Open Access
Add to BibSonomy
Abstract
As a famous spectroscopy method for substance detection and classification, laser-induced breakdown spectroscopy (LIBS) is not a nondestructive detection method. Considering the precious samples and the experimental environment, sometimes it is difficult to get enough spectra to build the classification model, which is important for qualitative analysis. In this paper, a spectral generation method for extending the spectral database of LIBS is proposed based on generative adversarial nets (GAN). After enough interactive training, the generated spectra looked very similar to the experimental spectra. Evaluated with unsupervised clustering methods PCA and K-means, the generated spectra could not be distinguished from the real spectra. For each type of sample, most of the simulated spectra and experimental spectra were clustered into the same class, which meant the proposed method was effective to extend the spectral database. Using the spectral database extended by this method as training set data to build the SVM model, the results showed that when there were only a few experimental spectra, the combination of the generated spectra and the experimental spectra for building the classification model could achieve better identification results.
© 2019 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Laser-induced breakdown spectroscopy (LIBS) is a new spectral detection methods has been used in classification and identification of elements [1] As a substance detection method, LIBS has been used in many science and industry fields, such as identification of explosives [2,3], clinical samples [4,5] and cultural relics [6,7] and application in alloy processing [8,9], space exploration [10,11], agriculture and food analysis [12,13]. In the above fields, LIBS has shown advantages in fast analysis and accurate identification.
However, in practical applications, we cannot get enough spectra to build the identification model in some special area. For instance, LIBS is an elemental analytical method which use a pulsed laser to ablate the sample and generate induced plasma emission spectra for analysis [14,15]. Although it can be seen as a slight destructive detection method, it still destroys the sample slightly with every laser shot. For some precious samples, it is very difficult to collect samples and the cost of consumption are very expensive. In some special area such as detecting the toxic and harmful samples, based on the safety consideration, we cannot detect the samples for a long time to get enough spectra for building the model neither. It is well-known that the classification or identification based on spectral analysis is a major application of qualitative analysis. And for most cases, to achieved the classification purposes, we need a spectral database to build the classification or identification model. Based on the above inconvenience of building LIBS database, it is an urgent demand to develop a way extending the spectral database with a small amount of existing spectra.
Commonly, the spectral database can be achieved from experiments. So, there’s few researches in extending the spectral database in special fields. S.K. Anubham et al proposed an approach to reduce the sample consumption for LIBS based identification of explosive materials [16]. They developed a method to construct synthetic spectra which simulated the real spectra. The synthetic spectra are performed by generating random numbers for each wavelength using mean and standard deviation(s) obtained from the experimental spectra, and then connecting the generated random numbers at all wavelengths in a single trial and smoothening by multiscale principal component analysis. However, the simulated spectra are still not very similar with experimental spectra, and especially for some samples like RDX.
In this paper, we proposed a method to extend spectral databases of LIBS based on the theory of generative adversarial nets (GAN). GAN was a generative model first proposed by Ian J. Goodfellow in 2014 for fake image constructing in computer science [17]. In which a generative model G and a discriminative model D are trained simultaneously. The G model generates fake data and the D model discriminates them and feedbacks the information about difference between real data and fake data. So that the G model can generates fake data which are closer to real ones.
In this way, we got simulated spectra and identified them with real spectra measured from experiments with unsupervised methods PCA and K-means. Furthermore, the generated spectra were used to build the SVM classification model with experimental spectra together. Meanwhile, another SVM model were built by only little experimental spectra. The correct classification rate (CCR) results were used to compare the identification ability of these two models and evaluated the proposed method.
2. Experiment
2.1. Experimental setup
The LIBS experimental setup used in this study has been reported in detail in our previous works [18], so only a brief description is given here. The schematics of our built setup was illustrated in Fig. 1. A flash-pumped Q-switched Nd: YAG laser (λ = 1064 nm, pulse frequency 1 Hz, pulse duration τ = 5 ns, beam diameter ∅6 mm, energy 50 mJ/pulse) was used to excite the sample’s surface. The laser beam was reflected by three mirrors and finally focused on the sample surface by a convex lens with a focal length of 100 mm. The plasma emission was focused into a fiber (∅ 600 μm diameter) through a convex lens with a focal length of 36 mm and transferred into a two-channel spectrometer (AvaSpec 2048-2-USB2, Avantes), which covered a range of 190 nm to 1100 nm with a resolution of 0.2~0.3 nm. The spectral acquisition delay time was set to 1.28 μs to reduce the interference from plasma continuum emission. The integration time gate width of the CCD detector was fixed at 2 ms.
