Open Access
Add to BibSonomy
Abstract
The laser-induced breakdown spectroscopy (LIBS) experimental platform was applied to obtain LIBS spectral the data of 10 CL60 wheel steel samples. The principle component analysis (PCA) was used to preliminarily analyze the macroscopic characteristics of LIBS spectral data. With the spectral intensity and spectral intensity combined with spectral intensity ratio as variables, three spectral correction methods including median filtering, baseline correction and multiple scattering correction (MSC) were used for pretreatment. And the support vector machine (SVM) qualitative model was established to determine the metallographic structure. It was found that the SVM model established by using the pre-processed data of MSC as the input variable has the best effect. The accuracy rate of calibration set is 100%, and the accuracy rate of prediction set is 98.4%. The research has shown that LIBS combined with SVM model can be used for discriminant analysis of different metallographic structures of train wheel steel.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
With the rapid development of modern high-speed railway, the operation safety and service reliability of trains have gradually become the focus of attention in the research and operation of high-speed railway. However, the successive occurrence of wheel-rail wear, deformation and fatigue damage has seriously affected the safety and reliability of railway [1,2]. The traction operation, braking and deceleration of trains all depend on the traction and braking force provided by the wheel friction pair, so the wheel plays a key role in maintaining the operation safety and service reliability of high-speed trains. Wheel damage forms can be roughly divided into two categories: (1) wheel wear damage, such as wheel tread wear, wheel tread width and wheel polygon wear [3–5]; (2) wheel fatigue damage, such as rolling contact fatigue [6,7], thermal fatigue crack [8,9], wheel abrasion or flat scar caused by tread braking or idling [10–12]. In the existing research, most of the research on wheels has been focused on wheel wear damage, such as the researches on image detection and ultrasonic detection of wheel tread wear [13,14], the research on displacement measurement and detection and vibration acceleration detection of polygon wear of wheels [15,16]. However, there are few researches on wheel fatigue damage detection. The main reason is that the wheel fatigue damage is caused by the changes in the metallographic structure of wheel steel. During the service process of high-speed train wheels, the surface characteristics such as metallographic structure and surface hardness will change at the micro level, and the accumulation of changes in the microstructure of train wheels will lead to the deformation of the wheel at the macro level, such as cracks, wheel out-of-round wear, etc. The metallographic structure detection method of metal materials is usually observed by metallographic microscope [17]. This method is difficult to apply to on-site detection because the method has high requirements on the smoothness of the metal surface and troublesome sample preparation. A fast, accurate and effective metallographic structure detection method that can be applied to the site is needed. Therefore, this paper introduces laser-induced breakdown spectroscopy (LIBS) technology to carry out the study on the detection of different metallographic structure of wheel steel.
LIBS technology is a rapidly developing atomic spectral analysis technology, which excites the sample surface to the plasma state through high-energy pulsed laser to analyze the emission spectrum of the generated elements [18,19]. It has the advantages of real-time online [20], remote sensing [21], no complicated sample preparation process [22,23], and near nondestructive analysis. Dai yuan et al. [24,25] studied the correlation between plasma characteristics with different metallographic structure by laser-induced breakdown spectroscopy. The experimental results showed that the spectral line intensity of different metallographic structures had a certain difference; and the correlation study between grain size grade and laser-induced breakdown spectroscopy was also carried out at the same time. The results show that the spectral line intensity of matrix elements Fe and Cr had a good correlation with the grain size grade of the sample. The ratio of ionic line strength to atomic line strength of Fe element has a certain correlation with the metallographic structure. Xue Bowen et al. [26] carried out a comparative analysis on the relationship between the spectral intensity of the matrix Fe element and the alloying element Mn and the metallographic structure of S45C steel. The results showed that the spectral line intensity of Fe element and Mn element were quite different in different metallographic structures. And the principal component analysis can be used to distinguish different metallographic structures in the wavelength range of 280nm∼320nm. Lu Shengzi et al. [27] used the method of combining LIBS and SVM to establish an SVM model by using the spectral line intensity and spectral line intensity ratio of the metallographic structure, and analyzed 10 different aging Spectral characteristics of grade T91 steel specimens. The results show that the ageing grade of the steel can be evaluated. Zhang Tianlong et al. [28] used the LIBS technology combined with SVM and PLS to quantitatively analyze 20 slag samples, then used PLS-DA to identify and classify open hearth slag and high titanium slag. The results show that this method is an effective method to realize on-line analysis and process control of slag. LIBS can be used as a new method to determine the microstructural changes of steel samples.
