Rapid Detection of Cement Raw Meal Composition Based on Near Infrared Spectroscopy

HUANG Bing, WANG Xiaohong, JIANG Ping＊, QIAO Jia

(1. School of Chemistry and Chemical Engineering, University of Jinan ,Jinan 250022,China; 2.School of Electrical Engineering, University of Jinan, Jinan 250022, China)

Abstract: The composition of cement raw materials was detected by near-infrared spectroscopy. It was found that the BiPLS-SiPLS method selected the NIR spectral band of cement raw materials, and the partial least squares regression algorithm was adopted to establish a quantitative correction model of cement raw materials with good prediction effect. The root-mean-square errors of SiO2, Al2O3, Fe2O3 and CaO calibration were 0.142, 0.072, 0.034 and 0.188 correspondingly. The results show that the NIR spectroscopy method can detect the composition of cement raw meal rapidly and accurately, which provides a new perspective for the composition detection of cement raw meal.

Key words: near infrared spectroscopy; cement raw meal; band selection; detection model

1 Introduction

Cement is one of the most important building materials, and the quality of which determines the strength of buildings. The cement industry, as a typical process industry, is greatly influenced by the cement raw meal.Thus, it is particularly important to detect the quality of cement raw meal preparation and its composition,which will have impact on the following production links and production guidance. Most cement plants in China use X-ray fluorescence detection and chemical analysis methods to detect the composition of cement raw meal[1-3]. Analysis process requires complicated sample preparation and long detection time, and can not guide the production in time. A few cement plants have introduced online analyzers of cement raw meal elements. They detect cement raw meal components online and reduces labor intensity by automatic proportioning of cement raw meal. However, high maintenance cost and radioactive sources are harmful to the environment and field workers.

As a quantitative analysis method, near infrared spectroscopy has the advantages of fast, safe, simultaneous analysis of multiple components, good repeatability, and no need for sample processing[4,5]. It has been widely used in industry, agriculture and other industries[6-8].Some scholars have studied the detection of near infrared of the inorganic compounds. Zhanget al[9]found PLS regression coefficients were used to obtain the effective wavelengths (EW) of the four phosphates as the input of LS-SVM to create a prediction model for detection of these phosphates and their phosphorus contents in the soil samples. Limet al[10]constructed an artificial neural network (ANN) classification model for the prediction of soil water content. There is little research on the composition analysis of cement raw meal components. Xiaoet al[11]found a rapid and accurate method to determine CaCO3, SiO2, Fe2O3, and Al2O3in cement raw meal using near infrared spectroscopy. Multiplicative scatter correction (MSC) was employed to eliminate the scattering signal and partial least squares regression was used to build the analysis model. But this method need many variables. Yanget al[12]found GA-biPLS can select less variables with better prediction performance by comparison with PLS and BiPLS. The NIR spectroscopy combined with the GA-biPLS algorithm is a fast, accurate and reliable alternative method for determination of oxides content in cement raw meal.

When modeling near infrared spectra, wavelength selection can effectively improve the quality of the model. When the model contains strongly correlated wavelengths, the prediction effect is generally not affected. Therefore, when there is a strong correlation for a continuous wavelength region in the spectrum,the band screening result is better. By contrast, when the strongly correlated wavelengths are discontinuous,the results of wavelength screening have advantages.Methods for band selection have been reported previously .

2 Experimental

2.1 Sample

Samples were collected from the actual production site for half a month, and a total of 96 samples of cement raw materials used were from actual production of Shandong Qufu Zhonglian Cement Co., Ltd. All the samples were individually packaged and labeled, and stored in the laboratory for one day at 28 ℃ with air humidity 30%.

2.2 Acquisition of spectral data

In this study, a MB3600 Fourier Transform near infrared spectrometer (ABB Co., Ltd Beijing) was used. In order to make the initial acquisition environment the same, the spectrometer was allowed to preheat for one hour, and polytetrafluoroethylene (PTFE) was taken as the background reference. Diffuse reflectance was used to collect spectra, and all the samples were scanned from 10 000-4 000 cm-1, which generated 3 113 variables. The spectra were obtained by 32 successive scans and a resolution of 4 cm-1. Each sample was measured three times, and the final spectrum corresponds to the average. During the process of spectral acquisition, the temperature and humidity in the laboratory were maintained at approximately 28 ℃ and 40%.

2.3 Sample division

Table 1 Test set cement raw meal sample composition

The samples were divided into a training set and a test set by the SPXY method, in which the number of training set samples was 80 and the number of test set samples was 16. The composition distribution of cement raw meal tested by the XRF test method of the sample is shown in Table 1. The composition range of the test set is included in the concentration range of the training set, which meets the modeling standard.