2.2. Sample preparation
In this work, five kinds of hazardous materials were used as samples to collect LIBS spectra, including three kinds of explosive materials (RDX, HMX and CL-20, in powder form) and two kinds of chemical poisoning agents (DMMP and TEP, in liquid form).
The explosive materials were pasted on the slides with double-sided tape and then smeared evenly. The chemical poisons were smeared on the aluminum plate formed 25 × 50 mm2 thin layers and dried naturally. In order to only detect the samples’ spectra, the focused laser energy was optimized to avoid ablating the substrates. Three-dimensional motorized stage was used to adjust the focus position of the laser on the samples. And for every hazardous sample, 100 spectra were collected, each on a fresh position.
3. Results and discussion
3.1. Spectra simulated based on generative adversarial nets
Generative adversarial nets (GAN) is a kind of generative model proposed by Ian J. Goodfellow in 2014 to build fake images in computer science fields [17]. Since 2017, GAN has become a very famous methods in deep learning field because of its ability to generate high quality images artificially and been tried in computer image processing field widely [19–22]. Basically, the GAN model contains two artificial neural networks (ANN) models, one works as a generator and the other works as a classifier and feedback. The ANN is a black-box modelling tool and a true multi-input multi-output system which has been used widely in the spectral analysis field [23]. So, we didn’t explain too much information of ANN details, our method focused on using the common ANN to achieved the GAN purpose and the detail process of the proposed method was described below.
In our case, two ANN models were trained simultaneously. One was known as the generative model G and the other was the discriminative model D. The basic unit of computation in an ANN model is the neuron, often called a node or unit. It receives input from some other nodes, or from an external source and computes an output. Each input has an associated weight, which is assigned on the basis of its relative importance to other inputs. Each node applies a function f to the weighted sum of its inputs. The function f is non-linear and is called the Activation Function. The purpose of the activation function is to introduce non-linearity into the output of a neuron. And the Sigmoid function was used as activation function in this paper. An ANN model often consists of three layers which represent three types of nodes, input layer, hidden layer and output layer. For the D model, the input layer nodes number was set to 4092 as the dimension of LIBS spectra, respectively. The hidden layer nodes number was set to 200. The output layer nodes number was set to 1, so that each sample spectra got an output class. Meanwhile, for the G model, the input layer nodes number was set to 200 (same as the hidden layer of D) to get the shared parameters from D. (For the ANN model, the shared parameters mentioned here are hidden ones can be used, the operator doesn’t need to know the exactly value of these parameters.) Without hidden layer, the output layer nodes number of G was set as the input nodes number of D, so that it could generate spectra of same dimension with experimental spectra.
The flowchart of this proposed method was shown in Fig. 2. According to the 100 real spectra for each type of sample as training data input, the G model first generated 100 random number sequences of the same dimension as test data output. The D model was used to estimate a spectrum whether came from G model or experimental spectra. In the discriminative model D, the labels of real spectra were different with generated spectra. Furthermore, the parameters of D were shared with G and the G model was trained with both real spectra and generated spectra. In generative model G, the labels of real spectra and generated spectra were same. Then based on the shared parameters, the new simulated spectra were generated by G model as G output.
This process can be considered as a two-player game and they learn from each other. Or more precisely, a teacher and a student. The student hands in a homework and the teacher gives him a standard homework template and tell him his deficiency. He starts to imitate the template and hands a new homework to the teacher. This process is repeated over and over again until he can hand a homework just like the standard template.
Taking CL-20 as an example, 100 artificial spectra were generated based on 100 experimental spectra. The average generated and the experimental LIBS spectra of CL-20 with different iterative training times were illustrated in Fig. 3. Five kinds of repeated interactive training times were chosen during the generated process, including 1, 50000, 100000, 150000 and 180000 times. As Fig. 3 showed, when the simulated spectra were first generated, they were just random noises. With the interactive training times increasing, the generated spectra were getting closer to the experimental spectra. When the repeated times were set as 180000, the generated spectra looked very similar as the experimental spectra.
