histogram shows slope values generally greater than unity (Fig. 16). All constituents (i.e. CaO, SiO₂, Al₂O₃, Fe₂O₃ and moisture) showed that signal pre-processing is the key determining factor for the offset variations.

The number of tests is large enough to permit the following general conclusions:

1) When using multivariate calibrations based on fine-grained material to predict the concentration of coarse-grained cement raw material as it comes out of the crusher the correlation between reference values and predicted values is preserved to a very high degree.

2) The main effect is an offset that depends on the distinct preprocessing technique and the chosen spectral bands.

3) The slopes associated with the predictions of the oxide components in the coarse-grained material show comparatively small deviations from their target value of 1.

4) Moisture calibrations done with fine-grained material show that for increasing moisture contents the predictions of moisture in coarse-grained material tend to systematically overpredict as the moisture level increases.

The practical consequence is that for an NIR On Line Analyzer it is possible to use fine-grained material samples to do the base calibration and the final adjustment to the actual raw material is limited to an offset correction for the oxides and a slope offset correction for moisture. A limited number of on site tests are sufficient to do this final model adaptation. Fine-grained samples are very often obtained from boreholes during a quarry campaign and are then already available at the cement plant. This type of approach was used for the SpectraFlow analyzer of the Raysut Cement company in Oman and the authors would like to express their gratitude for the provision of the various samples.