Fig.1 Near infrared spectrum of cement raw meal

The original near infrared spectrum of 83 cement raw meal samples are shown in Fig.1, in the spectral range of 10 000-4 000 cm-1, and the frequency doubling and combining peaks of SiO2, Al2O3, Fe2O3, and CaO are included.

3 Results and discussion

3.1 Selection of characteristic wave quantity

Due to the complex spectral information and low absorption intensity of the near infrared spectrum of cement raw meal in the wavelength band of 1 000-4 000 cm-1, it is necessary to select wavelength bands of the spectrum, eliminate redundant information, reduce the dimensionality of modeling data, and select the characteristic wavelength band related to the main components of cement raw meal to model so as to improve the accuracy of the model.

The near infrared spectra (3 113 wavelength points) of cement raw meal were divided intoksub-regions (k=60, 50, 40, 30, 20) by the BiPLS method, and the band screening results are shown in Table 2.

It can be seen from Table 2 that when the number of intervals is 30, the prediction of SiO2is the best. The selected intervals are the 2nd, 4th, 8th, 12th,19th, 20th, 24th, 27th and 30th intervals.RP2=0.89, and RMSEP=0.14. For the prediction of Al2O3, the model is optimal when the number of divided intervals is 50 and the 1st, 2nd, 6th, 15th, 18th, 24th, 27th, 29th, 44th,47th and 49th intervals are selected.RP2=0.71,RMSEP=0.11. For the prediction of Fe2O3, the model is optimal when the number of divided intervals is 30 and the 2nd, 14th, 23rd, 25th, 27th and 28th intervals are selected.RP2=0.61, and RMSEP=0.03. For the prediction of CaO, the model is optimal when the number of divided intervals is 40 and the 2nd, 5th, 9th, 22nd,35th and 36th intervals are selected.RP2=0.79, RMSEP=0.20. Band selection results are shown in Fig.2.

Table 2 BiPLS subinterval preferred partial result table

Table 3 SiPLS subinterval preferred partial result table

The SiPLS band selection method is used to select the selected bands for the second time. The selected bands are divided into 50 and 40 bands respectively,and then 4 and 3 bands are respectively selected for modeling.

As shown in Table 3, the detection model established by dividing the spectral bands of SiO2, Al2O3,Fe2O3, and CaO into 50 intervals and selecting 4 intervals is the best.

From Tables 2 and 3, it can be seen that after selecting BiPLS and SiPLS bands, theRP2of Al2O3,Fe2O3and CaO all increased, the RMSEP decreased, and only theRP2of SiO2decreased, but the value decreased less,and the RMSEP value increased slightly.

In order to make comparison, the near infrared spectrum of cement raw meal was selected by using BIPLS-UVE band selection method. The near infrared spectra of SiO2, Al2O3, Fe2O3and CaO were selected.

3.2 Modeling results

PLS linear regression models are respectively established for the original spectra, BiPLS, BiPLS-SiPLS,and BiPLS-UVE for each selected band variable. The results are shown in Table 5. It can be seen that using the original spectra to establish the model, its prediction is obviously lower than that of various band selection methods. Therefore, when using near infrared spectroscopy to detect the composition of cement raw meal, band selection should be carried out, and then a calibration model should be established.

Comparing the three band selection methods, it can be seen that the prediction effect of the detection model established by BiPLS-SiPLS band selection for Al2O3and CaO is better than the other two methods,and the band selected by this method is much smaller than the other two methods. The number of wave bands selected for the establishment of Al2O3is 56,RP2=0.88,and RMSEP=0.07 of the model. The number of bands selected for establishing CaO is 32,RP2=0.81 and RMSEP=0.19 of the model.

Fig.2 Band selection results(mscauto is used in the calculationons)

For SiO2, and Fe2O3, the prediction using BiPLS for band selection is better than the other two methods.However, the number of bands selected by this method is 930 and 624 respectively, while the number of bands selected by SiPLS method is only 76 and 52, and the prediction is not much different from the model established by BiPLS method.

According to the above results, the BiPLS-SiPLS band selection method was finally selected to model the near infrared spectra of cement raw meal components.As shown in Fig.3, the actual laboratory test value is compared with the predicted value of the model established by this method. The results of near infrared spectroscopy meet the requirements of on-site detection.

Fig.3 Comparison of the NIR prediction values with the reference values of validation set

Table 4 Band selection result

Table 5 Modeling results

4 Conclusions

The quantitative models of cement raw meal composition were developed by calibration of XRF reference value with near infrared spectra. The study found that the prediction after band selection is obviously better than that of the original spectra model. The number of bands in the near infrared spectra model of cement raw meal established by BiPLS-SiPLS is significantly less than that of BiPLS method and the BiPLS-UVEPLS method. The above results indicate that Bi-PLS-SiPLS shows better prediction effect , which can meet the requirements of cement raw meal production detection and provide technical and theoretical support for cement raw meal composition detection.