Fig. 3 The generated spectra after different interactive training times ((a), (b), (c), (d) and (e) are after 1, 50000, 100000, 150000 and 180000 times respectively) and the experimental spectrum.
For RDX, HMX, DMMP and TEP, the average generated spectra of 180000 repeated times and their corresponding real spectra were shown in Fig. 4. During our experiment, for all 5 kinds of hazardous samples, when the interactive training times were over 100000, the generated LIBS spectra started to look similar as real spectra. When the repeated times were set as 180000, the simulated spectra were very similar as the experimental spectra. The “similar” is just a visual description. In our experiments, we used a K-means algorithm to monitor the generated results after every interactive training. The K-means algorithm classified all the spectra into two clusters. The cluster with more than 50% percent of the experimental spectra was known as the experimental spectra cluster. The percentage of generated spectra which were classified into the same cluster with experimental spectra could be used as a quantitative factor to evaluate the results. The detail information about this factor of all 5 kinds of samples will be discussed in Section 3.2 Unsupervised verification of real spectra and generated spectra.
Fig. 4 Generated LIBS spectra after 180000 times interactive training ((a) RDX, (c) HMX, (e) DMMP, (g) TEP) and corresponding experimental spectra ((b) RDX, (d) HMX, (f) DMMP, (h) TEP).
Through the mentioned process, the generated spectra looked similar to the real spectra. However, there are still some differences according to the statistical property. For the spectra, two main parts of statistical property, average properties and fluctuations are often discussed [24]. For the experimental spectra, close to theoretical values, the average expectation values such as intensities of each spectral line are according to the sample composition and the function of atomic energy level radiation and the excitation energy. But there were also fluctuations because the laser energy and ambient environment were unstable. Although the generated spectrum looked very similar to the experimental spectrum, it did not represent the physical meaning behind it. The first generated data was a sequence of random numbers obeyed Gaussian distribution as mentioned before. Later, the generated spectra only simulated the characteristics of the experimental spectrum at each corresponding line. According to the times of interaction training, the similarity between generated spectra experimental spectra was different. The average sum spectral intensities for each kind of sample were illustrated in Table 1. It showed that for most kinds of samples, the average generated spectra were similar as experimental spectra. However, for some samples like CL-20, it was still a bit different. Meanwhile, there was a small fluctuation range for both two kinds of spectra. In order to compare the fluctuation between generated spectra and experimental spectra, the Standard Deviations of LIBS spectra of all 5 kinds of samples were calculated. The Standard Deviations of generated spectra and experimental spectra were listed in Table 1. Although generated and experimental spectra looked similar for each kind of sample, the values of variance for generated spectra were a little smaller, which showed that the generated spectra were more stable.
Table 1. Average variances and sum intensities of generated and experimental spectra for each kind of hazardous sample
3.2. Unsupervised verification of real spectra and generated spectra
In order to evaluated the similarity of LIBS spectra generated based on GAN method with experimental spectra, Principal Component Analysis (PCA) and K-means algorithms were used in this paper. PCA and K-means are two unsupervised clustering algorithms, which have been used in classification and identification field of spectra widely [25–28]. By projecting down into a less dimensional subspace through the linear transformation, PCA allows the selection of several principal components of the data representing the information for analysis and clustering [29]. For k classes of samples, K-means clustering analysis method finds k clustering centers by an initial cluster centers selection method and assign samples to the nearest center. By minimize the sum of the squared error of groups, center points are reselected and the final clustering is completed [30].
The scores of first two PCs of 100 generated spectra and 100 experimental spectra were illustrated in Fig. 5. As the Fig. 5 showed, the generated LIBS spectra could not be distinguished from the real spectra. Even the generated spectra of RDX, which was mentioned by S.K. Anubham et al that could not be simulated well before [16], were very similar as experimental spectra based on this method. It could also be seen that experimental spectra have a wider dispersion range than generated spectra in Fig. 5. The generated spectra were only widely distributed in the PC1 but concentrated in the PC2. However, it could be seen that the two kinds spectra cannot be distinguished under PC1 and PC2. For PC1, the distribution range of generated spectra and experimental spectra were similar and most of these values of spectra were between −1 and 1. For PC2, although the distribution range of generated spectra was much smaller than that of experimental spectra, it is completely included in the range of experimental spectra. So, the generated spectra could not be distinguished from experimental spectra. The fluctuation range of the experimental spectrum was large, so the ranges of variations of PC1 and PC2 were also large. The fluctuation range of the generated spectrum was small, so the ranges of variations of PC1 and PC2 were also small. However, the ranges of variations of the two spectra were overlapping, so they could not be distinguished by PCA. In order to make a further verification, we also added the PC3 for the analysis, but these two kinds of spectra were still cannot be distinguished. As shown in Fig. 5, the generated spectra distributed not as widely as experimental, but in the distribution range of the experimental spectra.