At present, there are few literature reports on the detection of different metallographic structures of train wheel steel. In view of the characteristics of LIBS technology such as real-time online and nondestructive analysis, a fast classification model for different metallographic structures of train wheel steels based on LIBS was proposed and applied to the field detection, which is of great significance to the development of fatigue damage detection technology of train wheels.
2. Materials and methods
2.1 Experimental materials
CL60 wheel steel is a rolling wheel steel widely used in trains in China. It has high strength, hardness and elasticity, low plasticity during cold deformation, poor machinability, poor welding and hardenability and other mechanical properties. Its main chemical composition is shown in Table 1. The sample used in this paper is 10 cylinders with a diameter of 20 mm and a thickness of 10 mm.
Table 1. Chemical composition of CL60 wheel steel (mass fraction, %)
In order to obtain wheel steel samples with different metallographic structures, a tubular furnace SK-G08123K was used for heat treatment process as shown in Table 2.
Table 2. Steel heat treatment process of CL60
The wheel steel specimens were taken from the wheel steel sample after the above heat treatment process, and the metallographic structure type was determined. First, the intercepted wheel steel specimens were ground on MP-2DE metallographic specimen grinder with sandpaper from 200 mesh to 2000 mesh for about 400 mesh gradient grinding. Then, the rough polishing was carried out with diamond polishing paste with a diameter of 3.5mm, and the fine polishing was carried out with a diamond polishing paste with a diameter of 1 mm, and cleaned and dried with anhydrous ethanol. Subsequently, 4% nitric acid alcohol solution was used for erosion. According to the change of polishing surface, the erosion degree was judged. After meeting the requirements, the sample was washed with flowing water, and then rinsed with anhydrous ethanol and dried. Finally, a metallographic microscope Leica DMI8 was used to observe the metallographic structure of the sample. The results are shown in Fig. 1.
Fig. 1. Metallographic determination results: (a)Rough sample metallography: flake pearlite + angular carbide; (b) Metallography of air-cooled samples: lamellar pearlite + secondary cementite distributed in a network along grain boundaries; (c) Metallography of tempered sample: tempered sorbsite with martensite orientation maintained; (d) Metallography of air-cooled samples: lamellar pearlite + reticular ferrite
The surface of the wheel steel sample needs to be treated before the LIBS detection experiment. It is placed on the MP-2A metallography sample polishing machine and polished with sandpaper from 200 mesh to 2000 mesh for about 400 mesh gradient grinding. Then it is polished to mirror surface with diamond polishing paste with a diameter of 3.5mm, and cleaned with anhydrous ethanol and dried. In this way, it ensures that the surface of the sample is highly smooth, thus reducing the potential fluctuation caused by the change of laser focal depth, and the influence of the surface oxide layer and external pollutants on the measurement of LIBS in the laser-induced plasma signal can also be reduced. The samples selected in this paper are 10 CL60 wheel steels, and their metallographic categories are shown in Table 3. Ferrite is the solid solution of carbon and alloying elements dissolved in α-Fe. Pearlite is the mechanical mixture of ferrite and carburized body formed by the eutectic reaction in iron-carbon alloy, the morphology of which is a lamellar complex with alternate overlapping thin layers of ferrite and carburized body, also called lamellar pearlite. Secondary cementite is precipitated by austenite and the morphology is distributed along the austenite grain boundaries and a network is formed. Tempered sorbsite is based on the ferrite with uniform carbide particles distributed on the matrix.
Table 3. Metallographic classification of CL60 wheel steel samples
2.2 LIBS equipment and parameters
The equipment used in the experiment is Marine optical MX2500+, and the schematic diagram of LIBS system is shown in Fig. 2. The sample is excited by a Q-switched Nd: YAG laser, which contains high energy with a pulse width of nanosecond order. After being reflected by a 45° plane transmitting mirror, the laser is focused on the sample surface by the lens to peel off trace substances on the sample surface to form plasma. The optical fiber is placed above the sample to collect the plasma spectral signal during the ablation process. The optical fiber is collected and transmitted to the multi-channel spectrometer with five channels, and then the parameters are set by the MaxLIBS software supporting the spectrometer, so as to collect the wavelength and signal intensity of the elements.