Fig. 5 PCA clustering results for LIBS spectra of the five kinds of samples after training 180000 times.
The K-means results of all 5 kinds of hazardous materials were showed in Fig. 6. In K-means analysis, if the generated spectra and experimental spectra could not be well divided into two categories, it indicated the spectra generated by the method showed good results. As Fig. 6 illustrated, most of the generated spectra were clustered into the same class with the experimental spectra for Cl-20 and TEP. This coincided with Fig. 5 which showed that for these two kinds of samples, the dispersion ranges were small and spectra were clustered tightly. For the other three kinds of samples, both the generated spectra and the experimental spectra were discriminated as two classes in Fig. 6, which mean these two kinds of spectra also could not be distinguished. This also coincided with Fig. 5, which showed although dispersion ranges were larger for RDX, HMX and DMMP, generated spectra and experimental spectra were still mixed together.
Fig. 6 K-means clustering results for LIBS spectra of the five kinds of samples after training 180000 times.
The percentage of generated spectra which were classified into the same cluster with experimental spectra was used to evaluate the similarity between these two kinds of spectra. The values of this factor for all five kinds of samples were listed in Table 2. And as Table 2 shown, in our experiments, this factor was achieved above 59% for all kinds of samples. For CL-20 and TEP, it achieved 100%.
Table 2. The percentage of generated spectra was classified into experimental spectra cluster (PGCE) for all five kinds of samples.
3.3. Application for classification model based on both real spectra and generated spectra
As mentioned before, the generated spectra were inseparable from experimental spectra. So that they could be used to build the supervised classification model with real spectra together. As a common model used in classification of LIBS spectra, an SVM classifier [21] was used to evaluate the usefulness of the generated spectra.
100 spectra for each kind of samples were measured from experiments. Among them, 10 random spectra were used as training data set and the other 90 were used to test the model. When the SVM classifier model was trained by only ten measured spectra for each kind of hazardous samples, the classification results were showed in Fig. 7(a) and the correct classification rate (CCR) achieved at only 88.89%. Using the proposed method, we got 100 generated spectra with the ten mentioned spectra for each kind of sample as an extended database and used these generated spectra to build a new SVM classifier with experimental spectra together. As the results illustrated in Fig. 7(b), the new SVM classifier was used to identify the same testing spectral data set and achieved a better CCR at 95.33%. The classification model trained by the extended spectral database was obviously better than only trained by the experimental spectral database.
Fig. 7 SVM classification results for 5 kinds of hazardous samples ((a) The model trained with 10 experimental spectra for each kind of sample; (b) The model trained with 10 experimental spectra and 100 generated spectra for each kind of sample.).
4. Conclusions
As a famous spectroscopy method for substance detection and classification, LIBS is not a non-destructive detection method. Based on this, it is difficult to get enough spectra for building the model in some special fields. The proposed artificially spectral generating method based on GAN can extend the spectral database of LIBS spectra with a small amount of experimental spectral data. After enough interactive training, the generated spectra looked very similar with the real spectra. The two common unsupervised learning methods in spectra analysis field, PCA and K-means were used to evaluated the generated spectra. Based on their clustering results, the generated spectra finally could not be distinguished from real spectra after enough interactive training times. The experimental results demonstrated that the proposed GAN spectral generation method can effectively extend the spectral database.
Using both generated and experimental sample spectra to build the SVM model, the results showed that when there were only a few experimental spectra, the combination of the generated spectra and the experimental spectra for building the classification model could achieve better identification results. Generating spectra based on the proposed method for extending the spectral database can improve the application of LIBS for qualitative analysis. The spectral generation method based on GAN is a practical method can be used to extend the spectral database. For LIBS experiments, it can save precious samples in this way.
Funding
National Natural Science Foundation of China (NSFC) (61775017).