The laser energy was set to 50 mJ, the wavelength range of the spectrometer was 198.71∼727.69 nm, the optical resolution was 0.1 nm, and the integration time was set to the minimum width of 1ms of the signal collected by the spectrometer. According to the optimal SNR, 2500 ns of delay time was selected. Considering the inhomogeneity of sample composition and the instability of laser energy when the laser action position changes, a special measurement scheme of 5×5 for the matrix positions (each position is ablated by laser pulse for 200 times) was designed and implemented, and the distance between two adjacent detection points was about 3 mm to reduce the mutual influence. The 5 × 5 for the matrix position means 25 random positions and 200 ablation times for each position. Then 200 sets of spectral data can be obtained. 200 sets of spectral data are processed: the first 30 sets of spectral data are eliminated and the last 170 groups of spectral data are averaged to represent the spectral data at this position.
2.3 Data processing
Due to tunnel effect, that is, the depth and shape of laser ablation crater change with the increase of the number of laser pulse ablation samples, which affects the laser ablation of the sample and the obtained spectral line intensity in the subsequent process [29]. Therefore, in order to reduce the influence of sample surface state on laser ablation, the spectral data obtained from the first 30 laser pulses ablation at all measurement points were removed, and the spectral line intensity used later was the average value of spectral data obtained from the 31st to the 200th laser pulses.
In order to reduce the fluctuation of laser energy, the difference of spectrometer resolution, the difference of the external environment, and and the influence of sample inhomogeneity, three different methods including median filter, baseline correction, and multiplicative scatter correction (MSC) were used to preprocess the input variables of the model before modeling.
Multivariate scattering correction [30] can eliminate the scattering effects caused by uneven particle distribution and particle size, and is widely used in solid diffuse reflection.
- 1) Calculate the average spectrum of the calibration set sample $\overline x$ (ideal spectrum).
- 2) Make linear regression with x and $\overline x$, that is, $x = {b_0} + \overline x b$, then to calculate the b0 and b by the least squares method.
- 3) ${x_{\textrm{MSC}}} = (x - {b_0})/b$
Baseline correction [31] is also known as background removal and background correction. Due to the changes of laser energy, plasma formation conditions, instrument electronic noise and thermal noise, LIBS spectral data often have large fluctuations. This will add more baseline drift and background interference to the collected spectral data, which even mask the information of the material and then greatly affect the accuracy of spectral analysis. The use of baseline correction can effectively solve the above problems. The principle is shown in formula (1):
In the formula, x represents all spectral values, $\min (x)$ represents the minimum value in all spectra, and y is the spectral value after baseline correction.Support vector machine is a relatively advanced and efficient classification model, and its core idea is to construct an optimal classification surface, which can not only classify the sample data of different categories in the high-dimensional data space, but also maximize the interval of sample data of different categories and minimize the structural risk [32].The commonly used kernel functions mainly include linear kernel and radial basis function (RBF). In this experiment, the classification of RBF kernel function test vectors was used. As shown in formula (2), where a represents the sample point; b represents the center point of the kernel function; ${\sigma ^2}$ indicates the variance of the RBF kernel function.
3. Results and discussion
3.1 Analysis of spectral characteristics
The representative wavelength ranges in the characteristic spectrum of sample 1# are 225-300 nm and 350-450 nm, as shown in Fig. 3, which is obtained by averaging the spectral data obtained after 200 pulse strikes at a randomly selected measurement point in a $5 \times 5$ measurement matrix. A large number of characteristic spectral lines of matrix elements (Fe) and alloying elements (Mn and Cr) can be clearly identified from the spectra.
In order to analyze the difference of spectral characteristics of different metallographic structures and samples, the characteristic spectral lines of matrix element Fe, alloying element Mn and Cr were selected in the experiment, and the spectral line intensity was analyzed and compared, and the results are shown in Fig. 4(a). It can be seen from the Fig. 4(a) that the intensity of the characteristic spectral lines of different metallographic samples has obvious differences, and the differences reflected by different spectral lines are not the same. Among them, the spectral line intensity of pearlite + cementite is the strongest, and the spectral line intensity difference of matrix element Fe II 259.8369 nm is obviously stronger than the rest of analytical spectral lines. However, the intensity variability of the alloying element Mn I 403.3068nm and Cr I 427.4806 nm are not significantly different from the intensity of the three types of metallographic structures other than pearlite + cementite. Therefore, the relationship between ions and atomic line intensity ratio and different metallographic structures was explored in this paper, and the following spectral lines are selected for analysis: Fe I 252.2849 nm, 373.7132 nm, FeII 249.3264 nm, Mn I 404.1357 nm, Mn II 294.9205 nm, Cr I 427.4806 nm, and the results were shown in Fig. 4(b). It can be seen from the figure that the ratio of spectral intensity to different metallographic structure has more obvious gap than the single spectral intensity. And the gap between singe spectral intensity to different metallographic structure is non-obvious. This result indicates that the spectral line intensity ratio can be used to distinguish different metallographic structures, but it cannot directly identify the metallographic types.
Fig. 4. (a) Relationship between Metallographic Structure and Spectral Line Strength (b) Relationship between Metallographic Structure and Spectral Line Strength Ratio
As mentioned above, the spectral line intensity cannot be used to directly distinguish the metallographic structure category. At the same time, it is found in the experiment that the linear relationship between the spectral line intensity ratio and the metallographic structure category is not ideal. Moreover, due to the different measurement points, the spectral line intensity will also have certain fluctuations, which has a great impact on the linear relationship. Therefore, Principal component analysis (PCA) dimensionality reduction multivariate analysis method was used to preliminarily analyze the characteristics of LIBS spectral data. Based on the research literature of a large number of scholars and the standard database of National Institute of Standards and Technology (NIST), 53 characteristic spectral lines of matrix elements Fe, alloying elements Mn and Cr were selected for PCA analysis. The selected characteristic spectral lines were listed in Table 4, and the number of principal components was set as 7. Due to the large dimension of selected data variables, in order to better distinguish the differences of the four types of metallographic structures from the figure, PC1, PC2 and PC3 were selected to establish a three-dimensional PCA classification model. The contribution rates of PC1, PC2 and PC3 were 68%, 10.8% and 7.5% respectively, as shown in Fig. 5. The analysis results show that the data points of samples with different types of metallographic structure have certain aggregation in the three-dimensional feature space after PCA dimensionality reduction. Besides, it can be seen that these data have obvious nonlinear separability, and it can be trusted that the data points of these samples can be better divided in a higher dimensional space. Since the PCA reduced-dimensional multivariate analysis method is an unsupervised learning method, there is no clear discriminant boundary between the established data point categories, and the spectral differentiation of a few samples is not straight forward. Therefore, after confirming that LIBS spectral data of different types of metallographic samples have a certain degree of separability, the chemometric analysis method was introduced to try to establish a rapid classification and detection model of different metallographic structures of wheel steel.
Table 4. Selected characteristic spectral lines for PCA analysis
3.2 Establishment of rapid classification model for different metallographic structures
3.2.1 Modeling by spectral line intensity
In order to conduct discriminant analysis on wheel steel samples with different metallographic structures, three different pretreatment methods of median filtering, baseline correction and MSC were used to process the full-band spectral line intensity, and then SVM algorithm was used to establish a qualitative model. The model uses K-S algorithm to divide 250 independent samples into calibration set and prediction set according to 3:1, that is, the calibration set consists of 188 samples, and the prediction set consists of 62 samples. The index results of the model are shown in Table 5. The results show that the combination of LIBS technology and chemometric analysis method can be used for rapid detection and analysis of wheel steel samples with different metallographic structures, and the SVM model established by MSC pretreatment with spectral line intensity as the variable has the best effect. One of the tempered sorbite samples in the calibration set was misjudged as pearlite + ferrite metallographic structure, and one of the pearlite + ferrite samples in the prediction set was misjudged as tempered sorbite microstructure. As shown in Fig. 6, the accuracy of calibration set was 99.5%, and the number of misjudged samples was 1; the accuracy of prediction set was 98.4%, and the number of misjudged samples was 1.