References
1. Z. Wang, F. Dong, and W. Zhou, “A Rising Force for the World-Wide Development of Laser-Induced Breakdown Spectroscopy,” Plasma Sci. Technol. 17(8), 617–620 (2015). [CrossRef]
2. S. A. Kalam, N. L. Murthy, P. Mathi, N. Kommu, A. K. Singh, and S. V. Rao, “Correlation of molecular, atomic emissions with detonation parameters in femtosecond and nanosecond LIBS plasma of high energy materials,” J. Anal. At. Spectrom. 32(8), 1535–1546 (2017). [CrossRef]
3. Q. Q. Wang, K. Liu, H. Zhao, C. H. Ge, and Z. W. Huang, “Detection of explosives with laser-induced breakdown spectroscopy,” Front. Phys-Beijing 7(6), 701–707 (2012).
4. R. Kanawade, F. Mahari, F. Klämpfl, M. Rohde, C. Knipfer, K. Tangermann-Gerk, W. Adler, M. Schmidt, and F. Stelzle, “Qualitative tissue differentiation by analysing the intensity ratios of atomic emission lines using laser induced breakdown spectroscopy (LIBS): prospects for a feedback mechanism for surgical laser systems,” J. Biophotonics 8(1-2), 153–161 (2015). [CrossRef] [PubMed]
5. S. J. Rehse, Q. I. Mohaidat, and S. Palchaudhuri, “Towards the clinical application of laser-induced breakdown spectroscopy for rapid pathogen diagnosis: the effect of mixed cultures and sample dilution on bacterial identification,” Appl. Opt. 49(13), C27–C35 (2010). [CrossRef]
6. D. Syvilay, X. S. Bai, N. Wilkie-Chancellier, A. Texier, L. Martinez, S. Serfaty, and V. Detalle, “Laser-induced emission, fluorescence and Raman hybrid setup: A versatile instrument to analyze materials from cultural heritage,” Spectrochim. Acta B At. Spectrosc. 140, 44–53 (2018). [CrossRef]
7. I. Osticioli, N. F. C. Mendes, S. Porcinai, A. Cagnini, and E. Castellucci, “Spectroscopic analysis of works of art using a single LIBS and pulsed Raman setup,” Anal. Bioanal. Chem. 394(4), 1033–1041 (2009). [CrossRef] [PubMed]
8. E. Vors, K. Tchepidjian, and J. B. Sirven, “Evaluation and optimization of the robustness of a multivariate analysis methodology for identification of alloys by laser induced breakdown spectroscopy,” Spectrochim. Acta B At. Spectrosc. 117, 16–22 (2016). [CrossRef]
9. J. Yang, X. Li, H. Lu, J. Xu, and H. Li, “An LIBS quantitative analysis method for alloy steel at high temperature based on transfer learning,” J. Anal. At. Spectrom. 33(7), 1184–1195 (2018). [CrossRef]
10. G. Bazalgette Courrèges-Lacoste, B. Ahlers, and F. R. Pérez, “Combined Raman spectrometer/laser-induced breakdown spectrometer for the next ESA mission to Mars,” Spectrochim. Acta A Mol. Biomol. Spectrosc. 68(4), 1023–1028 (2007). [CrossRef] [PubMed]
11. S. K. Sharma, A. K. Misra, P. G. Lucey, and R. C. Lentz, “A combined remote Raman and LIBS instrument for characterizing minerals with 532 nm laser excitation,” Spectrochim. Acta A Mol. Biomol. Spectrosc. 73(3), 468–476 (2009). [CrossRef] [PubMed]
12. V. C. Costa, D. V. de Babos, F. W. B. de Aquino, A. Virgílio, F. A. C. Amorim, and E. R. Pereira-Filho, “Direct Determination of Ca, K and Mg in Cassava Flour Samples by Laser-Induced Breakdown Spectroscopy (LIBS),” Food Anal. Methods 11(7), 1886–1896 (2018). [CrossRef]
13. M. Markiewicz-Keszycka, M. P. Casado-Gavalda, X. Cama-Moncunill, R. Cama-Moncunill, Y. Dixit, P. J. Cullen, and C. Sullivan, “Laser-induced breakdown spectroscopy (LIBS) for rapid analysis of ash, potassium and magnesium in gluten free flours,” Food Chem. 244, 324–330 (2018). [CrossRef] [PubMed]
14. D. W. Hahn and N. Omenetto, “Laser-Induced Breakdown Spectroscopy (LIBS), Part I: Review of Basic Diagnostics and Plasma-Particle Interactions: Still-Challenging Issues Within the Analytical Plasma Community,” Appl. Spectrosc. 64(12), 335–366 (2010). [CrossRef] [PubMed]
15. D. W. Hahn and N. Omenetto, “Laser-Induced Breakdown Spectroscopy (LIBS), Part II: Review of Instrumental and Methodological Approaches to Material Analysis and Applications to Different Fields,” Appl. Spectrosc. 66(4), 347–419 (2012). [CrossRef] [PubMed]
16. S. K. Anubham, R. Junjuri, A. K. Myakalwar, and M. K. Gundawar, “An Approach to Reduce the Sample Consumption for LIBS based Identification of Explosive Materials,” Def. Sci. J. 67(3), 254–259 (2017). [CrossRef]
17. I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” in 21st International Conference on Neural Information Processing (2014), pp. 2672–2680 (2014).
18. Q. Wang, G. Teng, X. Qiao, Y. Zhao, J. Kong, L. Dong, and X. Cui, “Importance evaluation of spectral lines in Laser-induced breakdown spectroscopy for classification of pathogenic bacteria,” Biomed. Opt. Express 9(11), 5837–5850 (2018). [CrossRef] [PubMed]
19. G. Yang, S. Yu, H. Dong, G. Slabaugh, P. L. Dragotti, X. Ye, F. Liu, S. Arridge, J. Keegan, Y. Guo, and D. Firmin, “DAGAN: Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction,” IEEE Trans. Med. Imaging 37(6), 1310–1321 (2018). [CrossRef] [PubMed]
20. Y. Liu, Z. Qin, T. Wan, and Z. Luo, “Auto-painter: Cartoon image generation from sketch by using conditional Wasserstein generative adversarial networks,” Neurocomputing 311, 78–87 (2018). [CrossRef]
21. F. Gao, F. Ma, J. Wang, J. Sun, E. Yang, and H. Zhou, “Semi-Supervised Generative Adversarial Nets with Multiple Generators for SAR Image Recognition,” Sensors (Basel) 18(8), 2706 (2018). [CrossRef] [PubMed]
22. L. Zhu, Y. Chen, P. Ghamisi, and J. A. Benediktsson, “Generative Adversarial Networks for Hyperspectral Image Classification,” IEEE T. Geosci. Remote 56(9), 5046–5063 (2018). [CrossRef]
23. T. Zhang, C. Yan, J. Qi, H. Tang, and H. Li, “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]
24. C. E. Porter, “Statistical properties of spectra,” Phys. Today 16(2), 26–36 (1963). [CrossRef]
25. A. C. Ranulfi, G. S. Senesi, J. B. Caetano, M. C. Meyer, A. B. Magalhães, P. R. Villas-Boas, and D. M. B. P. Milori, “Nutritional characterization of healthy and Aphelenchoides besseyi infected soybean leaves by laser-induced breakdown spectroscopy (LIBS),” Microchem. J. 141, 118–126 (2018). [CrossRef]
26. J. Wang, L. Li, P. Yang, Y. Chen, Y. Zhu, M. Tong, Z. Hao, and X. Li, “Identification of cervical cancer using laser-induced breakdown spectroscopy coupled with principal component analysis and support vector machine,” Lasers Med. Sci. 33(6), 1381–1386 (2018). [CrossRef] [PubMed]
27. Q. Q. Wang, L. A. He, Y. Zhao, Z. Peng, and L. Liu, “Study of cluster analysis used in explosives classification with laserinduced breakdown spectroscopy,” Laser Phys. 26(6), 065605 (2016). [CrossRef]
28. L. A. He, Q. Q. Wang, Y. Zhao, L. Liu, and Z. Peng, “Study on Cluster Analysis Used with Laser-Induced Breakdown Spectroscopy,” Plasma Sci. Technol. 18(6), 647–653 (2016). [CrossRef]
29. S. Wold, K. Esbensen, and P. Geladi, “Principal Component Analysis,” Chemometr. Intell. Lab. 2(1-3), 37–52 (1987). [CrossRef]
30. C. Lohrmann and P. Luukka, “A novel similarity classifier with multiple ideal vectors based on k-means clustering,” Decis. Support Syst. 111, 27–37 (2018). [CrossRef]