Fig. 6. Prediction results of full-band spectral line intensity using MSC pretreatment model:correction of the correlation between the classification variables and the predicted values;prediction of the correlation between the classification variables and the predicted values
Table 5. The evaluation index results of the established model
3.2.2 Modeling with spectral line intensity ratio
Both the spectral line intensity and the spectral line intensity ratio are related to the samples of different metallographic structures. Therefore, in order to figure out the best qualitative discriminant analysis model for different metallographic structures, spectral line intensity and along with few spectral intensity ratios are used as the input variables of the analysis model for comparative analysis. The characteristic spectral lines which are used to constitute the spectral line intensity ratio here are the characteristic spectral lines selected for PCA analysis in section 2.1. When the spectral lines are in different wavelength ranges of the spectrometer, they are used to construct the spectral intensity ratio, which is affected by the difference of the response characteristics of the spectrometer. Therefore, the spectral lines should be grouped according to the spectrometer channel where their wavelength is located. In this experiment, they are divided into three groups, 200 nm-240 nm, 240 nm-450 nm and 455 nm-660 nm, respectively. And then the spectral lines divided into the same group are paired and their corresponding spectral line intensity ratios are calculated, as shown in Table 6.
Table 6. Characteristic lines used to form spectral line intensity ratios
The spectral line intensity ratio was directly splinted after the variable spectral line intensity, and then the variables combined with spectral line intensity and spectral line intensity ratio were processed by using three different pretreatment methods as described above, and then the qualitative model was established by SVM algorithm. The K-S algorithm is still used to divide the calibration set and prediction set according to 3:1, and the model index results are shown in Table 7. The results show that the SVM model established by using the data after MSC preprocessing combined with the spectral line intensity ratio as a variable has the optimal performance. Among them, only one of the sorbite samples with concentrated tempering was misjudged as pearlite + ferrite microstructure, as shown in Fig. 7. The accuracy of calibration set was 100%, and the number of misjudged samples was 0; the accuracy of prediction set was 98.4%, and the number of misjudged samples was 1.
Fig. 7. Spectral Line Intensity Combined with Spectral Line Intensity Ratio Using MSC Pretreatment to Establish Model Results: (a) Corrected Correlation Graph between Classified Variables and Prediction Values; (b) Predicted Correlation Graph between Classified Variables and Prediction Values.
Table 7. The evaluation index results of the established model
By comparing the results of the two models with the same pretreatment method, it can be seen that the model with the ratio of spectral line intensity and spectral line intensity as the variable has a higher discriminant accuracy for metallographic structure. It is obviously found that the model with the ratio of spectral line intensity and spectral line intensity as the variable is better than the model with the spectral line intensity as the variable. The results show that the use of spectral line intensity combined with the intensity ratio of ion lines to atomic lines and the spectral line intensity ratio of alloying elements to matrix elements as the input variables of the model can significantly improve the SVM model for the solution of the rapid classification of metal material microstructure, and a quantitative analysis model with stronger correlation is established. At the same time, the result also indicates that the spectral line intensity ratio is essentially a relative intensity of the spectral line, which can reduce the influence of spectral signal fluctuation caused by the instability of the laser ablation sample [33].
4. Conclusion
LIBS technology can be used not only for elemental analysis, but also for rapid classification of metallographic structure of metal materials combined with chemometrics. It is found in the experiment that the spectral line intensity cannot distinguish the metallographic structure category directly, and the linear relationship between spectral line intensity ratio and metallographic structure is not ideal. Therefore, PCA reduced-dimensional multivariate analysis method was introduced for the preliminary analysis of the characteristics of LIBS spectral data. It is found that the data points of samples with different types of metallographic structures have certain aggregation in the three-dimensional feature space after PCA dimensionality reduction, and it can be seen that these data have obvious nonlinear separability. Then, the SVM discriminant analysis models were established by using spectral line intensity, spectral line intensity combined with spectral line intensity ratio as variables. Among the models with spectral line intensity as variables, the model established by MSC pretreatment showed the optimal performance. The accuracy of calibration set is 99.5%, and the number of misjudgment is 1; the accuracy of prediction set is 98.4%, and the number of misjudgment is 1. In the model with spectral line intensity combined with spectral line intensity ratio as variables, the model established after MSC pretreatment presented the best effect. The accuracy of calibration set was 100%, and the number of misjudgments was 0; the accuracy of prediction set was 98.4%, and the number of misjudgments was 1. It is found that the model with spectral line intensity combined with spectral line intensity ratio as variable has higher discrimination accuracy, which is significantly better than the model with spectral line intensity as variable. The study shows that it is feasible to quickly distinguish the microstructure of high-speed railway wheels by LIBS combined with SVM model, which can be used for on-site diagnosis and evaluation of wheel wear, thus providing a certain guarantee for maintaining the safe operation of trains.
Funding
Science and Technology Research Project of Jiangxi Education Department (GJJ210632); National Natural Science Foundation of China (2002017018); National 863 Program (SS2012AA101306).
Acknowledgments
The authors thank support provided by National 863 Program (SS2012AA101306), National Natural Science Foundation of China (2002017018) and Science and Technology Research Project of Jiangxi Education Department (GJJ210632).
Disclosures
The authors declare no conflicts of interest.
Data availability
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.
References
1. Z. Y. Shen, “Research on High-speed Railway and High-speed Train,” J. Vibration: Measurement and Diagnosis 1, 1–7 (1998).
2. X. S. Jin and Z. Y. Shen, “Recent progress in the research of wheel/rail rolling contact fatigue,” J. Journal of the China Railway Society 2, 92–108 (2002).
3. J. Nielsen and A. Johansson, “Out-of-round railway wheels-a literature survey,” J. Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit 214(2), 79–91 (2000). [CrossRef]
4. A. Johansson and C. Andersson, “Out-of-round railway wheels: a study of polygonalization through simulation of three-dimensional wheel-rail interaction and wear,” J. Vehicle System Dynamics 43(8), 539–559 (2005). [CrossRef]
5. A. Johansson, “Out-of-round railway wheels-assessment of wheel tread irregularities in train traffic,” Journal of Sound and Vibration 293(3-5), 795–806 (2006). [CrossRef]
6. B. Zhang, Failure Analysis and Damage Atlas of Railway Wheel and Hoop (China Railway Publishing, Beijing: 2002).
7. F. J. Franklin, T. Chung, and A. Kapoor, “Ratcheting and fatigue-led wear in rail-wheel contact,” J. Fatigue & Fracture of Engineering Materials & Structures 26(10), 949–955 (2003). [CrossRef]
8. C. G. He, Y. Z. Chen, Y. B. Huang, Q. Y. Liu, M. H. Zhu, and W. J. Wang, “On the surface scratch and thermal fatigue damage of wheel material under different braking speed conditions,” Eng. Failure Anal. 79, 889–901 (2017). [CrossRef]
9. A. Ekberg and E. Kabo, “Fatigue of railway wheels and rails under rolling contact and thermal loading-an overview,” Wear 258(7-8), 1288–1300 (2005). [CrossRef]
10. J. Jergéus, “Martensite formation and residual stresses around railway wheel flats,” J. Proceedings of the Institution of Mechanical Engineers Part C: Journal of Mechanical Engineering Science 212(1), 69–79 (1998). [CrossRef]
11. J. Jergéus, C. Odenmarck, and R. Lundén, “Full-scale railway wheel flat experiments,” J. Proceedings of the Institution of Mechanical Engineers Part F: Journal of Rail and Rapid Transit 213(1), 1–13 (1999). [CrossRef]
12. Y. Z. Chen, Research on Influence of Abrasion on Material Wear and Rolling Contact Fatigue of Wheel, (Southwest Jiaotong University, D. Chengdu: 2018).
13. P. Z. Li, Y. K. Wang, K. C. Wang, L. Kang, and S. J. Wang, “On-line flaw detection device for train wheelset tread based on electromagnetic ultrasonic,” J. Nondestructive testing 35(7), 70–74 (2013).
14. L. L. Tian, Z. D. Fang, and Y. Zhao, “Location method of peel and scratch in detection of wheel tread damage of railway freight car,” J. Journal of the China railway society 31(5), 31–36 (2009).
15. J. Yang, H. C. Xie, and Z. Y. Xing, “Portable Wheelset Measuring Instrument Based on Laser Displacement Sensor,” J. Instrument Technology and Sensor 4, 41–44 (2017).
16. Z. Wang, J Q Guo, X. R. Gao, and L. Wang, “Research on multi-channel Detection system for Train wheel Tread Defects,” J. Opto-electronic engineering 38(07), 92–98 (2011).
17. C. X. Kuang, “Application of metallographic microscope in steel testing,” J. Guangdong building materials 33(06), 25–26 (2017).
18. M. R. Dong, L. P. Wei, J. D. Lu, W. B. Li, S. Z. Lu, S. S. Li, C. Y. Liu, and J. H. Yoo, “A comparative model combining carbon atomic and molecular emissions based on partial least squares and support vector regression correction for carbon analysis in coal using LIBS,” J. Anal. At. Spectrom. 34(3), 480–488 (2019). [CrossRef]
19. N. Khajehzadeh, O. Haavisto, and L. Koresaar, “On-stream and quantitative mineral identification of tailing slurries using LIBS technique,” Miner. Eng. 98, 101–109 (2016). [CrossRef]
20. D. S. Meng, N. J. Zhao, Y. Y. Wang, M. J. Ma, L. Fang, Y. H. Gu, Y. Jia, and J. G. Liu, “On-line/on-site analysis of heavy metals in water and soils by laser induced breakdown spectroscopy,” Spectrochimica acta part B: atomic spectroscopy 137, 39–45 (2017). [CrossRef]
21. L. X. Sun, H. B. Yu, Z. B. Cong, Y. Xin, Y. Li, and L. F. Qi, “In situ analysis of steel melt by double-pulse laser-induced breakdown spectroscopy with a Cassegrain telescope,” Spectrochimica Acta Part B: Atomic Spectroscopy 112, 40–48 (2015). [CrossRef]
22. C. H. Ma, Y. H. Zhao, and F. W. Meng, “LIBS data acquisition and analysis system for molten steel,” J. Laser Journal 39(4), 30–33 (2018).
23. L. Gómez-Nubla, J. Aramendia, S. F. O. de Vallejuelo, and J. M. Madariaga, “Analytical methodology to elemental quantification of weathered terrestrial analogues to meteorites using a portable Laser-Induced Breakdown Spectroscopy (LIBS) instrument and Partial Least Squares (PLS) as multivariate calibration technique,” J. Microchemical Journal 137, 392–401 (2018). [CrossRef]
24. Y. Dai, X. Dong, W. L. Zhong, M. R. Dong, J. D. Lin, J. Li, S. C. Yao, and J. D. Lu, “Study on correlation between grain size grade of 12Cr1MoV and laser-induced breakdown spectroscopy,” J. Chinese journal of lasers 41(04), 269–274 (2014).
25. Y. Dai, J. Li, W. L. Zhong, X. Dong, W. S. Li, S. C. Yao, and D Lu J, “Plasma properties of 12Cr1MoV with different metallographic structures,” J. Acta optica sinica 34(03), 317–322 (2014).
26. B. W. Xue, M. C. Cui, C. X. Wang, Z. F. Miao, P. Liao, and S. D. Zhao, “Study on the classification of different metallographic structures based on laser technology,” Laser Technol. 42(06), 806–810 (2018).
27. S. Z. Lu, M. R. Dong, J. W. Huang, W. B. Li, J. D. Lu, and J. Li, “Estimation of the aging grade of T91 steel by laser-induced breakdown spectroscopy coupled with support vector machines,” J. Spectrochimica Acta Part B: Atomic Spectroscopy 140, 1–4 (2018). [CrossRef]
28. T. L. Zhang, S. Wu, D. Juan, J. Wei, K. Wang, H. S. Tang, X. F. Yang, and H. Li, “Quantitative and classification analysis of slag samples by laser induced breakdown spectroscopy (LIBS) coupled with support vector machine (SVM) and partial least square (PLS) methods,” J. Journal of Analytical Atomic Spectrometry 30(2), 1 (2015). [CrossRef]
29. D. N. Patel, R. P. Singh, and R. K. Thareja, “Craters and nanostructures with laser ablation of metal/metal alloy in air and liquid,” Appl. Surf. Sci. 288, 550–557 (2014). [CrossRef]
30. X. L. Chu, Molecular Spectroscopy Analytical Technology Combined with Chemometrics and its Applications, (Chemical Industry Press, 2011).
31. R. X. Xin, Plasma emission Spectrum Analysis (Chemical Industry Press, Beijing: 2011).
32. T. J. Li, J. J. Liu, G. F. Shao, and L. L. Fang, “A novel THz spectroscopy recognition method for transgenic organisms based on APSO combined with SVM,” Opt. Spectrosc. 120(4), 660–665 (2016). [CrossRef]
33. J. P. Castro and E. R. Pereira-Filho, “Twelve different types of data normalization for the proposition of classification, univariate and multivariate regression models for the direct analyses of alloys by Laser-induced breakdown spectroscopy (LIBS),” Journal of Analytical Atomic Spectrometry 31(10), 2005–2014 (2016). [CrossRef]
